phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-06-27 18:50:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Allow models to be registered as long as llama model is provided (#472)

Browse source 
This PR allows models to be registered with provider as long as the user
specifies a llama model, even though the model does not match our
prebuilt provider specific mapping.
Test:
pytest -v -s
llama_stack/providers/tests/inference/test_model_registration.py -m
"together" --env TOGETHER_API_KEY=<KEY>

---------

Co-authored-by: Dinesh Yeduguru <dineshyv@fb.com>
This commit is contained in:
Dinesh Yeduguru 2024-11-18 15:05:29 -08:00 • committed by GitHub
parent 2a31163178
commit 57a9b4d57f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 72 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

50
llama_stack/providers/tests/inference/test_model_registration.py
View file

@ -6,7 +6,6 @@
import pytest
from llama_models.datatypes import CoreModelId
# How to run this test:
#
@ -17,11 +16,22 @@ from llama_models.datatypes import CoreModelId
class TestModelRegistration:
@pytest.mark.asyncio
async def test_register_unsupported_model(self, inference_stack):
_, models_impl = inference_stack
async def test_register_unsupported_model(self, inference_stack, inference_model):
inference_impl, models_impl = inference_stack
provider = inference_impl.routing_table.get_provider_impl(inference_model)
if provider.__provider_spec__.provider_type not in (
"meta-reference",
"remote::ollama",
"remote::vllm",
"remote::tgi",
):
pytest.skip(
"Skipping test for remote inference providers since they can handle large models like 70B instruct"
)
# Try to register a model that's too large for local inference
with pytest.raises(Exception) as exc_info:
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="Llama3.1-70B-Instruct",
)
@ -37,21 +47,27 @@ class TestModelRegistration:
)
@pytest.mark.asyncio
async def test_update_model(self, inference_stack):
async def test_register_with_llama_model(self, inference_stack):
_, models_impl = inference_stack
# Register a model to update
model_id = CoreModelId.llama3_1_8b_instruct.value
old_model = await models_impl.register_model(model_id=model_id)
# Update the model
new_model_id = CoreModelId.llama3_2_3b_instruct.value
updated_model = await models_impl.update_model(
model_id=model_id, provider_model_id=new_model_id
_ = await models_impl.register_model(
model_id="custom-model",
metadata={"llama_model": "meta-llama/Llama-2-7b"},
)
# Retrieve the updated model to verify changes
assert updated_model.provider_resource_id != old_model.provider_resource_id
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="custom-model-2",
metadata={"llama_model": "meta-llama/Llama-2-7b"},
provider_model_id="custom-model",
)
# Cleanup
await models_impl.unregister_model(model_id=model_id)
@pytest.mark.asyncio
async def test_register_with_invalid_llama_model(self, inference_stack):
_, models_impl = inference_stack
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="custom-model-2",
metadata={"llama_model": "invalid-llama-model"},
)

5
llama_stack/providers/utils/inference/__init__.py
View file

@ -31,3 +31,8 @@ def supported_inference_models() -> List[str]:
or is_supported_safety_model(m)
)
]
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR = {
m.huggingface_repo: m.descriptor() for m in all_registered_models()
}

38
llama_stack/providers/utils/inference/model_registry.py
View file

@ -11,6 +11,10 @@ from llama_models.sku_list import all_registered_models
from llama_stack.providers.datatypes import Model, ModelsProtocolPrivate
from llama_stack.providers.utils.inference import (
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR,
)
ModelAlias = namedtuple("ModelAlias", ["provider_model_id", "aliases", "llama_model"])
@ -51,7 +55,7 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
if identifier in self.alias_to_provider_id_map:
return self.alias_to_provider_id_map[identifier]
else:
raise ValueError(f"Unknown model: `{identifier}`")
return None
def get_llama_model(self, provider_model_id: str) -> str:
if provider_model_id in self.provider_id_to_llama_model_map:
@ -60,8 +64,34 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
return None
async def register_model(self, model: Model) -> Model:
model.provider_resource_id = self.get_provider_model_id(
model.provider_resource_id
)
provider_resource_id = self.get_provider_model_id(model.provider_resource_id)
if provider_resource_id:
model.provider_resource_id = provider_resource_id
else:
if model.metadata.get("llama_model") is None:
raise ValueError(
f"Model '{model.provider_resource_id}' is not available and no llama_model was specified in metadata. "
"Please specify a llama_model in metadata or use a supported model identifier"
)
existing_llama_model = self.get_llama_model(model.provider_resource_id)
if existing_llama_model:
if existing_llama_model != model.metadata["llama_model"]:
raise ValueError(
f"Provider model id '{model.provider_resource_id}' is already registered to a different llama model: '{existing_llama_model}'"
)
else:
if (
model.metadata["llama_model"]
not in ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR
):
raise ValueError(
f"Invalid llama_model '{model.metadata['llama_model']}' specified in metadata. "
f"Must be one of: {', '.join(ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR.keys())}"
)
self.provider_id_to_llama_model_map[model.provider_resource_id] = (
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR[
model.metadata["llama_model"]
]
)
return model