9f2d8d2117
...
2024-04-04 00:11:31 -07:00
d42c3f6be1
x/build/blob: add fuzz test for ParseRef
2024-04-03 23:50:43 -07:00
4ea3e9efa6
x/build/blob: lock in zero allocs for ParseRef
2024-04-03 23:03:36 -07:00
2e1ea6ecaa
x/build/blob: move most commit value checks to emit func
2024-04-03 22:55:53 -07:00
6d2da77ce2
x/build/blob: add Parts for streaming ref parts
...
Also, make ParseRef use the new Parts method to parse the ref parts.
2024-04-03 22:27:55 -07:00
def4d902bf
... wip still broke
2024-04-03 22:15:58 -07:00
76a202c04e
...
2024-04-03 20:52:27 -07:00
f7cfe946dc
x/registry: fixing tests wip
2024-04-03 16:37:27 -07:00
005b6373e2
x/registry: fix startMinio
2024-04-03 16:19:50 -07:00
d54e0fb3b2
...
2024-04-03 16:14:22 -07:00
bdd05e0ae0
x/registry: skip ref test
2024-04-03 15:59:23 -07:00
1a346640db
x/registry: work on getting basic test passing
2024-04-03 15:58:04 -07:00
f5883070f8
x/registry: upload smoke test passing
2024-04-03 14:30:58 -07:00
adc23d5f96
Add 'x/' from commit 'a10a11b9d371f36b7c3510da32a1d70b74e27bd1'
...
git-subtree-dir: x
git-subtree-mainline: 7d05a6ee8fa10a11b9d3
2024-04-03 10:40:23 -07:00
a10a11b9d3
registry: initial work on multipart pushes
2024-04-03 10:39:30 -07:00
7d05a6ee8f
cmd: provide feedback if OLLAMA_MODELS is set on non-serve command ( #3470 )
...
This also moves the checkServerHeartbeat call out of the "RunE" Cobra
stuff (that's the only word I have for that) to on-site where it's after
the check for OLLAMA_MODELS, which allows the helpful error message to
be printed before the server heartbeat check. This also arguably makes
the code more readable without the magic/superfluous "pre" function
caller.
2024-04-02 22:11:13 -07:00
464d817824
Merge pull request #3464 from dhiltgen/subprocess
...
Fix numgpu opt miscomparison
2024-04-02 20:10:17 -07:00
531324a9be
feat: add OLLAMA_DEBUG in ollama server help message ( #3461 )
...
Co-authored-by: Pier Francesco Contino <pfcontino@gmail.com>
2024-04-02 18:20:03 -07:00
6589eb8a8c
Revert options as a ref in the server
2024-04-02 16:44:10 -07:00
a039e383cd
Merge pull request #3465 from ollama/mxyng/fix-metal
...
fix metal gpu
2024-04-02 16:29:58 -07:00
80163ebcb5
fix metal gpu
2024-04-02 16:06:45 -07:00
a57818d93e
Merge pull request #3343 from dhiltgen/bump_more2
...
Bump llama.cpp to b2581
2024-04-02 15:08:26 -07:00
94befe366a
...
2024-04-02 14:28:06 -07:00
c95f97689b
utils/upload: init
2024-04-02 14:15:21 -07:00
618eb5b909
registry: multipart push
2024-04-02 13:40:23 -07:00
841adda157
Fix windows lint CI flakiness
2024-04-02 12:22:16 -07:00
0035e31af8
Bump to b2581
2024-04-02 11:53:07 -07:00
eb75418be9
build/blob: test ParseRef round-trip
2024-04-02 11:45:01 -07:00
9959da05de
build/blob: break out test refs for other tests/fuzzing
2024-04-02 11:38:10 -07:00
c863c6a96d
Merge pull request #3218 from dhiltgen/subprocess
...
Switch back to subprocessing for llama.cpp
2024-04-02 10:49:44 -07:00
aff7970628
build: remove superfluous parseCompleteRef
2024-04-01 23:41:42 -07:00
628f1feb36
build: back to taking manifests as []byte
...
Its nicer to have the manifests be an opaque []byte, rather than a
struct. This way users of the build package don't need to know about the
internal structure of the manifests. The registry can interpret the
manifests as it sees fit, while letting build keep its own Go type of
manifest which is easier to work with in the build package.
2024-04-01 23:18:58 -07:00
ce3125afd5
registry: add New and take a minio client as argument
2024-04-01 22:53:49 -07:00
f488652ba7
build: make Build accept only refs without builds
2024-04-01 22:12:43 -07:00
2318ed2919
build: remove unused manifest()
2024-04-01 21:59:38 -07:00
b1b8be33d9
build: cleanup error names and other things
2024-04-01 21:57:34 -07:00
876f7eab81
build: move Manifest from internal/blobstore to build
...
It was getting confusing to have the arbirary handling of manifests in
the blobstore. It also prevented us from using model.Ref in the
blobstore because of cyclic dependencies.
This is much easier to grok now.
2024-04-01 21:43:30 -07:00
7cfc8a0838
build/blob: fix awkward Ref type
2024-04-01 21:25:18 -07:00
1f11b52511
Refined min memory from testing
2024-04-01 16:48:33 -07:00
526d4eb204
Release gpu discovery library after use
...
Leaving the cudart library loaded kept ~30m of memory
pinned in the GPU in the main process. This change ensures
we don't hold GPU resources when idle.
2024-04-01 16:48:33 -07:00
0a74cb31d5
Safeguard for noexec
...
We may have users that run into problems with our current
payload model, so this gives us an escape valve.
2024-04-01 16:48:33 -07:00
10ed1b6292
Detect too-old cuda driver
...
"cudart init failure: 35" isn't particularly helpful in the logs.
2024-04-01 16:48:33 -07:00
4fec5816d6
Integration test improvements
...
Cleaner shutdown logic, a bit of response hardening
2024-04-01 16:48:18 -07:00
0a0e9f3e0f
Apply 01-cache.diff
2024-04-01 16:48:18 -07:00
58d95cc9bd
Switch back to subprocessing for llama.cpp
...
This should resolve a number of memory leak and stability defects by allowing
us to isolate llama.cpp in a separate process and shutdown when idle, and
gracefully restart if it has problems. This also serves as a first step to be
able to run multiple copies to support multiple models concurrently.
2024-04-01 16:48:18 -07:00
3b6a9154dd
Simplify model conversion ( #3422 )
2024-04-01 16:14:53 -07:00
d6dd2ff839
Merge pull request #3241 from ollama/mxyng/mem
...
update memory estimations for gpu offloading
2024-04-01 13:59:14 -07:00
e57a6ba89f
Merge pull request #2926 from ollama/mxyng/decode-ggml-v2
...
refactor model parsing
2024-04-01 13:58:13 -07:00
12ec2346ef
Merge pull request #3442 from ollama/mxyng/generate-output
...
fix generate output
2024-04-01 13:56:09 -07:00
1ec0df1069
fix generate output
2024-04-01 13:47:34 -07:00
Blake Mizerany
Daniel Hiltgen
Pier Francesco Contino
Daniel Hiltgen
Michael Yang
Patrick Devine