Commit graph

72 commits

Author SHA1 Message Date
Blake Mizerany
cb42e607c5
llm: speed up gguf decoding by a lot (#5246)
Previously, some costly things were causing the loading of GGUF files
and their metadata and tensor information to be VERY slow:

  * Too many allocations when decoding strings
  * Hitting disk for each read of each key and value, resulting in a
    not-okay amount of syscalls/disk I/O.

The show API is now down to 33ms from 800ms+ for llama3 on a macbook pro
m3.

This commit also prevents collecting large arrays of values when
decoding GGUFs (if desired). When such keys are encountered, their
values are null, and are encoded as such in JSON.

Also, this fixes a broken test that was not encoding valid GGUF.
2024-06-24 21:47:52 -07:00
Daniel Hiltgen
5bf5aeec01 Refine mmap default logic on linux
If we try to use mmap when the model is larger than the system free space, loading is slower than the no-mmap approach.
2024-06-20 11:07:04 -07:00
Daniel Hiltgen
96624aa412
Merge pull request #5072 from dhiltgen/windows_path
Move libraries out of users path
2024-06-19 09:13:39 -07:00
Daniel Hiltgen
7784ca33ce Tighten up memory prediction logging
Prior to this change, we logged the memory prediction multiple times
as the scheduler iterates to find a suitable configuration, which can be
confusing since only the last log before the server starts is actually valid.
This now logs once just before starting the server on the final configuration.
It also reports what library instead of always saying "offloading to gpu" when
using CPU.
2024-06-18 09:15:35 -07:00
Daniel Hiltgen
171796791f Adjust mmap logic for cuda windows for faster model load
On Windows, recent llama.cpp changes make mmap slower in most
cases, so default to off.  This also implements a tri-state for
use_mmap so we can detect the difference between a user provided
value of true/false, or unspecified.
2024-06-17 16:54:30 -07:00
Daniel Hiltgen
b2799f111b Move libraries out of users path
We update the PATH on windows to get the CLI mapped, but this has
an unintended side effect of causing other apps that may use our bundled
DLLs to get terminated when we upgrade.
2024-06-17 13:12:18 -07:00
Daniel Hiltgen
da3bf23354 Workaround gfx900 SDMA bugs
Implement support for GPU env var workarounds, and leverage
this for the Vega RX 56 which needs
HSA_ENABLE_SDMA=0 set to work properly
2024-06-14 15:38:13 -07:00
Daniel Hiltgen
6f351bf586 review comments and coverage 2024-06-14 14:55:50 -07:00
Daniel Hiltgen
fc37c192ae Refine CPU load behavior with system memory visibility 2024-06-14 14:51:40 -07:00
Daniel Hiltgen
6fd04ca922 Improve multi-gpu handling at the limit
Still not complete, needs some refinement to our prediction to understand the
discrete GPUs available space so we can see how many layers fit in each one
since we can't split one layer across multiple GPUs we can't treat free space
as one logical block
2024-06-14 14:51:40 -07:00
Craig Hughes
b84aea1685
Critical fix from llama.cpp JSON grammar to forbid un-escaped escape characters inside strings, which breaks parsing. (#3782) 2024-06-09 10:57:09 -07:00
Michael Yang
e40145a39d lint 2024-06-04 11:13:30 -07:00
Michael Yang
c895a7d13f some gocritic 2024-06-04 11:13:30 -07:00
Michael Yang
829ff87bd1
revert tokenize ffi (#4761)
* Revert "use `int32_t` for call to tokenize (#4738)"

This reverts commit 763bb65dbb.

* Revert "vocab only"

This reverts commit bf54c845e9.

* Revert "use ffi for tokenizing/detokenizing"

This reverts commit 26a00a0410.
2024-05-31 18:54:21 -07:00
Jeffrey Morgan
a50a87a7b8
partial offloading: allow flash attention and disable mmap (#4734)
* partial offloading: allow flash attention and disable mmap

* allow mmap with num_gpu=0
2024-05-30 16:58:01 -07:00
Michael Yang
26a00a0410 use ffi for tokenizing/detokenizing 2024-05-29 11:26:47 -07:00
Daniel Hiltgen
92c81e8117 Give the final model loading more time
On some systems, 1 minute isn't sufficient to finish the load after it
hits 100% This creates 2 distinct timers, although they're both set to
the same value for now so we can refine the timeouts further.
2024-05-28 09:08:10 -07:00
Lei Jitang
7487229c34
llm/server.go: Fix 2 minor typos (#4661)
Signed-off-by: Lei Jitang <leijitang@outlook.com>
2024-05-27 17:21:10 -07:00
Daniel Hiltgen
0165ba1651
Merge pull request #4638 from dhiltgen/better_error
Report better warning on client closed abort of load
2024-05-25 14:32:28 -07:00
Daniel Hiltgen
c4209d6d21 Report better warning on client closed abort of load
If the client closes the connection before we finish loading the model
we abort, so lets make the log message clearer why to help users
understand this failure mode
2024-05-25 09:23:28 -07:00
Patrick Devine
4cc3be3035
Move envconfig and consolidate env vars (#4608) 2024-05-24 14:57:15 -07:00
Daniel Hiltgen
b37b496a12 Wire up load progress
This doesn't expose a UX yet, but wires the initial server portion
of progress reporting during load
2024-05-23 13:36:48 -07:00
Jeffrey Morgan
38255d2af1
Use flash attention flag for now (#4580)
* put flash attention behind flag for now

* add test

* remove print

* up timeout for sheduler tests
2024-05-22 21:52:09 -07:00
Sam
e15307fdf4
feat: add support for flash_attn (#4120)
* feat: enable flash attention if supported

* feat: enable flash attention if supported

* feat: enable flash attention if supported

* feat: add flash_attn support
2024-05-20 13:36:03 -07:00
Patrick Devine
d1692fd3e0
fix the cpu estimatedTotal memory + get the expiry time for loading models (#4461) 2024-05-15 15:43:16 -07:00
Daniel Hiltgen
853ae490e1 Sanitize the env var debug log
Only dump env vars we care about in the logs
2024-05-15 14:42:57 -07:00
Patrick Devine
6845988807
Ollama ps command for showing currently loaded models (#4327) 2024-05-13 17:17:36 -07:00
jmorganca
92ca2cca95 Revert "only forward some env vars"
This reverts commit ce3b212d12.
2024-05-10 22:53:21 -07:00
Daniel Hiltgen
c4014e73a2 Fall back to CPU runner with zero layers 2024-05-10 15:09:48 -07:00
Jeffrey Morgan
bb6fd02298
Don't clamp ctx size in PredictServerFit (#4317)
* dont clamp ctx size in `PredictServerFit`

* minimum 4 context

* remove context warning
2024-05-10 10:17:12 -07:00
Michael Yang
cf442cd57e fix typo 2024-05-09 16:23:37 -07:00
Michael Yang
ce3b212d12 only forward some env vars 2024-05-09 15:16:09 -07:00
Michael Yang
58876091f7 log clean up 2024-05-09 14:55:36 -07:00
Daniel Hiltgen
d0425f26cf
Merge pull request #4294 from dhiltgen/harden_subprocess_reaping
Harden subprocess reaping
2024-05-09 14:02:16 -07:00
Bruce MacDonald
cfa84b8470
add done_reason to the api (#4235) 2024-05-09 13:30:14 -07:00
Daniel Hiltgen
84ac7ce139 Refine subprocess reaping 2024-05-09 11:21:31 -07:00
Daniel Hiltgen
920a4b0794 Merge remote-tracking branch 'upstream/main' into pr3702 2024-05-08 16:44:35 -07:00
Daniel Hiltgen
ee49844d09
Merge pull request #4153 from dhiltgen/gpu_verbose_response
Add GPU usage
2024-05-08 16:39:11 -07:00
Daniel Hiltgen
bee2f4a3b0 Record GPU usage information
This records more GPU usage information for eventual UX inclusion.
2024-05-08 14:45:39 -07:00
Daniel Hiltgen
72700279e2 Detect noexec and report a better error
This will bubble up a much more informative error message if noexec
is preventing us from running the subprocess
2024-05-07 16:46:15 -07:00
Daniel Hiltgen
380378cc80 Use our libraries first
Trying to live off the land for cuda libraries was not the right strategy.  We need to use the version we compiled against to ensure things work properly
2024-05-06 14:23:29 -07:00
Jeffrey Morgan
ed740a2504
Fix no slots available error with concurrent requests (#4160) 2024-05-06 14:22:53 -07:00
Jeffrey Morgan
1b0e6c9c0e
Fix llava models not working after first request (#4164)
* fix llava models not working after first request

* individual requests only for llava models
2024-05-05 20:50:31 -07:00
Daniel Hiltgen
f56aa20014 Centralize server config handling
This moves all the env var reading into one central module
and logs the loaded config once at startup which should
help in troubleshooting user server logs
2024-05-05 16:49:50 -07:00
Mark Ward
321d57e1a0 Removing go routine calling .wait from load. 2024-05-01 18:51:10 +00:00
Mark Ward
ba26c7aa00 it will always return an error due to Kill() discarding Wait() errors 2024-05-01 18:51:10 +00:00
Mark Ward
63c763685f log when the waiting for the process to stop to help debug when other tasks execute during this wait.
expire timer clear the timer reference because it will not be reused.
close will clean up expireTimer if calling code has not already done this.
2024-05-01 18:51:10 +00:00
Mark Ward
948114e3e3 fix sched to wait for the runner to terminate to ensure following vram check will be more accurate 2024-05-01 18:51:10 +00:00
Jeffrey Morgan
7aa08a77ca
llm: dont cap context window limit to training context window (#3988) 2024-04-29 10:07:30 -04:00
Jeffrey Morgan
bb31def011
return code 499 when user cancels request while a model is loading (#3955) 2024-04-26 17:38:29 -04:00