phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-29 07:38:52 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into add-watsonx-inference-adapter

Browse source
This commit is contained in:
Sajikumar JS 2025-03-20 10:18:05 +05:30
commit 28e6c8478b
308 changed files with 33749 additions and 5102 deletions
Download patch file Download diff file Expand all files Collapse all files

13
llama_stack/providers/remote/inference/databricks/config.py
View file

@ -4,6 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from pydantic import BaseModel, Field
@ -20,3 +21,15 @@ class DatabricksImplConfig(BaseModel):
default=None,
description="The Databricks API token",
)
@classmethod
def sample_run_config(
cls,
url: str = "${env.DATABRICKS_URL}",
api_token: str = "${env.DATABRICKS_API_TOKEN}",
**kwargs: Any,
) -> Dict[str, Any]:
return {
"url": url,
"api_token": api_token,
}

4
llama_stack/providers/remote/inference/fireworks/models.py
View file

@ -24,10 +24,6 @@ MODEL_ENTRIES = [
"accounts/fireworks/models/llama-v3p1-405b-instruct",
CoreModelId.llama3_1_405b_instruct.value,
),
build_hf_repo_model_entry(
"accounts/fireworks/models/llama-v3p2-1b-instruct",
CoreModelId.llama3_2_1b_instruct.value,
),
build_hf_repo_model_entry(
"accounts/fireworks/models/llama-v3p2-3b-instruct",
CoreModelId.llama3_2_3b_instruct.value,

51
llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -6,6 +6,7 @@
import logging
import warnings
from functools import lru_cache
from typing import AsyncIterator, List, Optional, Union
from openai import APIConnectionError, AsyncOpenAI, BadRequestError
@ -82,12 +83,42 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
# )
self._config = config
# make sure the client lives longer than any async calls
self._client = AsyncOpenAI(
base_url=f"{self._config.url}/v1",
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
@lru_cache # noqa: B019
def _get_client(self, provider_model_id: str) -> AsyncOpenAI:
"""
For hosted models, https://integrate.api.nvidia.com/v1 is the primary base_url. However,
some models are hosted on different URLs. This function returns the appropriate client
for the given provider_model_id.
This relies on lru_cache and self._default_client to avoid creating a new client for each request
or for each model that is hosted on https://integrate.api.nvidia.com/v1.
:param provider_model_id: The provider model ID
:return: An OpenAI client
"""
@lru_cache # noqa: B019
def _get_client_for_base_url(base_url: str) -> AsyncOpenAI:
"""
Maintain a single OpenAI client per base_url.
"""
return AsyncOpenAI(
base_url=base_url,
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
special_model_urls = {
"meta/llama-3.2-11b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-11b-vision-instruct",
"meta/llama-3.2-90b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-90b-vision-instruct",
}
base_url = f"{self._config.url}/v1"
if _is_nvidia_hosted(self._config) and provider_model_id in special_model_urls:
base_url = special_model_urls[provider_model_id]
return _get_client_for_base_url(base_url)
async def completion(
self,
@ -105,9 +136,10 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
await check_health(self._config) # this raises errors
provider_model_id = self.get_provider_model_id(model_id)
request = convert_completion_request(
request=CompletionRequest(
model=self.get_provider_model_id(model_id),
model=provider_model_id,
content=content,
sampling_params=sampling_params,
response_format=response_format,
@ -118,7 +150,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._client.completions.create(**request)
response = await self._get_client(provider_model_id).completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e
@ -206,6 +238,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
await check_health(self._config) # this raises errors
provider_model_id = self.get_provider_model_id(model_id)
request = await convert_chat_completion_request(
request=ChatCompletionRequest(
model=self.get_provider_model_id(model_id),
@ -221,7 +254,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._client.chat.completions.create(**request)
response = await self._get_client(provider_model_id).chat.completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e

87
llama_stack/providers/remote/inference/passthrough/passthrough.py
View file

@ -4,12 +4,15 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import AsyncGenerator, List, Optional
from typing import Any, AsyncGenerator, Dict, List, Optional
from llama_stack_client import LlamaStackClient
from llama_stack_client import AsyncLlamaStackClient
from llama_stack.apis.common.content_types import InterleavedContent
from llama_stack.apis.inference import (
ChatCompletionResponse,
ChatCompletionResponseStreamChunk,
CompletionMessage,
EmbeddingsResponse,
EmbeddingTaskType,
Inference,
@ -24,6 +27,7 @@ from llama_stack.apis.inference import (
ToolPromptFormat,
)
from llama_stack.apis.models import Model
from llama_stack.distribution.library_client import convert_pydantic_to_json_value, convert_to_pydantic
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
from .config import PassthroughImplConfig
@ -46,7 +50,7 @@ class PassthroughInferenceAdapter(Inference):
async def register_model(self, model: Model) -> Model:
return model
def _get_client(self) -> LlamaStackClient:
def _get_client(self) -> AsyncLlamaStackClient:
passthrough_url = None
passthrough_api_key = None
provider_data = None
@ -71,7 +75,7 @@ class PassthroughInferenceAdapter(Inference):
)
passthrough_api_key = provider_data.passthrough_api_key
return LlamaStackClient(
return AsyncLlamaStackClient(
base_url=passthrough_url,
api_key=passthrough_api_key,
provider_data=provider_data,
@ -91,7 +95,7 @@ class PassthroughInferenceAdapter(Inference):
client = self._get_client()
model = await self.model_store.get_model(model_id)
params = {
request_params = {
"model_id": model.provider_resource_id,
"content": content,
"sampling_params": sampling_params,
@ -100,10 +104,13 @@ class PassthroughInferenceAdapter(Inference):
"logprobs": logprobs,
}
params = {key: value for key, value in params.items() if value is not None}
request_params = {key: value for key, value in request_params.items() if value is not None}
# cast everything to json dict
json_params = self.cast_value_to_json_dict(request_params)
# only pass through the not None params
return client.inference.completion(**params)
return await client.inference.completion(**json_params)
async def chat_completion(
self,
@ -120,10 +127,14 @@ class PassthroughInferenceAdapter(Inference):
) -> AsyncGenerator:
if sampling_params is None:
sampling_params = SamplingParams()
client = self._get_client()
model = await self.model_store.get_model(model_id)
params = {
# TODO: revisit this remove tool_calls from messages logic
for message in messages:
if hasattr(message, "tool_calls"):
message.tool_calls = None
request_params = {
"model_id": model.provider_resource_id,
"messages": messages,
"sampling_params": sampling_params,
@ -135,10 +146,41 @@ class PassthroughInferenceAdapter(Inference):
"logprobs": logprobs,
}
params = {key: value for key, value in params.items() if value is not None}
# only pass through the not None params
return client.inference.chat_completion(**params)
request_params = {key: value for key, value in request_params.items() if value is not None}
# cast everything to json dict
json_params = self.cast_value_to_json_dict(request_params)
if stream:
return self._stream_chat_completion(json_params)
else:
return await self._nonstream_chat_completion(json_params)
async def _nonstream_chat_completion(self, json_params: Dict[str, Any]) -> ChatCompletionResponse:
client = self._get_client()
response = await client.inference.chat_completion(**json_params)
return ChatCompletionResponse(
completion_message=CompletionMessage(
content=response.completion_message.content.text,
stop_reason=response.completion_message.stop_reason,
tool_calls=response.completion_message.tool_calls,
),
logprobs=response.logprobs,
)
async def _stream_chat_completion(self, json_params: Dict[str, Any]) -> AsyncGenerator:
client = self._get_client()
stream_response = await client.inference.chat_completion(**json_params)
async for chunk in stream_response:
chunk = chunk.to_dict()
# temporary hack to remove the metrics from the response
chunk["metrics"] = []
chunk = convert_to_pydantic(ChatCompletionResponseStreamChunk, chunk)
yield chunk
async def embeddings(
self,
@ -151,10 +193,29 @@ class PassthroughInferenceAdapter(Inference):
client = self._get_client()
model = await self.model_store.get_model(model_id)
return client.inference.embeddings(
return await client.inference.embeddings(
model_id=model.provider_resource_id,
contents=contents,
text_truncation=text_truncation,
output_dimension=output_dimension,
task_type=task_type,
)
def cast_value_to_json_dict(self, request_params: Dict[str, Any]) -> Dict[str, Any]:
json_params = {}
for key, value in request_params.items():
json_input = convert_pydantic_to_json_value(value)
if isinstance(json_input, dict):
json_input = {k: v for k, v in json_input.items() if v is not None}
elif isinstance(json_input, list):
json_input = [x for x in json_input if x is not None]
new_input = []
for x in json_input:
if isinstance(x, dict):
x = {k: v for k, v in x.items() if v is not None}
new_input.append(x)
json_input = new_input
json_params[key] = json_input
return json_params

3
llama_stack/providers/remote/inference/runpod/__init__.py
View file

@ -5,10 +5,11 @@
# the root directory of this source tree.
from .config import RunpodImplConfig
from .runpod import RunpodInferenceAdapter
async def get_adapter_impl(config: RunpodImplConfig, _deps):
from .runpod import RunpodInferenceAdapter
assert isinstance(config, RunpodImplConfig), f"Unexpected config type: {type(config)}"
impl = RunpodInferenceAdapter(config)
await impl.initialize()

9
llama_stack/providers/remote/inference/runpod/config.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Optional
from typing import Any, Dict, Optional
from pydantic import BaseModel, Field
@ -21,3 +21,10 @@ class RunpodImplConfig(BaseModel):
default=None,
description="The API token",
)
@classmethod
def sample_run_config(cls, **kwargs: Any) -> Dict[str, Any]:
return {
"url": "${env.RUNPOD_URL:}",
"api_token": "${env.RUNPOD_API_TOKEN:}",
}

1
llama_stack/providers/remote/inference/runpod/runpod.py
View file

@ -8,7 +8,6 @@ from typing import AsyncGenerator
from openai import OpenAI
from llama_stack.apis.inference import * # noqa: F403
from llama_stack.models.llama.datatypes import Message
# from llama_stack.providers.datatypes import ModelsProtocolPrivate
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper

10
llama_stack/providers/remote/inference/sambanova/sambanova.py
View file

@ -42,9 +42,7 @@ from llama_stack.models.llama.datatypes import (
TopKSamplingStrategy,
TopPSamplingStrategy,
)
from llama_stack.providers.utils.inference.model_registry import (
ModelRegistryHelper,
)
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
from llama_stack.providers.utils.inference.openai_compat import (
process_chat_completion_stream_response,
)
@ -293,14 +291,12 @@ class SambaNovaInferenceAdapter(ModelRegistryHelper, Inference):
if not tool_calls:
return []
for call in tool_calls:
call_function_arguments = json.loads(call.function.arguments)
compitable_tool_calls = [
ToolCall(
call_id=call.id,
tool_name=call.function.name,
arguments=call_function_arguments,
arguments=json.loads(call.function.arguments),
arguments_json=call.function.arguments,
)
for call in tool_calls
]

17
llama_stack/providers/remote/inference/sample/__init__.py
View file

@ -1,17 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any
from .config import SampleConfig
async def get_adapter_impl(config: SampleConfig, _deps) -> Any:
from .sample import SampleInferenceImpl
impl = SampleInferenceImpl(config)
await impl.initialize()
return impl

12
llama_stack/providers/remote/inference/sample/config.py
View file

@ -1,12 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from pydantic import BaseModel
class SampleConfig(BaseModel):
host: str = "localhost"
port: int = 9999

23
llama_stack/providers/remote/inference/sample/sample.py
View file

@ -1,23 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.apis.inference import Inference
from llama_stack.apis.models import Model
from .config import SampleConfig
class SampleInferenceImpl(Inference):
def __init__(self, config: SampleConfig):
self.config = config
async def register_model(self, model: Model) -> None:
# these are the model names the Llama Stack will use to route requests to this provider
# perform validation here if necessary
pass
async def initialize(self):
pass

2
llama_stack/providers/remote/inference/together/config.py
View file

@ -26,5 +26,5 @@ class TogetherImplConfig(BaseModel):
def sample_run_config(cls, **kwargs) -> Dict[str, Any]:
return {
"url": "https://api.together.xyz/v1",
"api_key": "${env.TOGETHER_API_KEY}",
"api_key": "${env.TOGETHER_API_KEY:}",
}

71
llama_stack/providers/remote/inference/together/together.py
View file

@ -6,7 +6,7 @@
from typing import AsyncGenerator, List, Optional, Union
from together import Together
from together import AsyncTogether
from llama_stack.apis.common.content_types import (
InterleavedContent,
@ -59,12 +59,15 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
def __init__(self, config: TogetherImplConfig) -> None:
ModelRegistryHelper.__init__(self, MODEL_ENTRIES)
self.config = config
self._client = None
async def initialize(self) -> None:
pass
async def shutdown(self) -> None:
pass
if self._client:
await self._client.close()
self._client = None
async def completion(
self,
@ -91,35 +94,32 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
else:
return await self._nonstream_completion(request)
def _get_client(self) -> Together:
together_api_key = None
config_api_key = self.config.api_key.get_secret_value() if self.config.api_key else None
if config_api_key:
together_api_key = config_api_key
else:
provider_data = self.get_request_provider_data()
if provider_data is None or not provider_data.together_api_key:
raise ValueError(
'Pass Together API Key in the header X-LlamaStack-Provider-Data as { "together_api_key": <your api key>}'
)
together_api_key = provider_data.together_api_key
return Together(api_key=together_api_key)
def _get_client(self) -> AsyncTogether:
if not self._client:
together_api_key = None
config_api_key = self.config.api_key.get_secret_value() if self.config.api_key else None
if config_api_key:
together_api_key = config_api_key
else:
provider_data = self.get_request_provider_data()
if provider_data is None or not provider_data.together_api_key:
raise ValueError(
'Pass Together API Key in the header X-LlamaStack-Provider-Data as { "together_api_key": <your api key>}'
)
together_api_key = provider_data.together_api_key
self._client = AsyncTogether(api_key=together_api_key)
return self._client
async def _nonstream_completion(self, request: CompletionRequest) -> ChatCompletionResponse:
params = await self._get_params(request)
r = self._get_client().completions.create(**params)
client = self._get_client()
r = await client.completions.create(**params)
return process_completion_response(r)
async def _stream_completion(self, request: CompletionRequest) -> AsyncGenerator:
params = await self._get_params(request)
# if we shift to TogetherAsyncClient, we won't need this wrapper
async def _to_async_generator():
s = self._get_client().completions.create(**params)
for chunk in s:
yield chunk
stream = _to_async_generator()
client = await self._get_client()
stream = await client.completions.create(**params)
async for chunk in process_completion_stream_response(stream):
yield chunk
@ -184,25 +184,21 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
async def _nonstream_chat_completion(self, request: ChatCompletionRequest) -> ChatCompletionResponse:
params = await self._get_params(request)
client = self._get_client()
if "messages" in params:
r = self._get_client().chat.completions.create(**params)
r = await client.chat.completions.create(**params)
else:
r = self._get_client().completions.create(**params)
r = await client.completions.create(**params)
return process_chat_completion_response(r, request)
async def _stream_chat_completion(self, request: ChatCompletionRequest) -> AsyncGenerator:
params = await self._get_params(request)
client = self._get_client()
if "messages" in params:
stream = await client.chat.completions.create(**params)
else:
stream = await client.completions.create(**params)
# if we shift to TogetherAsyncClient, we won't need this wrapper
async def _to_async_generator():
if "messages" in params:
s = self._get_client().chat.completions.create(**params)
else:
s = self._get_client().completions.create(**params)
for chunk in s:
yield chunk
stream = _to_async_generator()
async for chunk in process_chat_completion_stream_response(stream, request):
yield chunk
@ -240,7 +236,8 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
assert all(not content_has_media(content) for content in contents), (
"Together does not support media for embeddings"
)
r = self._get_client().embeddings.create(
client = self._get_client()
r = await client.embeddings.create(
model=model.provider_resource_id,
input=[interleaved_content_as_str(content) for content in contents],
)

5
llama_stack/providers/remote/inference/vllm/config.py
View file

@ -25,6 +25,10 @@ class VLLMInferenceAdapterConfig(BaseModel):
default="fake",
description="The API token",
)
tls_verify: bool = Field(
default=True,
description="Whether to verify TLS certificates",
)
@classmethod
def sample_run_config(
@ -36,4 +40,5 @@ class VLLMInferenceAdapterConfig(BaseModel):
"url": url,
"max_tokens": "${env.VLLM_MAX_TOKENS:4096}",
"api_token": "${env.VLLM_API_TOKEN:fake}",
"tls_verify": "${env.VLLM_TLS_VERIFY:true}",
}

15
llama_stack/providers/remote/inference/vllm/vllm.py
View file

@ -7,6 +7,7 @@ import json
import logging
from typing import AsyncGenerator, List, Optional, Union
import httpx
from openai import AsyncOpenAI
from openai.types.chat.chat_completion_chunk import (
ChatCompletionChunk as OpenAIChatCompletionChunk,
@ -89,15 +90,12 @@ def _convert_to_vllm_tool_calls_in_response(
if not tool_calls:
return []
call_function_arguments = None
for call in tool_calls:
call_function_arguments = json.loads(call.function.arguments)
return [
ToolCall(
call_id=call.id,
tool_name=call.function.name,
arguments=call_function_arguments,
arguments=json.loads(call.function.arguments),
arguments_json=call.function.arguments,
)
for call in tool_calls
]
@ -182,6 +180,7 @@ async def _process_vllm_chat_completion_stream_response(
call_id=tool_call_buf.call_id,
tool_name=tool_call_buf.tool_name,
arguments=args,
arguments_json=args_str,
),
parse_status=ToolCallParseStatus.succeeded,
),
@ -229,7 +228,11 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
async def initialize(self) -> None:
log.info(f"Initializing VLLM client with base_url={self.config.url}")
self.client = AsyncOpenAI(base_url=self.config.url, api_key=self.config.api_token)
self.client = AsyncOpenAI(
base_url=self.config.url,
api_key=self.config.api_token,
http_client=None if self.config.tls_verify else httpx.AsyncClient(verify=False),
)
async def shutdown(self) -> None:
pass