phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-04 12:07:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

refactor: convert providers to be installed via package

Browse source 
currently providers have a `pip_package` list. Rather than make our own form of python dependency management, we should use `pyproject.toml` files in each provider declaring the dependencies in a more trackable manner.
Each provider can then be installed using the already in place `module` field in the ProviderSpec, pointing to the directory the provider lives in
we can then simply `uv pip install` this directory as opposed to installing the dependencies one by one

Signed-off-by: Charlie Doern <cdoern@redhat.com>
This commit is contained in:
Charlie Doern 2025-07-29 15:18:54 -04:00
parent a1301911e4
commit 41431d8bdd
76 changed files with 1294 additions and 134 deletions
Download patch file Download diff file Expand all files Collapse all files

40
llama_stack/providers/registry/post_training.py
View file

@ -5,15 +5,12 @@
# the root directory of this source tree.
from typing import cast
from llama_stack.providers.datatypes import Api, InlineProviderSpec, ProviderSpec, RemoteProviderSpec
# We provide two versions of these providers so that distributions can package the appropriate version of torch.
# The CPU version is used for distributions that don't have GPU support -- they result in smaller container images.
torchtune_def = dict(
api=Api.post_training,
pip_packages=["numpy"],
module="llama_stack.providers.inline.post_training.torchtune",
config_class="llama_stack.providers.inline.post_training.torchtune.TorchtunePostTrainingConfig",
api_dependencies=[
@ -27,28 +24,32 @@ torchtune_def = dict(
def available_providers() -> list[ProviderSpec]:
return [
InlineProviderSpec(
**{ # type: ignore
**torchtune_def,
"provider_type": "inline::torchtune-cpu",
"pip_packages": (
cast(list[str], torchtune_def["pip_packages"])
+ ["torch torchtune>=0.5.0 torchao>=0.12.0 --extra-index-url https://download.pytorch.org/whl/cpu"]
),
},
api=Api.post_training,
provider_type="inline::torchtune-cpu",
module="llama_stack.providers.inline.post_training.torchtune",
config_class="llama_stack.providers.inline.post_training.torchtune.TorchtunePostTrainingConfig",
api_dependencies=[
Api.datasetio,
Api.datasets,
],
description="TorchTune-based post-training provider for fine-tuning and optimizing models using Meta's TorchTune framework (CPU).",
package_extras=["cpu"],
),
InlineProviderSpec(
**{ # type: ignore
**torchtune_def,
"provider_type": "inline::torchtune-gpu",
"pip_packages": (
cast(list[str], torchtune_def["pip_packages"]) + ["torch torchtune>=0.5.0 torchao>=0.12.0"]
),
},
api=Api.post_training,
provider_type="inline::torchtune-gpu",
module="llama_stack.providers.inline.post_training.torchtune",
config_class="llama_stack.providers.inline.post_training.torchtune.TorchtunePostTrainingConfig",
api_dependencies=[
Api.datasetio,
Api.datasets,
],
description="TorchTune-based post-training provider for fine-tuning and optimizing models using Meta's TorchTune framework (GPU).",
package_extras=["gpu"],
),
InlineProviderSpec(
api=Api.post_training,
provider_type="inline::huggingface-gpu",
pip_packages=["trl", "transformers", "peft", "datasets>=4.0.0", "torch"],
module="llama_stack.providers.inline.post_training.huggingface",
config_class="llama_stack.providers.inline.post_training.huggingface.HuggingFacePostTrainingConfig",
api_dependencies=[
@ -61,7 +62,6 @@ def available_providers() -> list[ProviderSpec]:
api=Api.post_training,
adapter_type="nvidia",
provider_type="remote::nvidia",
pip_packages=["requests", "aiohttp"],
module="llama_stack.providers.remote.post_training.nvidia",
config_class="llama_stack.providers.remote.post_training.nvidia.NvidiaPostTrainingConfig",
description="NVIDIA's post-training provider for fine-tuning models on NVIDIA's platform.",