Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
package main
|
|
|
|
|
|
|
|
import (
|
|
|
|
"context"
|
|
|
|
"encoding/json"
|
|
|
|
"errors"
|
|
|
|
"flag"
|
|
|
|
"fmt"
|
|
|
|
"log"
|
|
|
|
"log/slog"
|
|
|
|
"net"
|
|
|
|
"net/http"
|
|
|
|
"os"
|
|
|
|
"path/filepath"
|
|
|
|
"regexp"
|
|
|
|
"runtime"
|
2024-11-14 23:01:48 +00:00
|
|
|
"runtime/debug"
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
"strconv"
|
|
|
|
"strings"
|
|
|
|
"sync"
|
|
|
|
"time"
|
2024-10-21 18:07:19 +00:00
|
|
|
"unicode/utf8"
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
"golang.org/x/sync/semaphore"
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
"github.com/ollama/ollama/api"
|
|
|
|
"github.com/ollama/ollama/llama"
|
|
|
|
)
|
|
|
|
|
|
|
|
// input is an element of the prompt to process, either
|
|
|
|
// a token or an image embedding (generated from a vision projector)
|
|
|
|
type input struct {
|
|
|
|
token int
|
|
|
|
|
|
|
|
// embed is an image embedding
|
|
|
|
embed []float32
|
|
|
|
}
|
|
|
|
|
|
|
|
type Sequence struct {
|
|
|
|
// batch index
|
|
|
|
iBatch int
|
|
|
|
|
|
|
|
// number of tokens predicted so far
|
|
|
|
numPredicted int
|
|
|
|
|
|
|
|
// prompt inputs left to evaluate
|
|
|
|
inputs []input
|
|
|
|
|
|
|
|
// tokens that have been generated but not returned yet (e.g. for stop sequences)
|
|
|
|
pendingResponses []string
|
|
|
|
|
|
|
|
// input cache being used by this sequence
|
|
|
|
cache *InputCacheSlot
|
|
|
|
|
2024-10-31 17:55:31 +00:00
|
|
|
// does this sequence require cross-attention layers to be processed? - if we have seen
|
|
|
|
// an image for certain multi-modal models
|
|
|
|
crossAttention bool
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
// channel to send responses over
|
|
|
|
responses chan string
|
|
|
|
|
|
|
|
// channel to stop decoding (such as if the remote connection is closed)
|
|
|
|
quit chan bool
|
|
|
|
|
|
|
|
// number of tokens to predict
|
|
|
|
numPredict int
|
|
|
|
|
|
|
|
samplingCtx *llama.SamplingContext
|
|
|
|
|
|
|
|
// channel to send back the embedding if embedding only
|
|
|
|
embedding chan []float32
|
|
|
|
|
|
|
|
// stop sequences
|
|
|
|
stop []string
|
|
|
|
|
|
|
|
// number of inputs to keep at the beginning when shifting context window
|
|
|
|
numKeep int
|
|
|
|
|
|
|
|
// true if an embedding are to be returned instead of text generation
|
|
|
|
embeddingOnly bool
|
|
|
|
|
|
|
|
doneReason string
|
|
|
|
|
|
|
|
// Metrics
|
|
|
|
startProcessingTime time.Time
|
|
|
|
startGenerationTime time.Time
|
|
|
|
numDecoded int
|
|
|
|
numPromptInputs int
|
|
|
|
}
|
|
|
|
|
|
|
|
type NewSequenceParams struct {
|
|
|
|
numPredict int
|
|
|
|
stop []string
|
|
|
|
numKeep int
|
|
|
|
samplingParams *llama.SamplingParams
|
|
|
|
embedding bool
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) NewSequence(prompt string, images []ImageData, params NewSequenceParams) (*Sequence, error) {
|
|
|
|
s.ready.Wait()
|
|
|
|
|
|
|
|
startTime := time.Now()
|
|
|
|
|
|
|
|
inputs, err := s.inputs(prompt, images)
|
|
|
|
if err != nil {
|
|
|
|
return nil, fmt.Errorf("failed to process inputs: %w", err)
|
|
|
|
} else if len(inputs) == 0 {
|
|
|
|
return nil, errors.New("no input provided")
|
|
|
|
}
|
|
|
|
|
|
|
|
if params.numKeep < 0 {
|
|
|
|
params.numKeep = len(inputs)
|
|
|
|
}
|
|
|
|
|
2024-11-08 19:10:56 +00:00
|
|
|
if s.model.AddBOSToken() {
|
|
|
|
params.numKeep += 1
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
2024-11-08 19:10:56 +00:00
|
|
|
// Ensure that at least 1 input can be discarded during shift
|
|
|
|
params.numKeep = min(params.numKeep, s.cache.numCtx-1)
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
if len(inputs) > s.cache.numCtx {
|
2024-11-08 19:10:56 +00:00
|
|
|
slog.Warn("input exceeds context length", "prompt", len(inputs), "limit", s.cache.numCtx)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
var sc *llama.SamplingContext
|
|
|
|
if params.samplingParams != nil {
|
2024-11-01 22:50:53 +00:00
|
|
|
sc, err = llama.NewSamplingContext(s.model, *params.samplingParams)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
for _, input := range inputs {
|
|
|
|
if input.embed == nil {
|
2024-10-17 18:59:52 +00:00
|
|
|
sc.Accept(input.token, false)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return &Sequence{
|
|
|
|
inputs: inputs,
|
|
|
|
numPromptInputs: len(inputs),
|
|
|
|
startProcessingTime: startTime,
|
|
|
|
numPredict: params.numPredict,
|
|
|
|
pendingResponses: make([]string, 0),
|
|
|
|
responses: make(chan string, 100),
|
|
|
|
quit: make(chan bool, 1),
|
|
|
|
embedding: make(chan []float32, 1),
|
|
|
|
samplingCtx: sc,
|
|
|
|
embeddingOnly: params.embedding,
|
|
|
|
stop: params.stop,
|
|
|
|
numKeep: params.numKeep,
|
|
|
|
}, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// inputs processes the prompt and images into a list of inputs
|
|
|
|
// by splitting the prompt on [img-<n>] tags, tokenizing text and
|
|
|
|
// generating image embeddings for each image
|
|
|
|
func (s *Server) inputs(prompt string, images []ImageData) ([]input, error) {
|
|
|
|
var inputs []input
|
|
|
|
|
|
|
|
re := regexp.MustCompile(`\[img-(\d+)\]`)
|
|
|
|
parts := re.Split(prompt, -1)
|
|
|
|
matches := re.FindAllStringSubmatch(prompt, -1)
|
|
|
|
|
|
|
|
for i, part := range parts {
|
|
|
|
// text - tokenize
|
2024-11-14 00:49:01 +00:00
|
|
|
tokens, err := s.lc.Model().Tokenize(part, i == 0, true)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
2024-11-14 00:49:01 +00:00
|
|
|
for _, t := range tokens {
|
|
|
|
inputs = append(inputs, input{token: t})
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// image - generate image embedding
|
|
|
|
if i < len(matches) {
|
|
|
|
n, _ := strconv.Atoi(matches[i][1])
|
|
|
|
|
|
|
|
imageIndex := -1
|
|
|
|
for j := range images {
|
|
|
|
if images[j].ID == n {
|
|
|
|
imageIndex = j
|
|
|
|
break
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if imageIndex < 0 {
|
|
|
|
return nil, fmt.Errorf("invalid image index: %d", n)
|
|
|
|
}
|
|
|
|
|
2024-11-01 22:50:53 +00:00
|
|
|
embed, err := s.image.NewEmbed(s.lc, images[imageIndex].Data, images[imageIndex].AspectRatioID)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
for _, e := range embed {
|
|
|
|
inputs = append(inputs, input{embed: e})
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return inputs, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
type Server struct {
|
2024-11-12 19:23:46 +00:00
|
|
|
// is the server ready to process requests?
|
|
|
|
// protects access to model and image
|
|
|
|
ready sync.WaitGroup
|
|
|
|
|
|
|
|
// loaded model
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
model *llama.Model
|
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// image model context for multi-modal models
|
2024-10-11 22:34:01 +00:00
|
|
|
image *ImageContext
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// status for external health reporting - loading, ready to serve, etc.
|
|
|
|
status ServerStatus
|
|
|
|
|
|
|
|
// current progress on loading the model
|
|
|
|
progress float32
|
|
|
|
|
|
|
|
// number of simultaneous requests to handle
|
|
|
|
parallel int
|
|
|
|
|
|
|
|
// maximum number of elements in a batch (per sequence)
|
2024-11-01 21:29:57 +00:00
|
|
|
// TODO (jmorganca): make this n_batch
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
batchSize int
|
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// protects access to everything below this line
|
|
|
|
// this is context state needed for decoding
|
|
|
|
mu sync.Mutex
|
|
|
|
|
|
|
|
// indicates that data is ready for processing
|
|
|
|
cond *sync.Cond
|
|
|
|
|
|
|
|
// decoding state
|
|
|
|
lc *llama.Context
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// the list of simultaneous sequences being evaluated
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
seqs []*Sequence
|
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// seqs can have a maximum of parallel entries, which
|
|
|
|
// is enfoced by seqSem
|
|
|
|
seqsSem *semaphore.Weighted
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
// KV cache
|
|
|
|
cache *InputCache
|
|
|
|
|
|
|
|
// next sequence for prompt processing to avoid starvation
|
|
|
|
nextSeq int
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) allNil() bool {
|
|
|
|
for _, item := range s.seqs {
|
|
|
|
if item != nil {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
func flushPending(seq *Sequence) bool {
|
2024-10-21 18:07:19 +00:00
|
|
|
joined := strings.Join(seq.pendingResponses, "")
|
|
|
|
seq.pendingResponses = []string{}
|
|
|
|
|
|
|
|
// Check if there are any partial UTF-8 characters remaining.
|
|
|
|
// We already check and queue as we are generating but some may
|
|
|
|
// still make it here:
|
|
|
|
// - Sequence is ending, e.g. generation limit has been hit
|
|
|
|
// - Invalid characters in the middle of a string
|
|
|
|
// This is a stricter check to ensure we never output invalid Unicode.
|
|
|
|
for !utf8.ValidString(joined) {
|
|
|
|
joined = joined[:len(joined)-1]
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
2024-10-21 18:07:19 +00:00
|
|
|
if len(joined) == 0 {
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
select {
|
|
|
|
case seq.responses <- joined:
|
|
|
|
return true
|
|
|
|
case <-seq.quit:
|
|
|
|
return false
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) removeSequence(seqIndex int, reason string) {
|
|
|
|
seq := s.seqs[seqIndex]
|
|
|
|
|
|
|
|
flushPending(seq)
|
|
|
|
seq.doneReason = reason
|
|
|
|
close(seq.responses)
|
|
|
|
close(seq.embedding)
|
|
|
|
seq.cache.InUse = false
|
|
|
|
s.seqs[seqIndex] = nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) run(ctx context.Context) {
|
|
|
|
s.ready.Wait()
|
|
|
|
|
2024-11-01 21:29:57 +00:00
|
|
|
// Logically these batches are used only within the context of processBatch
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
// but it is better for performance to allocate them once here
|
2024-11-01 22:50:53 +00:00
|
|
|
tokenBatch, err := llama.NewBatch(s.batchSize, len(s.seqs), 0)
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
defer tokenBatch.Free()
|
|
|
|
|
2024-11-01 21:29:57 +00:00
|
|
|
var embedBatch *llama.Batch
|
|
|
|
embedBatchSize := s.image.BatchSize(s.batchSize)
|
|
|
|
if embedBatchSize != 0 {
|
2024-11-01 22:50:53 +00:00
|
|
|
embedBatch, err = llama.NewBatch(embedBatchSize, len(s.seqs), s.image.EmbedSize(s.lc))
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
2024-11-01 21:29:57 +00:00
|
|
|
defer embedBatch.Free()
|
|
|
|
} else {
|
|
|
|
embedBatch = &llama.Batch{}
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
for {
|
|
|
|
select {
|
|
|
|
case <-ctx.Done():
|
|
|
|
return
|
|
|
|
default:
|
|
|
|
s.processBatch(tokenBatch, embedBatch)
|
|
|
|
tokenBatch.Clear()
|
|
|
|
embedBatch.Clear()
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO (jmorganca): processBatch should be simplified, removing:
|
|
|
|
// * sampling
|
|
|
|
// * stop token checking
|
|
|
|
// * metrics
|
|
|
|
// these should instead be handled by the handlers
|
|
|
|
// it should only be responsible for accepting tokens or embeddings and
|
|
|
|
// processing batches as fast as possible
|
|
|
|
func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch) {
|
2024-11-14 23:01:48 +00:00
|
|
|
// Try to keep going even if we hit a panic so that corner cases don't take the whole
|
|
|
|
// runner down. In most cases, this will result in dropping the tokens that we are currently
|
|
|
|
// processing and then continuing with what is remaining.
|
|
|
|
defer func() {
|
|
|
|
if err := recover(); err != nil {
|
|
|
|
slog.Error("error while processing batch", "error", err, "stack", debug.Stack())
|
|
|
|
}
|
|
|
|
}()
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.mu.Lock()
|
|
|
|
for s.allNil() {
|
|
|
|
s.cond.Wait() // Wait until an item is added
|
|
|
|
}
|
|
|
|
defer s.mu.Unlock()
|
|
|
|
|
|
|
|
var batch *llama.Batch
|
2024-10-31 17:55:31 +00:00
|
|
|
crossAttention := false
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
seqIdx := s.nextSeq - 1
|
|
|
|
for range s.seqs {
|
|
|
|
seqIdx = (seqIdx + 1) % len(s.seqs)
|
|
|
|
seq := s.seqs[seqIdx]
|
|
|
|
|
|
|
|
if seq == nil {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
2024-11-14 23:01:48 +00:00
|
|
|
// If an error occurred during the processing of a previous batch then we may have emptied the inputs
|
|
|
|
// without adding a new one. In this case, end the sequence rather than infinite looping.
|
|
|
|
if len(seq.inputs) == 0 {
|
|
|
|
slog.Error("removing sequence due to no input tokens", "index", seqIdx, "cache id", seq.cache.Id)
|
|
|
|
s.removeSequence(seqIdx, "error")
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
// if past the num predict limit
|
2024-11-12 18:41:44 +00:00
|
|
|
if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.removeSequence(seqIdx, "limit")
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
var numInputsProcessed int
|
2024-11-08 19:10:56 +00:00
|
|
|
shifted := false
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
for i, input := range seq.inputs {
|
2024-11-08 19:10:56 +00:00
|
|
|
if len(seq.cache.Inputs)+1 > s.cache.numCtx {
|
|
|
|
if !shifted {
|
|
|
|
s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
|
|
|
|
shifted = true
|
|
|
|
} else {
|
|
|
|
break
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
embedding := input.embed != nil
|
|
|
|
|
|
|
|
// If we don't currently have a batch, use one of the correct type and
|
|
|
|
// fill it up as much as possible across all sequences. If we encounter an
|
|
|
|
// input of the opppsite type, stop for that sequence but then pick up from
|
|
|
|
// there for the next batch, ensuring that we alternate types
|
|
|
|
if batch == nil {
|
|
|
|
if !embedding {
|
|
|
|
batch = tokenBatch
|
|
|
|
} else {
|
|
|
|
batch = embedBatch
|
2024-10-31 17:55:31 +00:00
|
|
|
seq.crossAttention = s.image.NeedCrossAttention(input)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
2024-10-31 17:55:31 +00:00
|
|
|
} else if embedding != batch.IsEmbedding() || crossAttention != seq.crossAttention {
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.nextSeq = seqIdx
|
|
|
|
break
|
|
|
|
}
|
|
|
|
|
2024-11-01 21:29:57 +00:00
|
|
|
if i >= batch.Size() {
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
break
|
|
|
|
}
|
|
|
|
|
2024-10-31 17:55:31 +00:00
|
|
|
crossAttention = seq.crossAttention
|
2024-11-08 19:10:56 +00:00
|
|
|
batch.Add(input.token, input.embed, len(seq.cache.Inputs), i+1 == len(seq.inputs), seq.cache.Id)
|
|
|
|
seq.cache.Inputs = append(seq.cache.Inputs, input)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
numInputsProcessed++
|
|
|
|
}
|
|
|
|
|
|
|
|
if numInputsProcessed > 0 {
|
|
|
|
seq.inputs = seq.inputs[numInputsProcessed:]
|
|
|
|
seq.iBatch = batch.NumTokens() - 1
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if batch == nil || batch.NumTokens() == 0 {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2024-10-31 17:55:31 +00:00
|
|
|
s.lc.SetCrossAttention(crossAttention)
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
err := s.lc.Decode(batch)
|
|
|
|
if err != nil {
|
|
|
|
slog.Error("failed to decode batch", "error", err)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2024-11-13 22:12:30 +00:00
|
|
|
if crossAttention {
|
|
|
|
// synchronize state to ensure the cross attention batch is complete.
|
|
|
|
// needed specifically for multi-GPU systems otherwise an inflight
|
|
|
|
// task may be incorrectly invalidated causing a crash
|
|
|
|
s.lc.Synchronize()
|
|
|
|
}
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
for i, seq := range s.seqs {
|
|
|
|
if seq == nil {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
// don't sample prompt processing
|
|
|
|
if len(seq.inputs) != 0 {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
seq.numDecoded += 1
|
|
|
|
if seq.numDecoded == 1 {
|
|
|
|
seq.startGenerationTime = time.Now()
|
|
|
|
}
|
|
|
|
|
|
|
|
// if done processing the prompt, generate an embedding and return
|
|
|
|
if seq.embeddingOnly {
|
|
|
|
embed := s.lc.GetEmbeddingsSeq(i)
|
|
|
|
if embed == nil {
|
|
|
|
embed = s.lc.GetEmbeddingsIth(seq.iBatch)
|
|
|
|
}
|
|
|
|
|
|
|
|
seq.embedding <- embed
|
|
|
|
s.removeSequence(i, "")
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
// sample a token
|
2024-10-17 18:59:52 +00:00
|
|
|
token := seq.samplingCtx.Sample(s.lc, seq.iBatch)
|
|
|
|
seq.samplingCtx.Accept(token, true)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
piece := s.model.TokenToPiece(token)
|
|
|
|
|
|
|
|
seq.numPredicted++
|
|
|
|
|
|
|
|
// if it's an end of sequence token, break
|
|
|
|
if s.model.TokenIsEog(token) {
|
|
|
|
// TODO (jmorganca): we should send this back
|
|
|
|
// as it's important for the /api/generate context
|
|
|
|
// seq.responses <- piece
|
|
|
|
|
|
|
|
s.removeSequence(i, "stop")
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
seq.inputs = []input{{token: token}}
|
|
|
|
|
|
|
|
seq.pendingResponses = append(seq.pendingResponses, piece)
|
|
|
|
sequence := strings.Join(seq.pendingResponses, "")
|
|
|
|
|
|
|
|
if ok, stop := findStop(sequence, seq.stop); ok {
|
2024-10-09 23:12:23 +00:00
|
|
|
slog.Debug("hit stop token", "pending", seq.pendingResponses, "stop", stop)
|
|
|
|
|
|
|
|
var tokenTruncated bool
|
|
|
|
origLen := len(seq.pendingResponses)
|
|
|
|
seq.pendingResponses, tokenTruncated = truncateStop(seq.pendingResponses, stop)
|
|
|
|
newLen := len(seq.pendingResponses)
|
|
|
|
|
|
|
|
// Update the cache based on the tokens that will be returned:
|
|
|
|
// - We have 1 token more than is currently in the cache because
|
|
|
|
// the last one generated wasn't submitted to Decode
|
|
|
|
// - Remove any stop sequences that we stripped out
|
|
|
|
// - If truncateStop removed a portion of a token, drop that
|
|
|
|
// - As defense-in-depth, if truncatedToken didn't find a stop token
|
|
|
|
// remove the extra one that we added to the cache len
|
|
|
|
tokenLen := len(seq.cache.Inputs) + 1
|
|
|
|
tokenLen -= origLen - newLen
|
|
|
|
if tokenTruncated || origLen == newLen {
|
|
|
|
tokenLen--
|
|
|
|
}
|
|
|
|
seq.cache.Inputs = seq.cache.Inputs[:tokenLen]
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
s.removeSequence(i, "stop")
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
if containsStopSuffix(sequence, seq.stop) {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
if incompleteUnicode(sequence) {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
if !flushPending(seq) {
|
|
|
|
s.removeSequence(i, "connection")
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO (jmorganca): use structs from the api package to avoid duplication
|
|
|
|
// this way the api acts as a proxy instead of using a different api for the
|
|
|
|
// runner
|
|
|
|
type Options struct {
|
|
|
|
api.Runner
|
|
|
|
|
|
|
|
NumKeep int `json:"n_keep"`
|
|
|
|
Seed int `json:"seed"`
|
|
|
|
NumPredict int `json:"n_predict"`
|
|
|
|
TopK int `json:"top_k"`
|
|
|
|
TopP float32 `json:"top_p"`
|
|
|
|
MinP float32 `json:"min_p"`
|
|
|
|
TFSZ float32 `json:"tfs_z"`
|
|
|
|
TypicalP float32 `json:"typical_p"`
|
|
|
|
RepeatLastN int `json:"repeat_last_n"`
|
|
|
|
Temperature float32 `json:"temperature"`
|
|
|
|
RepeatPenalty float32 `json:"repeat_penalty"`
|
|
|
|
PresencePenalty float32 `json:"presence_penalty"`
|
|
|
|
FrequencyPenalty float32 `json:"frequency_penalty"`
|
|
|
|
Mirostat int `json:"mirostat"`
|
|
|
|
MirostatTau float32 `json:"mirostat_tau"`
|
|
|
|
MirostatEta float32 `json:"mirostat_eta"`
|
|
|
|
PenalizeNewline bool `json:"penalize_nl"`
|
|
|
|
Stop []string `json:"stop"`
|
|
|
|
}
|
|
|
|
|
|
|
|
type ImageData struct {
|
2024-10-18 23:12:35 +00:00
|
|
|
Data []byte `json:"data"`
|
|
|
|
ID int `json:"id"`
|
|
|
|
AspectRatioID int `json:"aspect_ratio_id"`
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
type CompletionRequest struct {
|
|
|
|
Prompt string `json:"prompt"`
|
|
|
|
Images []ImageData `json:"image_data"`
|
|
|
|
Grammar string `json:"grammar"`
|
|
|
|
CachePrompt bool `json:"cache_prompt"`
|
|
|
|
|
|
|
|
Options
|
|
|
|
}
|
|
|
|
|
|
|
|
type Timings struct {
|
|
|
|
PredictedN int `json:"predicted_n"`
|
|
|
|
PredictedMS float64 `json:"predicted_ms"`
|
|
|
|
PromptN int `json:"prompt_n"`
|
|
|
|
PromptMS float64 `json:"prompt_ms"`
|
|
|
|
}
|
|
|
|
|
|
|
|
type CompletionResponse struct {
|
|
|
|
Content string `json:"content"`
|
|
|
|
Stop bool `json:"stop"`
|
|
|
|
|
|
|
|
Model string `json:"model,omitempty"`
|
|
|
|
Prompt string `json:"prompt,omitempty"`
|
|
|
|
StoppedLimit bool `json:"stopped_limit,omitempty"`
|
|
|
|
PredictedN int `json:"predicted_n,omitempty"`
|
|
|
|
PredictedMS float64 `json:"predicted_ms,omitempty"`
|
|
|
|
PromptN int `json:"prompt_n,omitempty"`
|
|
|
|
PromptMS float64 `json:"prompt_ms,omitempty"`
|
|
|
|
|
|
|
|
Timings Timings `json:"timings"`
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
|
|
|
|
var req CompletionRequest
|
|
|
|
req.Options = Options(api.DefaultOptions())
|
|
|
|
if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
|
|
|
|
http.Error(w, "Bad request", http.StatusBadRequest)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set the headers to indicate streaming
|
|
|
|
w.Header().Set("Content-Type", "application/json")
|
|
|
|
w.Header().Set("Transfer-Encoding", "chunked")
|
|
|
|
|
|
|
|
flusher, ok := w.(http.Flusher)
|
|
|
|
if !ok {
|
|
|
|
http.Error(w, "Streaming not supported", http.StatusInternalServerError)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
var samplingParams llama.SamplingParams
|
|
|
|
samplingParams.TopK = req.TopK
|
|
|
|
samplingParams.TopP = req.TopP
|
|
|
|
samplingParams.MinP = req.MinP
|
|
|
|
samplingParams.TfsZ = req.TFSZ
|
|
|
|
samplingParams.TypicalP = req.TypicalP
|
|
|
|
samplingParams.Temp = req.Temperature
|
|
|
|
samplingParams.RepeatLastN = req.RepeatLastN
|
|
|
|
samplingParams.PenaltyRepeat = req.RepeatPenalty
|
|
|
|
samplingParams.PenaltyFreq = req.FrequencyPenalty
|
|
|
|
samplingParams.PenaltyPresent = req.PresencePenalty
|
|
|
|
samplingParams.Mirostat = req.Mirostat
|
|
|
|
samplingParams.MirostatTau = req.MirostatTau
|
|
|
|
samplingParams.MirostatEta = req.MirostatEta
|
|
|
|
samplingParams.PenalizeNl = req.PenalizeNewline
|
|
|
|
samplingParams.Seed = uint32(req.Seed)
|
|
|
|
samplingParams.Grammar = req.Grammar
|
|
|
|
|
|
|
|
seq, err := s.NewSequence(req.Prompt, req.Images, NewSequenceParams{
|
|
|
|
numPredict: req.NumPredict,
|
|
|
|
stop: req.Stop,
|
|
|
|
numKeep: req.NumKeep,
|
|
|
|
samplingParams: &samplingParams,
|
|
|
|
embedding: false,
|
|
|
|
})
|
|
|
|
if err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// Ensure that a place to put the sequence is available
|
|
|
|
if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
|
|
|
|
slog.Error("Failed to acquire semaphore", "error", err)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
defer s.seqsSem.Release(1)
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.mu.Lock()
|
|
|
|
for i, sq := range s.seqs {
|
|
|
|
if sq == nil {
|
2024-11-08 19:10:56 +00:00
|
|
|
seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
if err != nil {
|
|
|
|
s.mu.Unlock()
|
|
|
|
http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
|
|
|
|
return
|
|
|
|
}
|
2024-10-11 22:34:01 +00:00
|
|
|
|
2024-10-31 17:55:31 +00:00
|
|
|
seq.crossAttention = s.image.NeedCrossAttention(seq.cache.Inputs...)
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.seqs[i] = seq
|
|
|
|
s.cond.Signal()
|
|
|
|
break
|
|
|
|
}
|
|
|
|
}
|
|
|
|
s.mu.Unlock()
|
|
|
|
|
|
|
|
for {
|
|
|
|
select {
|
|
|
|
case <-r.Context().Done():
|
|
|
|
close(seq.quit)
|
|
|
|
return
|
|
|
|
case content, ok := <-seq.responses:
|
|
|
|
if ok {
|
|
|
|
if err := json.NewEncoder(w).Encode(&CompletionResponse{
|
|
|
|
Content: content,
|
|
|
|
}); err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
|
|
|
|
close(seq.quit)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
flusher.Flush()
|
|
|
|
} else {
|
|
|
|
// Send the final response
|
|
|
|
if err := json.NewEncoder(w).Encode(&CompletionResponse{
|
|
|
|
Stop: true,
|
|
|
|
StoppedLimit: seq.doneReason == "limit",
|
|
|
|
Timings: Timings{
|
|
|
|
PromptN: seq.numPromptInputs,
|
|
|
|
PromptMS: float64(seq.startGenerationTime.Sub(seq.startProcessingTime).Milliseconds()),
|
|
|
|
PredictedN: seq.numDecoded,
|
|
|
|
PredictedMS: float64(time.Since(seq.startGenerationTime).Milliseconds()),
|
|
|
|
},
|
|
|
|
}); err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("failed to encode final response: %v", err), http.StatusInternalServerError)
|
|
|
|
}
|
|
|
|
|
|
|
|
return
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
type EmbeddingRequest struct {
|
|
|
|
Content string `json:"content"`
|
|
|
|
CachePrompt bool `json:"cache_prompt"`
|
|
|
|
}
|
|
|
|
|
|
|
|
type EmbeddingResponse struct {
|
|
|
|
Embedding []float32 `json:"embedding"`
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
|
|
|
|
var req EmbeddingRequest
|
|
|
|
if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("bad request: %s", err), http.StatusBadRequest)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
w.Header().Set("Content-Type", "application/json")
|
|
|
|
|
|
|
|
slog.Debug("embedding request", "content", req.Content)
|
|
|
|
|
|
|
|
seq, err := s.NewSequence(req.Content, nil, NewSequenceParams{embedding: true})
|
|
|
|
if err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2024-11-12 19:23:46 +00:00
|
|
|
// Ensure that a place to put the sequence is available
|
|
|
|
if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
|
|
|
|
slog.Error("Failed to acquire semaphore", "error", err)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
defer s.seqsSem.Release(1)
|
|
|
|
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
s.mu.Lock()
|
|
|
|
for i, sq := range s.seqs {
|
|
|
|
if sq == nil {
|
2024-11-08 19:10:56 +00:00
|
|
|
seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
if err != nil {
|
|
|
|
s.mu.Unlock()
|
|
|
|
http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
s.seqs[i] = seq
|
|
|
|
s.cond.Signal()
|
|
|
|
break
|
|
|
|
}
|
|
|
|
}
|
|
|
|
s.mu.Unlock()
|
|
|
|
|
|
|
|
embedding := <-seq.embedding
|
|
|
|
|
|
|
|
if err := json.NewEncoder(w).Encode(&EmbeddingResponse{
|
|
|
|
Embedding: embedding,
|
|
|
|
}); err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
type HealthResponse struct {
|
|
|
|
Status string `json:"status"`
|
|
|
|
Progress float32 `json:"progress"`
|
|
|
|
}
|
|
|
|
|
|
|
|
type ServerStatus int
|
|
|
|
|
|
|
|
const (
|
|
|
|
ServerStatusReady ServerStatus = iota
|
|
|
|
ServerStatusLoadingModel
|
|
|
|
ServerStatusError
|
|
|
|
)
|
|
|
|
|
|
|
|
func (s ServerStatus) ToString() string {
|
|
|
|
switch s {
|
|
|
|
case ServerStatusReady:
|
|
|
|
return "ok"
|
|
|
|
case ServerStatusLoadingModel:
|
|
|
|
return "loading model"
|
|
|
|
default:
|
|
|
|
return "server error"
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) health(w http.ResponseWriter, r *http.Request) {
|
|
|
|
w.Header().Set("Content-Type", "application/json")
|
|
|
|
if err := json.NewEncoder(w).Encode(&HealthResponse{
|
|
|
|
Status: s.status.ToString(),
|
|
|
|
Progress: s.progress,
|
|
|
|
}); err != nil {
|
|
|
|
http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *Server) loadModel(
|
|
|
|
params llama.ModelParams,
|
|
|
|
mpath string,
|
|
|
|
lpath string,
|
|
|
|
ppath string,
|
|
|
|
kvSize int,
|
|
|
|
flashAttention bool,
|
|
|
|
threads int,
|
|
|
|
multiUserCache bool,
|
|
|
|
) {
|
|
|
|
llama.BackendInit()
|
|
|
|
|
2024-10-22 21:57:46 +00:00
|
|
|
var err error
|
|
|
|
s.model, err = llama.LoadModelFromFile(mpath, params)
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
ctxParams := llama.NewContextParams(kvSize, s.batchSize*s.parallel, s.parallel, threads, flashAttention)
|
2024-10-22 21:57:46 +00:00
|
|
|
s.lc, err = llama.NewContextWithModel(s.model, ctxParams)
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
if lpath != "" {
|
|
|
|
err := s.model.ApplyLoraFromFile(s.lc, lpath, 1.0, threads)
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if ppath != "" {
|
2024-10-18 23:12:35 +00:00
|
|
|
var err error
|
2024-10-11 22:34:01 +00:00
|
|
|
s.image, err = NewImageContext(s.lc, ppath)
|
2024-10-18 23:12:35 +00:00
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
}
|
|
|
|
|
2024-11-08 19:10:56 +00:00
|
|
|
s.cache, err = NewInputCache(s.lc, kvSize, s.parallel, multiUserCache)
|
|
|
|
if err != nil {
|
|
|
|
panic(err)
|
|
|
|
}
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
s.status = ServerStatusReady
|
|
|
|
s.ready.Done()
|
|
|
|
}
|
|
|
|
|
|
|
|
func main() {
|
|
|
|
mpath := flag.String("model", "", "Path to model binary file")
|
|
|
|
ppath := flag.String("mmproj", "", "Path to projector binary file")
|
|
|
|
parallel := flag.Int("parallel", 1, "Number of sequences to handle simultaneously")
|
|
|
|
batchSize := flag.Int("batch-size", 512, "Batch size")
|
|
|
|
nGpuLayers := flag.Int("n-gpu-layers", 0, "Number of layers to offload to GPU")
|
|
|
|
mainGpu := flag.Int("main-gpu", 0, "Main GPU")
|
|
|
|
flashAttention := flag.Bool("flash-attn", false, "Enable flash attention")
|
|
|
|
kvSize := flag.Int("ctx-size", 2048, "Context (or KV cache) size")
|
|
|
|
lpath := flag.String("lora", "", "Path to lora layer file")
|
|
|
|
port := flag.Int("port", 8080, "Port to expose the server on")
|
|
|
|
threads := flag.Int("threads", runtime.NumCPU(), "Number of threads to use during generation")
|
|
|
|
verbose := flag.Bool("verbose", false, "verbose output (default: disabled)")
|
|
|
|
noMmap := flag.Bool("no-mmap", false, "do not memory-map model (slower load but may reduce pageouts if not using mlock)")
|
|
|
|
mlock := flag.Bool("mlock", false, "force system to keep model in RAM rather than swapping or compressing")
|
|
|
|
tensorSplit := flag.String("tensor-split", "", "fraction of the model to offload to each GPU, comma-separated list of proportions")
|
|
|
|
multiUserCache := flag.Bool("multiuser-cache", false, "optimize input cache algorithm for multiple users")
|
|
|
|
requirements := flag.Bool("requirements", false, "print json requirement information")
|
|
|
|
|
|
|
|
flag.Parse()
|
|
|
|
if *requirements {
|
|
|
|
printRequirements(os.Stdout)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
level := slog.LevelInfo
|
|
|
|
if *verbose {
|
|
|
|
level = slog.LevelDebug
|
|
|
|
}
|
|
|
|
handler := slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{
|
|
|
|
Level: level,
|
|
|
|
AddSource: true,
|
|
|
|
ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
|
|
|
|
if attr.Key == slog.SourceKey {
|
|
|
|
source := attr.Value.Any().(*slog.Source)
|
|
|
|
source.File = filepath.Base(source.File)
|
|
|
|
}
|
|
|
|
return attr
|
|
|
|
},
|
|
|
|
})
|
|
|
|
slog.SetDefault(slog.New(handler))
|
|
|
|
slog.Info("starting go runner")
|
2024-10-30 19:28:36 +00:00
|
|
|
slog.Info("system", "info", llama.PrintSystemInfo(), "threads", *threads)
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
|
|
|
|
server := &Server{
|
|
|
|
batchSize: *batchSize,
|
|
|
|
parallel: *parallel,
|
|
|
|
seqs: make([]*Sequence, *parallel),
|
2024-11-12 19:23:46 +00:00
|
|
|
seqsSem: semaphore.NewWeighted(int64(*parallel)),
|
Re-introduce the `llama` package (#5034)
* Re-introduce the llama package
This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:
- C APIs can be called directly from Go without needing to use the previous
"server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
a go generate ./... step, making it easy to get up and running to hack on
parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source
This is a big PR, but much of it is vendor code except for:
- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* cache: Clear old KV cache entries when evicting a slot
When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.
This change fixes two issues:
- The KV cache fills up and runs out of space even though we think
we are managing it correctly
- Performance gets worse over time as we use new cache entries that
are not hot in the processor caches
* doc: explain golang objc linker warning (#6830)
* llama: gather transitive dependencies for rocm for dist packaging (#6848)
* Refine go server makefiles to be more DRY (#6924)
This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles. This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.
When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.
* llama: don't create extraneous directories (#6988)
* llama: Exercise the new build in CI (#6989)
Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet.
* llama: Refine developer docs for Go server (#6842)
This enhances the documentation for development focusing on the new Go
server. After we complete the transition further doc refinements
can remove the "transition" discussion.
* runner.go: Allocate batches for all sequences during init
We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.
* llama.go: Don't return nil from Tokenize on zero length input
Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.
* runner.go: Remove stop tokens from cache
If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.
However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.
This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.
By trimming the cache to the tokens that we actually return this
issue can be avoided.
* runner.go: Simplify flushing of pending tokens
* runner.go: Update TODOs
* runner.go: Don't panic when processing sequences
If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.
Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: More accurately capture timings
Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.
* runner.go: Support for vision models
In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
- Cache prompting works with images, avoiding the need to re-decode
embeddings for every message in a conversation
- Parallelism is supported, avoiding the need to restrict to one
sequence at a time. (Though for now Ollama will not schedule
them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* runner.go: Move Unicode checking code and add tests
* runner.go: Export external cache members
Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.
* runner.go: Image embedding cache
Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.
This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.
* llama: catch up on patches
Carry forward solar-pro and cli-unicode patches
* runner.go: Don't re-allocate memory for every batch
We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.
This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.
* runner.go: Default to classic input cache policy
The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.
However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).
This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.
For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.
* runner.go: Increase size of response channel
Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.
* llama: Add CI to verify all vendored changes have patches (#7066)
Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.
* llama: adjust clip patch for mingw utf-16 (#7065)
* llama: adjust clip patch for mingw utf-16
* llama: ensure static linking of runtime libs
Avoid runtime dependencies on non-standard libraries
* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)
These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.
* llm: Don't add BOS/EOS for tokenize requests
This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.
* runner.go: Don't cache prompts for embeddings
Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.
Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.
* runner.go: Adjust debug log levels
Add system info printed at startup and quiet down noisier logging.
* llama: fix compiler flag differences (#7082)
Adjust the flags for the new Go server to more closely match the
generate flow
* llama: refine developer docs (#7121)
* llama: doc and example clean up (#7122)
* llama: doc and example clean up
* llama: Move new dockerfile into llama dir
Temporary home until we fully transition to the Go server
* llama: runner doc cleanup
* llama.go: Add description for Tokenize error case
---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
2024-10-08 15:53:54 +00:00
|
|
|
status: ServerStatusLoadingModel,
|
|
|
|
}
|
|
|
|
|
|
|
|
var tensorSplitFloats []float32
|
|
|
|
if *tensorSplit != "" {
|
|
|
|
stringFloats := regexp.MustCompile(",").Split(*tensorSplit, -1)
|
|
|
|
|
|
|
|
tensorSplitFloats = make([]float32, 0, len(stringFloats))
|
|
|
|
for _, s := range stringFloats {
|
|
|
|
f, _ := strconv.ParseFloat(s, 32)
|
|
|
|
tensorSplitFloats = append(tensorSplitFloats, float32(f))
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
params := llama.ModelParams{
|
|
|
|
NumGpuLayers: *nGpuLayers,
|
|
|
|
MainGpu: *mainGpu,
|
|
|
|
UseMmap: !*noMmap && *lpath == "",
|
|
|
|
UseMlock: *mlock,
|
|
|
|
TensorSplit: tensorSplitFloats,
|
|
|
|
Progress: func(progress float32) {
|
|
|
|
server.progress = progress
|
|
|
|
},
|
|
|
|
}
|
|
|
|
|
|
|
|
server.ready.Add(1)
|
|
|
|
go server.loadModel(params, *mpath, *lpath, *ppath, *kvSize, *flashAttention, *threads, *multiUserCache)
|
|
|
|
|
|
|
|
server.cond = sync.NewCond(&server.mu)
|
|
|
|
|
|
|
|
ctx, cancel := context.WithCancel(context.Background())
|
|
|
|
go server.run(ctx)
|
|
|
|
|
|
|
|
addr := "127.0.0.1:" + strconv.Itoa(*port)
|
|
|
|
listener, err := net.Listen("tcp", addr)
|
|
|
|
if err != nil {
|
|
|
|
fmt.Println("Listen error:", err)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
defer listener.Close()
|
|
|
|
|
|
|
|
mux := http.NewServeMux()
|
|
|
|
mux.HandleFunc("/embedding", server.embeddings)
|
|
|
|
mux.HandleFunc("/completion", server.completion)
|
|
|
|
mux.HandleFunc("/health", server.health)
|
|
|
|
|
|
|
|
httpServer := http.Server{
|
|
|
|
Handler: mux,
|
|
|
|
}
|
|
|
|
|
|
|
|
log.Println("Server listening on", addr)
|
|
|
|
if err := httpServer.Serve(listener); err != nil {
|
|
|
|
log.Fatal("server error:", err)
|
|
|
|
}
|
|
|
|
|
|
|
|
cancel()
|
|
|
|
}
|