2023-07-17 00:02:22 +00:00
|
|
|
package server
|
|
|
|
|
|
|
|
import (
|
|
|
|
"bytes"
|
2024-04-12 20:55:12 +00:00
|
|
|
"cmp"
|
2023-07-25 21:08:51 +00:00
|
|
|
"context"
|
2023-07-17 00:02:22 +00:00
|
|
|
"crypto/sha256"
|
2024-04-30 18:02:08 +00:00
|
|
|
"encoding/base64"
|
2023-08-29 03:50:24 +00:00
|
|
|
"encoding/hex"
|
2023-07-17 00:02:22 +00:00
|
|
|
"encoding/json"
|
|
|
|
"errors"
|
|
|
|
"fmt"
|
|
|
|
"io"
|
|
|
|
"log"
|
2024-01-18 18:52:01 +00:00
|
|
|
"log/slog"
|
server: allow mixed-case model names on push, pull, cp, and create (#7676)
This change allows for mixed-case model names to be pushed, pulled,
copied, and created, which was previously disallowed because the Ollama
registry was backed by a Docker registry that enforced a naming
convention that disallowed mixed-case names, which is no longer the
case.
This does not break existing, intended, behaviors.
Also, make TestCase test a story of creating, updating, pulling, and
copying a model with case variations, ensuring the model's manifest is
updated correctly, and not duplicated across different files with
different case variations.
2024-11-19 23:05:57 +00:00
|
|
|
"net"
|
2023-07-17 00:02:22 +00:00
|
|
|
"net/http"
|
2023-08-22 01:38:31 +00:00
|
|
|
"net/url"
|
2023-07-17 00:02:22 +00:00
|
|
|
"os"
|
|
|
|
"path/filepath"
|
2023-08-22 01:24:42 +00:00
|
|
|
"runtime"
|
2024-05-22 04:30:52 +00:00
|
|
|
"slices"
|
2024-02-14 19:29:49 +00:00
|
|
|
"strconv"
|
2023-07-17 00:02:22 +00:00
|
|
|
"strings"
|
|
|
|
|
2024-03-26 20:04:17 +00:00
|
|
|
"github.com/ollama/ollama/api"
|
2024-04-30 18:02:08 +00:00
|
|
|
"github.com/ollama/ollama/auth"
|
2024-06-04 18:53:23 +00:00
|
|
|
"github.com/ollama/ollama/envconfig"
|
2024-03-13 18:03:56 +00:00
|
|
|
"github.com/ollama/ollama/format"
|
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/llama"
|
2024-03-26 20:04:17 +00:00
|
|
|
"github.com/ollama/ollama/llm"
|
2024-05-20 18:26:45 +00:00
|
|
|
"github.com/ollama/ollama/parser"
|
2024-06-10 21:54:42 +00:00
|
|
|
"github.com/ollama/ollama/template"
|
2024-04-30 18:02:08 +00:00
|
|
|
"github.com/ollama/ollama/types/errtypes"
|
2024-04-16 23:22:38 +00:00
|
|
|
"github.com/ollama/ollama/types/model"
|
2024-03-26 20:04:17 +00:00
|
|
|
"github.com/ollama/ollama/version"
|
2023-07-17 00:02:22 +00:00
|
|
|
)
|
|
|
|
|
2024-06-21 02:13:36 +00:00
|
|
|
var (
|
|
|
|
errCapabilities = errors.New("does not support")
|
|
|
|
errCapabilityCompletion = errors.New("completion")
|
|
|
|
errCapabilityTools = errors.New("tools")
|
|
|
|
errCapabilityInsert = errors.New("insert")
|
|
|
|
)
|
2024-06-17 17:38:55 +00:00
|
|
|
|
2024-06-11 21:03:42 +00:00
|
|
|
type Capability string
|
|
|
|
|
2024-06-20 20:45:47 +00:00
|
|
|
const (
|
|
|
|
CapabilityCompletion = Capability("completion")
|
|
|
|
CapabilityTools = Capability("tools")
|
2024-06-21 02:13:36 +00:00
|
|
|
CapabilityInsert = Capability("insert")
|
2024-06-20 20:45:47 +00:00
|
|
|
)
|
2024-06-11 21:03:42 +00:00
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
type registryOptions struct {
|
|
|
|
Insecure bool
|
|
|
|
Username string
|
|
|
|
Password string
|
|
|
|
Token string
|
2024-07-25 22:58:30 +00:00
|
|
|
|
|
|
|
CheckRedirect func(req *http.Request, via []*http.Request) error
|
2024-02-14 19:29:49 +00:00
|
|
|
}
|
|
|
|
|
2023-07-17 00:02:22 +00:00
|
|
|
type Model struct {
|
2023-11-30 18:30:23 +00:00
|
|
|
Name string `json:"name"`
|
2023-12-01 19:37:17 +00:00
|
|
|
Config ConfigV2
|
2023-11-30 18:30:23 +00:00
|
|
|
ShortName string
|
|
|
|
ModelPath string
|
2024-01-25 20:12:36 +00:00
|
|
|
ParentModel string
|
2023-11-30 18:30:23 +00:00
|
|
|
AdapterPaths []string
|
|
|
|
ProjectorPaths []string
|
|
|
|
System string
|
|
|
|
License []string
|
|
|
|
Digest string
|
|
|
|
Options map[string]interface{}
|
2024-06-19 21:14:28 +00:00
|
|
|
Messages []api.Message
|
2024-06-10 21:54:42 +00:00
|
|
|
|
|
|
|
Template *template.Template
|
2024-01-25 20:12:36 +00:00
|
|
|
}
|
|
|
|
|
2024-06-17 17:38:55 +00:00
|
|
|
// CheckCapabilities checks if the model has the specified capabilities returning an error describing
|
|
|
|
// any missing or unknown capabilities
|
|
|
|
func (m *Model) CheckCapabilities(caps ...Capability) error {
|
|
|
|
var errs []error
|
2024-06-11 21:03:42 +00:00
|
|
|
for _, cap := range caps {
|
|
|
|
switch cap {
|
|
|
|
case CapabilityCompletion:
|
2024-06-14 21:57:49 +00:00
|
|
|
f, err := os.Open(m.ModelPath)
|
|
|
|
if err != nil {
|
|
|
|
slog.Error("couldn't open model file", "error", err)
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
defer f.Close()
|
|
|
|
|
|
|
|
// TODO(mxyng): decode the GGML into model to avoid doing this multiple times
|
|
|
|
ggml, _, err := llm.DecodeGGML(f, 0)
|
|
|
|
if err != nil {
|
|
|
|
slog.Error("couldn't decode ggml", "error", err)
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
|
2024-06-17 17:38:55 +00:00
|
|
|
errs = append(errs, errCapabilityCompletion)
|
2024-06-11 21:03:42 +00:00
|
|
|
}
|
2024-06-20 20:45:47 +00:00
|
|
|
case CapabilityTools:
|
|
|
|
if !slices.Contains(m.Template.Vars(), "tools") {
|
2024-06-21 02:13:36 +00:00
|
|
|
errs = append(errs, errCapabilityTools)
|
|
|
|
}
|
|
|
|
case CapabilityInsert:
|
|
|
|
vars := m.Template.Vars()
|
|
|
|
if !slices.Contains(vars, "suffix") {
|
|
|
|
errs = append(errs, errCapabilityInsert)
|
2024-06-20 20:45:47 +00:00
|
|
|
}
|
2024-06-11 21:03:42 +00:00
|
|
|
default:
|
|
|
|
slog.Error("unknown capability", "capability", cap)
|
2024-06-17 17:38:55 +00:00
|
|
|
return fmt.Errorf("unknown capability: %s", cap)
|
2024-06-11 21:03:42 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-06-17 17:38:55 +00:00
|
|
|
if err := errors.Join(errs...); err != nil {
|
2024-06-21 02:13:36 +00:00
|
|
|
return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
|
2024-06-17 17:38:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return nil
|
2024-02-21 02:37:29 +00:00
|
|
|
}
|
|
|
|
|
2024-04-30 17:55:19 +00:00
|
|
|
func (m *Model) String() string {
|
2024-05-20 18:26:45 +00:00
|
|
|
var modelfile parser.File
|
2024-04-30 17:55:19 +00:00
|
|
|
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-04-30 17:55:19 +00:00
|
|
|
Name: "model",
|
|
|
|
Args: m.ModelPath,
|
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
|
2024-05-08 19:42:48 +00:00
|
|
|
for _, adapter := range m.AdapterPaths {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-05-08 19:42:48 +00:00
|
|
|
Name: "adapter",
|
|
|
|
Args: adapter,
|
2024-04-30 17:55:19 +00:00
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
2024-05-08 19:42:48 +00:00
|
|
|
for _, projector := range m.ProjectorPaths {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-05-08 19:42:48 +00:00
|
|
|
Name: "model",
|
|
|
|
Args: projector,
|
2024-04-30 17:55:19 +00:00
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
2024-06-10 21:54:42 +00:00
|
|
|
if m.Template != nil {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-05-08 19:42:48 +00:00
|
|
|
Name: "template",
|
2024-06-10 21:54:42 +00:00
|
|
|
Args: m.Template.String(),
|
2024-04-30 17:55:19 +00:00
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
2024-05-08 19:42:48 +00:00
|
|
|
if m.System != "" {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-05-08 19:42:48 +00:00
|
|
|
Name: "system",
|
|
|
|
Args: m.System,
|
2024-04-30 17:55:19 +00:00
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for k, v := range m.Options {
|
|
|
|
switch v := v.(type) {
|
|
|
|
case []any:
|
|
|
|
for _, s := range v {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-04-30 17:55:19 +00:00
|
|
|
Name: k,
|
|
|
|
Args: fmt.Sprintf("%v", s),
|
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
default:
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-04-30 17:55:19 +00:00
|
|
|
Name: k,
|
|
|
|
Args: fmt.Sprintf("%v", v),
|
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
for _, license := range m.License {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-04-30 17:55:19 +00:00
|
|
|
Name: "license",
|
|
|
|
Args: license,
|
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for _, msg := range m.Messages {
|
2024-05-20 18:26:45 +00:00
|
|
|
modelfile.Commands = append(modelfile.Commands, parser.Command{
|
2024-04-30 17:55:19 +00:00
|
|
|
Name: "message",
|
2024-07-31 23:52:09 +00:00
|
|
|
Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content),
|
2024-04-30 17:55:19 +00:00
|
|
|
})
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
2024-04-30 17:55:19 +00:00
|
|
|
return modelfile.String()
|
2024-04-26 23:59:31 +00:00
|
|
|
}
|
|
|
|
|
2023-07-17 00:02:22 +00:00
|
|
|
type ConfigV2 struct {
|
2023-12-09 10:05:43 +00:00
|
|
|
ModelFormat string `json:"model_format"`
|
|
|
|
ModelFamily string `json:"model_family"`
|
|
|
|
ModelFamilies []string `json:"model_families"`
|
|
|
|
ModelType string `json:"model_type"`
|
|
|
|
FileType string `json:"file_type"`
|
|
|
|
|
2023-07-21 20:33:56 +00:00
|
|
|
// required by spec
|
2023-07-17 00:02:22 +00:00
|
|
|
Architecture string `json:"architecture"`
|
|
|
|
OS string `json:"os"`
|
2023-12-01 19:37:17 +00:00
|
|
|
RootFS RootFS `json:"rootfs"`
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
type RootFS struct {
|
|
|
|
Type string `json:"type"`
|
|
|
|
DiffIDs []string `json:"diff_ids"`
|
|
|
|
}
|
|
|
|
|
2024-06-10 15:47:13 +00:00
|
|
|
func GetManifest(mp ModelPath) (*Manifest, string, error) {
|
2023-10-27 14:19:59 +00:00
|
|
|
fp, err := mp.GetManifestPath()
|
2023-07-17 18:03:55 +00:00
|
|
|
if err != nil {
|
2023-08-29 03:50:24 +00:00
|
|
|
return nil, "", err
|
2023-07-17 18:03:55 +00:00
|
|
|
}
|
2023-07-17 21:21:27 +00:00
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
f, err := os.Open(fp)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2024-08-14 21:37:51 +00:00
|
|
|
return nil, "", err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2024-08-14 21:37:51 +00:00
|
|
|
defer f.Close()
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
sha256sum := sha256.New()
|
2023-08-29 03:50:24 +00:00
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
var manifest Manifest
|
|
|
|
if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
|
2023-08-29 03:50:24 +00:00
|
|
|
return nil, "", err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
func GetModel(name string) (*Model, error) {
|
2023-08-22 16:39:42 +00:00
|
|
|
mp := ParseModelPath(name)
|
2023-08-29 03:50:24 +00:00
|
|
|
manifest, digest, err := GetManifest(mp)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
model := &Model{
|
2023-10-19 14:39:58 +00:00
|
|
|
Name: mp.GetFullTagname(),
|
|
|
|
ShortName: mp.GetShortTagname(),
|
|
|
|
Digest: digest,
|
2024-06-10 21:54:42 +00:00
|
|
|
Template: template.DefaultTemplate,
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-08-06 00:13:52 +00:00
|
|
|
if manifest.Config.Digest != "" {
|
|
|
|
filename, err := GetBlobsPath(manifest.Config.Digest)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
2023-12-01 19:37:17 +00:00
|
|
|
|
2024-08-06 00:13:52 +00:00
|
|
|
configFile, err := os.Open(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
defer configFile.Close()
|
2023-12-01 19:37:17 +00:00
|
|
|
|
2024-08-06 00:13:52 +00:00
|
|
|
if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
2023-12-01 19:37:17 +00:00
|
|
|
}
|
|
|
|
|
2023-07-17 00:02:22 +00:00
|
|
|
for _, layer := range manifest.Layers {
|
2023-07-18 05:44:21 +00:00
|
|
|
filename, err := GetBlobsPath(layer.Digest)
|
2023-07-17 18:03:55 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
2023-07-17 00:02:22 +00:00
|
|
|
switch layer.MediaType {
|
|
|
|
case "application/vnd.ollama.image.model":
|
|
|
|
model.ModelPath = filename
|
2024-01-25 20:12:36 +00:00
|
|
|
model.ParentModel = layer.From
|
2023-08-04 22:56:40 +00:00
|
|
|
case "application/vnd.ollama.image.embed":
|
2023-10-16 15:07:37 +00:00
|
|
|
// Deprecated in versions > 0.1.2
|
|
|
|
// TODO: remove this warning in a future version
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
|
2023-08-04 00:16:05 +00:00
|
|
|
case "application/vnd.ollama.image.adapter":
|
|
|
|
model.AdapterPaths = append(model.AdapterPaths, filename)
|
2023-11-30 18:30:23 +00:00
|
|
|
case "application/vnd.ollama.image.projector":
|
|
|
|
model.ProjectorPaths = append(model.ProjectorPaths, filename)
|
2024-06-10 21:54:42 +00:00
|
|
|
case "application/vnd.ollama.image.prompt",
|
|
|
|
"application/vnd.ollama.image.template":
|
2023-07-17 21:21:27 +00:00
|
|
|
bts, err := os.ReadFile(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
2024-06-10 21:54:42 +00:00
|
|
|
model.Template, err = template.Parse(string(bts))
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
2024-06-10 21:54:42 +00:00
|
|
|
case "application/vnd.ollama.image.system":
|
2023-07-20 02:43:00 +00:00
|
|
|
bts, err := os.ReadFile(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
2024-06-10 21:54:42 +00:00
|
|
|
model.System = string(bts)
|
2023-07-17 00:02:22 +00:00
|
|
|
case "application/vnd.ollama.image.params":
|
2023-07-17 19:08:10 +00:00
|
|
|
params, err := os.Open(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
defer params.Close()
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-07-31 19:07:04 +00:00
|
|
|
// parse model options parameters into a map so that we can see which fields have been specified explicitly
|
2023-08-01 17:36:31 +00:00
|
|
|
if err = json.NewDecoder(params).Decode(&model.Options); err != nil {
|
2023-07-31 19:07:04 +00:00
|
|
|
return nil, err
|
|
|
|
}
|
2024-01-25 20:12:36 +00:00
|
|
|
case "application/vnd.ollama.image.messages":
|
|
|
|
msgs, err := os.Open(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
defer msgs.Close()
|
|
|
|
|
|
|
|
if err = json.NewDecoder(msgs).Decode(&model.Messages); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
2023-09-06 18:04:17 +00:00
|
|
|
case "application/vnd.ollama.image.license":
|
|
|
|
bts, err := os.ReadFile(filename)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
model.License = append(model.License, string(bts))
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return model, nil
|
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
func realpath(rel, from string) string {
|
2023-11-21 20:43:17 +00:00
|
|
|
abspath, err := filepath.Abs(from)
|
2023-11-14 20:30:34 +00:00
|
|
|
if err != nil {
|
2023-11-21 20:43:17 +00:00
|
|
|
return from
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
|
|
|
|
2023-11-14 20:30:34 +00:00
|
|
|
home, err := os.UserHomeDir()
|
2023-07-20 04:55:15 +00:00
|
|
|
if err != nil {
|
2023-11-14 20:30:34 +00:00
|
|
|
return abspath
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2023-11-21 20:43:17 +00:00
|
|
|
if from == "~" {
|
2023-11-14 20:30:34 +00:00
|
|
|
return home
|
2023-11-21 20:43:17 +00:00
|
|
|
} else if strings.HasPrefix(from, "~/") {
|
|
|
|
return filepath.Join(home, from[2:])
|
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
if _, err := os.Stat(filepath.Join(rel, from)); err == nil {
|
2023-11-21 20:43:17 +00:00
|
|
|
// this is a file relative to the Modelfile
|
2024-04-12 20:55:12 +00:00
|
|
|
return filepath.Join(rel, from)
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2023-11-14 20:30:34 +00:00
|
|
|
return abspath
|
|
|
|
}
|
|
|
|
|
2024-05-08 21:36:08 +00:00
|
|
|
func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantization string, modelfile *parser.File, fn func(resp api.ProgressResponse)) (err error) {
|
2023-07-21 20:33:56 +00:00
|
|
|
config := ConfigV2{
|
|
|
|
OS: "linux",
|
2023-11-14 20:30:34 +00:00
|
|
|
Architecture: "amd64",
|
2023-11-22 21:28:49 +00:00
|
|
|
RootFS: RootFS{
|
|
|
|
Type: "layers",
|
|
|
|
},
|
2023-07-21 20:33:56 +00:00
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
var messages []*api.Message
|
|
|
|
parameters := make(map[string]any)
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2024-08-07 21:22:17 +00:00
|
|
|
var layers []Layer
|
2024-08-23 18:29:56 +00:00
|
|
|
var baseLayers []*layerGGML
|
2024-04-30 17:55:19 +00:00
|
|
|
for _, c := range modelfile.Commands {
|
2023-11-14 20:30:34 +00:00
|
|
|
mediatype := fmt.Sprintf("application/vnd.ollama.image.%s", c.Name)
|
2024-08-23 18:29:56 +00:00
|
|
|
command := c.Name
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2024-08-23 18:29:56 +00:00
|
|
|
switch command {
|
2024-04-12 20:55:12 +00:00
|
|
|
case "model", "adapter":
|
2024-08-23 18:29:56 +00:00
|
|
|
if name := model.ParseName(c.Args); name.IsValid() && command == "model" {
|
2024-04-23 22:18:45 +00:00
|
|
|
baseLayers, err = parseFromModel(ctx, name, fn)
|
2023-11-15 18:59:38 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
} else if strings.HasPrefix(c.Args, "@") {
|
2024-05-10 22:48:41 +00:00
|
|
|
digest := strings.TrimPrefix(c.Args, "@")
|
2024-05-20 21:58:27 +00:00
|
|
|
if ib, ok := intermediateBlobs[digest]; ok {
|
|
|
|
p, err := GetBlobsPath(ib)
|
2024-05-10 22:48:41 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
if _, err := os.Stat(p); errors.Is(err, os.ErrNotExist) {
|
|
|
|
// pass
|
|
|
|
} else if err != nil {
|
|
|
|
return err
|
|
|
|
} else {
|
2024-05-20 21:58:27 +00:00
|
|
|
fn(api.ProgressResponse{Status: fmt.Sprintf("using cached layer %s", ib)})
|
|
|
|
digest = ib
|
2024-05-10 22:48:41 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
blobpath, err := GetBlobsPath(digest)
|
2024-04-12 20:55:12 +00:00
|
|
|
if err != nil {
|
2023-11-14 20:30:34 +00:00
|
|
|
return err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2023-07-21 20:33:56 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
blob, err := os.Open(blobpath)
|
2023-08-18 04:52:11 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
defer blob.Close()
|
2023-08-18 04:52:11 +00:00
|
|
|
|
2024-08-23 18:29:56 +00:00
|
|
|
baseLayers, err = parseFromFile(ctx, command, baseLayers, blob, digest, fn)
|
2023-08-18 04:52:11 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
|
|
|
|
defer file.Close()
|
2023-08-18 04:52:11 +00:00
|
|
|
|
2024-08-23 18:29:56 +00:00
|
|
|
baseLayers, err = parseFromFile(ctx, command, baseLayers, file, "", fn)
|
2024-04-12 20:55:12 +00:00
|
|
|
if err != nil {
|
2023-08-18 04:52:11 +00:00
|
|
|
return err
|
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
} else {
|
|
|
|
return fmt.Errorf("invalid model reference: %s", c.Args)
|
|
|
|
}
|
2023-08-18 04:52:11 +00:00
|
|
|
|
2024-04-25 15:53:08 +00:00
|
|
|
for _, baseLayer := range baseLayers {
|
2024-04-25 16:01:20 +00:00
|
|
|
if quantization != "" &&
|
|
|
|
baseLayer.MediaType == "application/vnd.ollama.image.model" &&
|
|
|
|
baseLayer.GGML != nil &&
|
|
|
|
baseLayer.GGML.Name() == "gguf" {
|
2024-05-08 00:44:03 +00:00
|
|
|
want, err := llm.ParseFileType(quantization)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2024-04-25 15:53:08 +00:00
|
|
|
return err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2024-05-08 00:44:03 +00:00
|
|
|
ft := baseLayer.GGML.KV().FileType()
|
|
|
|
if !slices.Contains([]string{"F16", "F32"}, ft.String()) {
|
2024-04-25 15:53:08 +00:00
|
|
|
return errors.New("quantization is only supported for F16 and F32 models")
|
2024-05-08 00:44:03 +00:00
|
|
|
} else if want != ft {
|
|
|
|
fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", ft, quantization)})
|
|
|
|
|
|
|
|
blob, err := GetBlobsPath(baseLayer.Digest)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
temp, err := os.CreateTemp(filepath.Dir(blob), quantization)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
defer temp.Close()
|
|
|
|
defer os.Remove(temp.Name())
|
|
|
|
|
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 := llama.Quantize(blob, temp.Name(), uint32(want)); err != nil {
|
2024-05-08 00:44:03 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-06-21 20:30:43 +00:00
|
|
|
layer, err := NewLayer(temp, baseLayer.MediaType)
|
2024-05-08 00:44:03 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-05-17 18:29:04 +00:00
|
|
|
|
2024-06-21 20:30:43 +00:00
|
|
|
if _, err := temp.Seek(0, io.SeekStart); err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-06-25 04:47:52 +00:00
|
|
|
ggml, _, err := llm.DecodeGGML(temp, 0)
|
2024-06-21 20:30:43 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
2024-05-17 18:29:04 +00:00
|
|
|
}
|
|
|
|
|
2024-06-21 20:30:43 +00:00
|
|
|
baseLayer.Layer = layer
|
|
|
|
baseLayer.GGML = ggml
|
2024-04-12 20:55:12 +00:00
|
|
|
}
|
2023-11-24 19:57:20 +00:00
|
|
|
}
|
2023-08-04 00:16:05 +00:00
|
|
|
|
2024-04-25 15:53:08 +00:00
|
|
|
if baseLayer.GGML != nil {
|
|
|
|
config.ModelFormat = cmp.Or(config.ModelFormat, baseLayer.GGML.Name())
|
|
|
|
config.ModelFamily = cmp.Or(config.ModelFamily, baseLayer.GGML.KV().Architecture())
|
|
|
|
config.ModelType = cmp.Or(config.ModelType, format.HumanNumber(baseLayer.GGML.KV().ParameterCount()))
|
2024-05-08 00:44:03 +00:00
|
|
|
config.FileType = cmp.Or(config.FileType, baseLayer.GGML.KV().FileType().String())
|
2024-04-25 15:53:08 +00:00
|
|
|
config.ModelFamilies = append(config.ModelFamilies, baseLayer.GGML.KV().Architecture())
|
2023-11-24 19:57:20 +00:00
|
|
|
}
|
|
|
|
|
2024-04-25 15:53:08 +00:00
|
|
|
layers = append(layers, baseLayer.Layer)
|
2023-12-01 18:50:55 +00:00
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
case "license", "template", "system":
|
2024-07-19 22:24:29 +00:00
|
|
|
if c.Name == "template" {
|
|
|
|
if _, err := template.Parse(c.Args); err != nil {
|
|
|
|
return fmt.Errorf("%w: %s", errBadTemplate, err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
if c.Name != "license" {
|
|
|
|
// replace
|
2024-08-07 21:22:17 +00:00
|
|
|
layers = slices.DeleteFunc(layers, func(layer Layer) bool {
|
2024-05-08 21:36:08 +00:00
|
|
|
if layer.MediaType != mediatype {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
if err := layer.Remove(); err != nil {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
return true
|
2024-04-12 20:55:12 +00:00
|
|
|
})
|
2024-03-08 23:38:53 +00:00
|
|
|
}
|
|
|
|
|
2024-05-08 21:36:08 +00:00
|
|
|
blob := strings.NewReader(c.Args)
|
|
|
|
layer, err := NewLayer(blob, mediatype)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
layers = append(layers, layer)
|
|
|
|
case "message":
|
|
|
|
role, content, ok := strings.Cut(c.Args, ": ")
|
|
|
|
if !ok {
|
|
|
|
return fmt.Errorf("invalid message: %s", c.Args)
|
|
|
|
}
|
|
|
|
|
|
|
|
messages = append(messages, &api.Message{Role: role, Content: content})
|
|
|
|
default:
|
|
|
|
ps, err := api.FormatParams(map[string][]string{c.Name: {c.Args}})
|
2023-08-04 00:16:05 +00:00
|
|
|
if err != nil {
|
2023-11-14 20:30:34 +00:00
|
|
|
return err
|
2023-08-04 00:16:05 +00:00
|
|
|
}
|
2023-08-08 20:56:48 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
for k, v := range ps {
|
|
|
|
if ks, ok := parameters[k].([]string); ok {
|
|
|
|
parameters[k] = append(ks, v.([]string)...)
|
|
|
|
} else if vs, ok := v.([]string); ok {
|
|
|
|
parameters[k] = vs
|
|
|
|
} else {
|
|
|
|
parameters[k] = v
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2023-11-22 21:28:49 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
var err2 error
|
2024-08-07 21:22:17 +00:00
|
|
|
layers = slices.DeleteFunc(layers, func(layer Layer) bool {
|
2024-04-12 20:55:12 +00:00
|
|
|
switch layer.MediaType {
|
|
|
|
case "application/vnd.ollama.image.message":
|
|
|
|
// if there are new messages, remove the inherited ones
|
|
|
|
if len(messages) > 0 {
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
return false
|
|
|
|
case "application/vnd.ollama.image.params":
|
|
|
|
// merge inherited parameters with new ones
|
|
|
|
r, err := layer.Open()
|
2023-08-08 20:56:48 +00:00
|
|
|
if err != nil {
|
2024-04-12 20:55:12 +00:00
|
|
|
err2 = err
|
|
|
|
return false
|
2023-08-08 20:56:48 +00:00
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
defer r.Close()
|
2023-08-08 20:56:48 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
var ps map[string]any
|
|
|
|
if err := json.NewDecoder(r).Decode(&ps); err != nil {
|
|
|
|
err2 = err
|
|
|
|
return false
|
|
|
|
}
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
for k, v := range ps {
|
|
|
|
if _, ok := parameters[k]; !ok {
|
|
|
|
parameters[k] = v
|
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2023-07-17 21:21:27 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
return true
|
2023-07-17 00:02:22 +00:00
|
|
|
default:
|
2024-04-12 20:55:12 +00:00
|
|
|
return false
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
})
|
|
|
|
|
|
|
|
if err2 != nil {
|
|
|
|
return err2
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-01-25 20:12:36 +00:00
|
|
|
if len(messages) > 0 {
|
|
|
|
var b bytes.Buffer
|
2024-04-12 20:55:12 +00:00
|
|
|
if err := json.NewEncoder(&b).Encode(messages); err != nil {
|
2024-01-25 20:12:36 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
layer, err := NewLayer(&b, "application/vnd.ollama.image.messages")
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
layers = append(layers, layer)
|
2024-01-25 20:12:36 +00:00
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
if len(parameters) > 0 {
|
2023-11-14 20:30:34 +00:00
|
|
|
var b bytes.Buffer
|
2024-04-12 20:55:12 +00:00
|
|
|
if err := json.NewEncoder(&b).Encode(parameters); err != nil {
|
2023-08-04 22:56:40 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2023-11-22 21:28:49 +00:00
|
|
|
layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2023-11-14 20:30:34 +00:00
|
|
|
return err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
layers = append(layers, layer)
|
2023-08-04 22:56:40 +00:00
|
|
|
}
|
|
|
|
|
2024-04-12 20:55:12 +00:00
|
|
|
digests := make([]string, len(layers))
|
|
|
|
for i, layer := range layers {
|
2023-11-22 21:28:49 +00:00
|
|
|
digests[i] = layer.Digest
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2023-11-22 21:28:49 +00:00
|
|
|
config.RootFS.DiffIDs = digests
|
2023-11-14 20:30:34 +00:00
|
|
|
|
2023-11-22 21:28:49 +00:00
|
|
|
var b bytes.Buffer
|
|
|
|
if err := json.NewEncoder(&b).Encode(config); err != nil {
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-08-07 18:44:25 +00:00
|
|
|
configLayer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
|
2023-11-22 21:28:49 +00:00
|
|
|
if err != nil {
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-08-07 18:44:25 +00:00
|
|
|
for _, layer := range append(layers, configLayer) {
|
2024-04-12 20:55:12 +00:00
|
|
|
if layer.status != "" {
|
|
|
|
fn(api.ProgressResponse{Status: layer.status})
|
2023-07-17 18:03:55 +00:00
|
|
|
}
|
2024-04-12 20:55:12 +00:00
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2024-05-08 21:36:08 +00:00
|
|
|
old, _ := ParseNamedManifest(name)
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-11-22 21:28:49 +00:00
|
|
|
fn(api.ProgressResponse{Status: "writing manifest"})
|
2024-08-07 18:44:25 +00:00
|
|
|
if err := WriteManifest(name, configLayer, layers); err != nil {
|
2023-10-27 14:19:59 +00:00
|
|
|
return err
|
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2024-07-04 00:22:13 +00:00
|
|
|
if !envconfig.NoPrune() && old != nil {
|
2024-05-08 21:36:08 +00:00
|
|
|
if err := old.RemoveLayers(); err != nil {
|
2023-11-22 21:28:49 +00:00
|
|
|
return err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-11-22 21:28:49 +00:00
|
|
|
fn(api.ProgressResponse{Status: "success"})
|
|
|
|
return nil
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-04-16 23:22:38 +00:00
|
|
|
func CopyModel(src, dst model.Name) error {
|
2024-04-26 20:08:32 +00:00
|
|
|
if !dst.IsFullyQualified() {
|
|
|
|
return model.Unqualified(dst)
|
|
|
|
}
|
|
|
|
if !src.IsFullyQualified() {
|
|
|
|
return model.Unqualified(src)
|
|
|
|
}
|
|
|
|
|
2024-04-29 03:47:49 +00:00
|
|
|
if src.Filepath() == dst.Filepath() {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2024-04-16 23:22:38 +00:00
|
|
|
manifests, err := GetManifestPath()
|
2023-08-22 04:56:56 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-04-26 20:08:32 +00:00
|
|
|
dstpath := filepath.Join(manifests, dst.Filepath())
|
2024-04-16 23:22:38 +00:00
|
|
|
if err := os.MkdirAll(filepath.Dir(dstpath), 0o755); err != nil {
|
2023-10-27 14:19:59 +00:00
|
|
|
return err
|
|
|
|
}
|
2023-07-24 15:27:28 +00:00
|
|
|
|
2024-04-26 20:08:32 +00:00
|
|
|
srcpath := filepath.Join(manifests, src.Filepath())
|
2024-04-16 23:22:38 +00:00
|
|
|
srcfile, err := os.Open(srcpath)
|
2023-07-24 15:27:28 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-04-16 23:22:38 +00:00
|
|
|
defer srcfile.Close()
|
2023-07-24 15:27:28 +00:00
|
|
|
|
2024-04-16 23:22:38 +00:00
|
|
|
dstfile, err := os.Create(dstpath)
|
2023-07-24 15:27:28 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2024-04-16 23:22:38 +00:00
|
|
|
defer dstfile.Close()
|
2023-07-24 15:27:28 +00:00
|
|
|
|
2024-04-16 23:22:38 +00:00
|
|
|
_, err = io.Copy(dstfile, srcfile)
|
|
|
|
return err
|
2023-07-24 15:27:28 +00:00
|
|
|
}
|
|
|
|
|
2024-08-14 23:36:07 +00:00
|
|
|
func deleteUnusedLayers(deleteMap map[string]struct{}) error {
|
2024-11-05 22:21:45 +00:00
|
|
|
// Ignore corrupt manifests to avoid blocking deletion of layers that are freshly orphaned
|
|
|
|
manifests, err := Manifests(true)
|
2023-07-20 23:09:23 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2023-08-30 18:31:12 +00:00
|
|
|
|
2024-08-14 23:36:07 +00:00
|
|
|
for _, manifest := range manifests {
|
2023-08-30 18:31:12 +00:00
|
|
|
for _, layer := range manifest.Layers {
|
|
|
|
delete(deleteMap, layer.Digest)
|
|
|
|
}
|
|
|
|
|
|
|
|
delete(deleteMap, manifest.Config.Digest)
|
2023-07-31 22:26:18 +00:00
|
|
|
}
|
2023-07-20 23:09:23 +00:00
|
|
|
|
|
|
|
// only delete the files which are still in the deleteMap
|
2023-11-14 20:30:34 +00:00
|
|
|
for k := range deleteMap {
|
|
|
|
fp, err := GetBlobsPath(k)
|
|
|
|
if err != nil {
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("couldn't get file path for '%s': %v", k, err))
|
2023-11-14 20:30:34 +00:00
|
|
|
continue
|
|
|
|
}
|
2024-05-09 23:35:20 +00:00
|
|
|
if err := os.Remove(fp); err != nil {
|
|
|
|
slog.Info(fmt.Sprintf("couldn't remove file '%s': %v", fp, err))
|
|
|
|
continue
|
2023-07-20 23:09:23 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-09-11 18:46:35 +00:00
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func PruneLayers() error {
|
2023-11-14 20:30:34 +00:00
|
|
|
deleteMap := make(map[string]struct{})
|
2023-09-11 18:46:35 +00:00
|
|
|
p, err := GetBlobsPath("")
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
blobs, err := os.ReadDir(p)
|
|
|
|
if err != nil {
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("couldn't read dir '%s': %v", p, err))
|
2023-09-11 18:46:35 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
for _, blob := range blobs {
|
|
|
|
name := blob.Name()
|
2024-03-15 03:18:06 +00:00
|
|
|
name = strings.ReplaceAll(name, "-", ":")
|
2024-05-09 23:35:20 +00:00
|
|
|
|
|
|
|
_, err := GetBlobsPath(name)
|
|
|
|
if err != nil {
|
|
|
|
if errors.Is(err, ErrInvalidDigestFormat) {
|
|
|
|
// remove invalid blobs (e.g. partial downloads)
|
|
|
|
if err := os.Remove(filepath.Join(p, blob.Name())); err != nil {
|
|
|
|
slog.Error("couldn't remove blob", "blob", blob.Name(), "error", err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
continue
|
2023-11-14 22:27:51 +00:00
|
|
|
}
|
2024-05-09 23:35:20 +00:00
|
|
|
|
|
|
|
deleteMap[name] = struct{}{}
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
|
|
|
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("total blobs: %d", len(deleteMap)))
|
2023-09-11 18:46:35 +00:00
|
|
|
|
2024-08-14 23:36:07 +00:00
|
|
|
if err := deleteUnusedLayers(deleteMap); err != nil {
|
2024-08-06 00:13:52 +00:00
|
|
|
slog.Error(fmt.Sprintf("couldn't remove unused layers: %v", err))
|
2024-08-01 22:05:16 +00:00
|
|
|
return nil
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
|
|
|
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("total unused blobs removed: %d", len(deleteMap)))
|
2023-09-11 18:46:35 +00:00
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2023-09-27 00:28:14 +00:00
|
|
|
func PruneDirectory(path string) error {
|
|
|
|
info, err := os.Lstat(path)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
|
|
|
|
entries, err := os.ReadDir(path)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
for _, entry := range entries {
|
|
|
|
if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
entries, err = os.ReadDir(path)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
if len(entries) > 0 {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
return os.Remove(path)
|
|
|
|
}
|
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
|
2023-08-22 16:39:42 +00:00
|
|
|
mp := ParseModelPath(name)
|
2023-07-19 01:51:30 +00:00
|
|
|
fn(api.ProgressResponse{Status: "retrieving manifest"})
|
|
|
|
|
2023-08-22 16:39:42 +00:00
|
|
|
if mp.ProtocolScheme == "http" && !regOpts.Insecure {
|
2024-08-01 21:52:15 +00:00
|
|
|
return errors.New("insecure protocol http")
|
2023-08-22 16:39:42 +00:00
|
|
|
}
|
|
|
|
|
2023-08-29 03:50:24 +00:00
|
|
|
manifest, _, err := GetManifest(mp)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2023-07-19 01:51:30 +00:00
|
|
|
fn(api.ProgressResponse{Status: "couldn't retrieve manifest"})
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-08-07 21:22:17 +00:00
|
|
|
var layers []Layer
|
2023-08-01 01:37:40 +00:00
|
|
|
layers = append(layers, manifest.Layers...)
|
2024-08-06 00:13:52 +00:00
|
|
|
if manifest.Config.Digest != "" {
|
2024-08-07 21:22:17 +00:00
|
|
|
layers = append(layers, manifest.Config)
|
2024-08-06 00:13:52 +00:00
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
|
|
|
for _, layer := range layers {
|
2023-10-09 17:24:27 +00:00
|
|
|
if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("error uploading blob: %v", err))
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
2023-07-19 01:51:30 +00:00
|
|
|
}
|
|
|
|
|
2023-07-23 00:31:26 +00:00
|
|
|
fn(api.ProgressResponse{Status: "pushing manifest"})
|
2023-08-22 01:38:31 +00:00
|
|
|
requestURL := mp.BaseURL()
|
|
|
|
requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
|
2023-07-17 00:02:22 +00:00
|
|
|
|
|
|
|
manifestJSON, err := json.Marshal(manifest)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2023-08-22 01:24:42 +00:00
|
|
|
headers := make(http.Header)
|
|
|
|
headers.Set("Content-Type", "application/vnd.docker.distribution.manifest.v2+json")
|
2023-11-02 20:10:58 +00:00
|
|
|
resp, err := makeRequestWithRetry(ctx, http.MethodPut, requestURL, headers, bytes.NewReader(manifestJSON), regOpts)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
defer resp.Body.Close()
|
|
|
|
|
2023-07-23 00:31:26 +00:00
|
|
|
fn(api.ProgressResponse{Status: "success"})
|
2023-07-17 00:02:22 +00:00
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
|
2023-08-22 16:39:42 +00:00
|
|
|
mp := ParseModelPath(name)
|
|
|
|
|
2023-09-11 18:46:35 +00:00
|
|
|
// build deleteMap to prune unused layers
|
2023-11-14 20:30:34 +00:00
|
|
|
deleteMap := make(map[string]struct{})
|
2024-08-14 21:37:51 +00:00
|
|
|
manifest, _, err := GetManifest(mp)
|
|
|
|
if errors.Is(err, os.ErrNotExist) {
|
|
|
|
// noop
|
2024-11-05 22:21:45 +00:00
|
|
|
} else if err != nil {
|
|
|
|
slog.Warn("pulling model with bad existing manifest", "name", name, "error", err)
|
2024-08-14 21:37:51 +00:00
|
|
|
} else {
|
|
|
|
for _, l := range manifest.Layers {
|
|
|
|
deleteMap[l.Digest] = struct{}{}
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
2024-08-14 21:37:51 +00:00
|
|
|
if manifest.Config.Digest != "" {
|
|
|
|
deleteMap[manifest.Config.Digest] = struct{}{}
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-08-22 16:39:42 +00:00
|
|
|
if mp.ProtocolScheme == "http" && !regOpts.Insecure {
|
2024-08-01 21:52:15 +00:00
|
|
|
return errors.New("insecure protocol http")
|
2023-08-22 04:56:56 +00:00
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-07-19 01:51:30 +00:00
|
|
|
fn(api.ProgressResponse{Status: "pulling manifest"})
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-09-11 18:46:35 +00:00
|
|
|
manifest, err = pullModelManifest(ctx, mp, regOpts)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2023-07-24 21:48:17 +00:00
|
|
|
return fmt.Errorf("pull model manifest: %s", err)
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-08-07 21:22:17 +00:00
|
|
|
var layers []Layer
|
2023-07-20 18:18:00 +00:00
|
|
|
layers = append(layers, manifest.Layers...)
|
2024-08-06 00:13:52 +00:00
|
|
|
if manifest.Config.Digest != "" {
|
2024-08-07 21:22:17 +00:00
|
|
|
layers = append(layers, manifest.Config)
|
2024-08-06 00:13:52 +00:00
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2024-05-24 15:40:40 +00:00
|
|
|
skipVerify := make(map[string]bool)
|
2023-07-17 00:02:22 +00:00
|
|
|
for _, layer := range layers {
|
2024-05-24 15:40:40 +00:00
|
|
|
cacheHit, err := downloadBlob(ctx, downloadOpts{
|
|
|
|
mp: mp,
|
|
|
|
digest: layer.Digest,
|
|
|
|
regOpts: regOpts,
|
|
|
|
fn: fn,
|
|
|
|
})
|
|
|
|
if err != nil {
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
2024-05-24 15:40:40 +00:00
|
|
|
skipVerify[layer.Digest] = cacheHit
|
2023-09-11 18:46:35 +00:00
|
|
|
delete(deleteMap, layer.Digest)
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
2023-09-11 18:46:35 +00:00
|
|
|
delete(deleteMap, manifest.Config.Digest)
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-07-20 18:44:05 +00:00
|
|
|
fn(api.ProgressResponse{Status: "verifying sha256 digest"})
|
|
|
|
for _, layer := range layers {
|
2024-05-24 15:40:40 +00:00
|
|
|
if skipVerify[layer.Digest] {
|
|
|
|
continue
|
|
|
|
}
|
2023-07-20 18:44:05 +00:00
|
|
|
if err := verifyBlob(layer.Digest); err != nil {
|
2023-07-24 18:53:01 +00:00
|
|
|
if errors.Is(err, errDigestMismatch) {
|
|
|
|
// something went wrong, delete the blob
|
|
|
|
fp, err := GetBlobsPath(layer.Digest)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
if err := os.Remove(fp); err != nil {
|
|
|
|
// log this, but return the original error
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("couldn't remove file with digest mismatch '%s': %v", fp, err))
|
2023-07-24 18:53:01 +00:00
|
|
|
}
|
|
|
|
}
|
2023-07-20 18:44:05 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-19 01:51:30 +00:00
|
|
|
fn(api.ProgressResponse{Status: "writing manifest"})
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-07-17 18:03:55 +00:00
|
|
|
manifestJSON, err := json.Marshal(manifest)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2023-10-27 14:19:59 +00:00
|
|
|
fp, err := mp.GetManifestPath()
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2023-10-27 14:19:59 +00:00
|
|
|
if err := os.MkdirAll(filepath.Dir(fp), 0o755); err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-07-20 18:18:00 +00:00
|
|
|
err = os.WriteFile(fp, manifestJSON, 0o644)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("couldn't write to %s", fp))
|
2023-07-17 00:02:22 +00:00
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
if !envconfig.NoPrune() && len(deleteMap) > 0 {
|
|
|
|
fn(api.ProgressResponse{Status: "removing unused layers"})
|
2024-08-14 23:36:07 +00:00
|
|
|
if err := deleteUnusedLayers(deleteMap); err != nil {
|
2024-08-01 22:05:16 +00:00
|
|
|
fn(api.ProgressResponse{Status: fmt.Sprintf("couldn't remove unused layers: %v", err)})
|
2023-09-11 18:46:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-19 01:51:30 +00:00
|
|
|
fn(api.ProgressResponse{Status: "success"})
|
2023-07-17 00:02:22 +00:00
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2024-06-10 15:47:13 +00:00
|
|
|
func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
|
2023-08-22 01:38:31 +00:00
|
|
|
requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
|
2023-07-17 00:02:22 +00:00
|
|
|
|
2023-08-22 01:24:42 +00:00
|
|
|
headers := make(http.Header)
|
|
|
|
headers.Set("Accept", "application/vnd.docker.distribution.manifest.v2+json")
|
2023-11-02 20:13:32 +00:00
|
|
|
resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, regOpts)
|
2023-07-17 00:02:22 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
defer resp.Body.Close()
|
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
var m Manifest
|
2023-07-17 00:02:22 +00:00
|
|
|
if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
2024-08-14 21:37:51 +00:00
|
|
|
return &m, err
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// GetSHA256Digest returns the SHA256 hash of a given buffer and returns it, and the size of buffer
|
2023-09-28 17:00:34 +00:00
|
|
|
func GetSHA256Digest(r io.Reader) (string, int64) {
|
2023-07-19 00:14:12 +00:00
|
|
|
h := sha256.New()
|
|
|
|
n, err := io.Copy(h, r)
|
|
|
|
if err != nil {
|
|
|
|
log.Fatal(err)
|
|
|
|
}
|
|
|
|
|
2023-09-28 17:00:34 +00:00
|
|
|
return fmt.Sprintf("sha256:%x", h.Sum(nil)), n
|
2023-07-17 00:02:22 +00:00
|
|
|
}
|
|
|
|
|
2024-08-01 21:52:15 +00:00
|
|
|
var errUnauthorized = errors.New("unauthorized: access denied")
|
2024-04-30 18:02:08 +00:00
|
|
|
|
|
|
|
// getTokenSubject returns the subject of a JWT token, it does not validate the token
|
|
|
|
func getTokenSubject(token string) string {
|
|
|
|
parts := strings.Split(token, ".")
|
|
|
|
if len(parts) != 3 {
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
|
|
|
|
payload := parts[1]
|
|
|
|
payloadBytes, err := base64.RawURLEncoding.DecodeString(payload)
|
|
|
|
if err != nil {
|
|
|
|
slog.Error(fmt.Sprintf("failed to decode jwt payload: %v", err))
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
|
|
|
|
var payloadMap map[string]interface{}
|
|
|
|
if err := json.Unmarshal(payloadBytes, &payloadMap); err != nil {
|
|
|
|
slog.Error(fmt.Sprintf("failed to unmarshal payload JSON: %v", err))
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
|
|
|
|
sub, ok := payloadMap["sub"]
|
|
|
|
if !ok {
|
|
|
|
slog.Error("jwt does not contain 'sub' field")
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
|
|
|
|
return fmt.Sprintf("%s", sub)
|
|
|
|
}
|
2023-11-16 21:44:18 +00:00
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.ReadSeeker, regOpts *registryOptions) (*http.Response, error) {
|
2024-04-30 18:02:08 +00:00
|
|
|
anonymous := true // access will default to anonymous if no user is found associated with the public key
|
2024-05-22 05:21:04 +00:00
|
|
|
for range 2 {
|
2024-02-14 19:29:49 +00:00
|
|
|
resp, err := makeRequest(ctx, method, requestURL, headers, body, regOpts)
|
2023-08-17 19:35:29 +00:00
|
|
|
if err != nil {
|
2024-01-12 21:32:24 +00:00
|
|
|
if !errors.Is(err, context.Canceled) {
|
2024-01-18 18:52:01 +00:00
|
|
|
slog.Info(fmt.Sprintf("request failed: %v", err))
|
2024-01-12 21:32:24 +00:00
|
|
|
}
|
|
|
|
|
2023-08-17 19:35:29 +00:00
|
|
|
return nil, err
|
|
|
|
}
|
2024-01-12 21:32:24 +00:00
|
|
|
|
|
|
|
switch {
|
|
|
|
case resp.StatusCode == http.StatusUnauthorized:
|
2024-09-26 19:00:31 +00:00
|
|
|
resp.Body.Close()
|
|
|
|
|
2024-01-12 21:32:24 +00:00
|
|
|
// Handle authentication error with one retry
|
2024-02-14 19:29:49 +00:00
|
|
|
challenge := parseRegistryChallenge(resp.Header.Get("www-authenticate"))
|
|
|
|
token, err := getAuthorizationToken(ctx, challenge)
|
2023-08-17 19:35:29 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
2024-04-30 18:02:08 +00:00
|
|
|
anonymous = getTokenSubject(token) == "anonymous"
|
2024-01-12 21:32:24 +00:00
|
|
|
regOpts.Token = token
|
|
|
|
if body != nil {
|
|
|
|
_, err = body.Seek(0, io.SeekStart)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
}
|
|
|
|
case resp.StatusCode == http.StatusNotFound:
|
2024-09-26 19:00:31 +00:00
|
|
|
resp.Body.Close()
|
2024-01-12 21:32:24 +00:00
|
|
|
return nil, os.ErrNotExist
|
|
|
|
case resp.StatusCode >= http.StatusBadRequest:
|
2024-09-26 19:00:31 +00:00
|
|
|
defer resp.Body.Close()
|
2024-01-12 21:32:24 +00:00
|
|
|
responseBody, err := io.ReadAll(resp.Body)
|
|
|
|
if err != nil {
|
|
|
|
return nil, fmt.Errorf("%d: %s", resp.StatusCode, err)
|
|
|
|
}
|
|
|
|
return nil, fmt.Errorf("%d: %s", resp.StatusCode, responseBody)
|
|
|
|
default:
|
|
|
|
return resp, nil
|
2023-08-17 19:35:29 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-04-30 18:02:08 +00:00
|
|
|
if anonymous {
|
|
|
|
// no user is associated with the public key, and the request requires non-anonymous access
|
|
|
|
pubKey, nestedErr := auth.GetPublicKey()
|
|
|
|
if nestedErr != nil {
|
|
|
|
slog.Error(fmt.Sprintf("couldn't get public key: %v", nestedErr))
|
|
|
|
return nil, errUnauthorized
|
|
|
|
}
|
|
|
|
return nil, &errtypes.UnknownOllamaKey{Key: pubKey}
|
|
|
|
}
|
|
|
|
// user is associated with the public key, but is not authorized to make the request
|
2024-01-12 21:32:24 +00:00
|
|
|
return nil, errUnauthorized
|
2023-08-17 19:35:29 +00:00
|
|
|
}
|
|
|
|
|
server: allow mixed-case model names on push, pull, cp, and create (#7676)
This change allows for mixed-case model names to be pushed, pulled,
copied, and created, which was previously disallowed because the Ollama
registry was backed by a Docker registry that enforced a naming
convention that disallowed mixed-case names, which is no longer the
case.
This does not break existing, intended, behaviors.
Also, make TestCase test a story of creating, updating, pulling, and
copying a model with case variations, ensuring the model's manifest is
updated correctly, and not duplicated across different files with
different case variations.
2024-11-19 23:05:57 +00:00
|
|
|
// testMakeRequestDialContext specifies the dial function for the http client in
|
|
|
|
// makeRequest. It can be used to resolve hosts in model names to local
|
|
|
|
// addresses for testing. For example, the model name ("example.com/my/model")
|
|
|
|
// can be directed to push/pull from "127.0.0.1:1234".
|
|
|
|
//
|
|
|
|
// This is not safe to set across goroutines. It should be set in
|
|
|
|
// the main test goroutine, and not by tests marked to run in parallel with
|
|
|
|
// t.Parallel().
|
|
|
|
//
|
|
|
|
// It should be cleared after use, otherwise it will affect other tests.
|
|
|
|
//
|
|
|
|
// Ideally we would have some set this up the stack, but the code is not
|
|
|
|
// structured in a way that makes this easy, so this will have to do for now.
|
|
|
|
var testMakeRequestDialContext func(ctx context.Context, network, addr string) (net.Conn, error)
|
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
func makeRequest(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.Reader, regOpts *registryOptions) (*http.Response, error) {
|
|
|
|
if requestURL.Scheme != "http" && regOpts != nil && regOpts.Insecure {
|
|
|
|
requestURL.Scheme = "http"
|
|
|
|
}
|
|
|
|
|
|
|
|
req, err := http.NewRequestWithContext(ctx, method, requestURL.String(), body)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
if headers != nil {
|
|
|
|
req.Header = headers
|
|
|
|
}
|
|
|
|
|
|
|
|
if regOpts != nil {
|
|
|
|
if regOpts.Token != "" {
|
|
|
|
req.Header.Set("Authorization", "Bearer "+regOpts.Token)
|
|
|
|
} else if regOpts.Username != "" && regOpts.Password != "" {
|
|
|
|
req.SetBasicAuth(regOpts.Username, regOpts.Password)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-09-05 21:26:00 +00:00
|
|
|
req.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
|
2024-02-14 19:29:49 +00:00
|
|
|
|
|
|
|
if s := req.Header.Get("Content-Length"); s != "" {
|
|
|
|
contentLength, err := strconv.ParseInt(s, 10, 64)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
req.ContentLength = contentLength
|
|
|
|
}
|
|
|
|
|
2024-07-25 22:58:30 +00:00
|
|
|
resp, err := (&http.Client{
|
server: allow mixed-case model names on push, pull, cp, and create (#7676)
This change allows for mixed-case model names to be pushed, pulled,
copied, and created, which was previously disallowed because the Ollama
registry was backed by a Docker registry that enforced a naming
convention that disallowed mixed-case names, which is no longer the
case.
This does not break existing, intended, behaviors.
Also, make TestCase test a story of creating, updating, pulling, and
copying a model with case variations, ensuring the model's manifest is
updated correctly, and not duplicated across different files with
different case variations.
2024-11-19 23:05:57 +00:00
|
|
|
Transport: &http.Transport{
|
|
|
|
DialContext: testMakeRequestDialContext,
|
|
|
|
},
|
2024-07-25 22:58:30 +00:00
|
|
|
CheckRedirect: regOpts.CheckRedirect,
|
|
|
|
}).Do(req)
|
2024-02-14 19:29:49 +00:00
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
return resp, nil
|
|
|
|
}
|
|
|
|
|
2023-08-10 18:34:25 +00:00
|
|
|
func getValue(header, key string) string {
|
|
|
|
startIdx := strings.Index(header, key+"=")
|
|
|
|
if startIdx == -1 {
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
|
|
|
|
// Move the index to the starting quote after the key.
|
|
|
|
startIdx += len(key) + 2
|
|
|
|
endIdx := startIdx
|
|
|
|
|
|
|
|
for endIdx < len(header) {
|
|
|
|
if header[endIdx] == '"' {
|
|
|
|
if endIdx+1 < len(header) && header[endIdx+1] != ',' { // If the next character isn't a comma, continue
|
|
|
|
endIdx++
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
break
|
|
|
|
}
|
|
|
|
endIdx++
|
|
|
|
}
|
|
|
|
return header[startIdx:endIdx]
|
|
|
|
}
|
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
func parseRegistryChallenge(authStr string) registryChallenge {
|
2023-08-10 18:34:25 +00:00
|
|
|
authStr = strings.TrimPrefix(authStr, "Bearer ")
|
|
|
|
|
2024-02-14 19:29:49 +00:00
|
|
|
return registryChallenge{
|
2023-08-10 18:34:25 +00:00
|
|
|
Realm: getValue(authStr, "realm"),
|
|
|
|
Service: getValue(authStr, "service"),
|
|
|
|
Scope: getValue(authStr, "scope"),
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-04-21 02:11:06 +00:00
|
|
|
var errDigestMismatch = errors.New("digest mismatch, file must be downloaded again")
|
2023-07-24 18:53:01 +00:00
|
|
|
|
2023-07-20 18:44:05 +00:00
|
|
|
func verifyBlob(digest string) error {
|
|
|
|
fp, err := GetBlobsPath(digest)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
f, err := os.Open(fp)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
defer f.Close()
|
|
|
|
|
|
|
|
fileDigest, _ := GetSHA256Digest(f)
|
|
|
|
if digest != fileDigest {
|
2023-07-24 18:53:01 +00:00
|
|
|
return fmt.Errorf("%w: want %s, got %s", errDigestMismatch, digest, fileDigest)
|
2023-07-20 18:44:05 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|