phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-05 02:17:31 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into chroma

Browse source
This commit is contained in:
kimbwook 2025-10-15 00:14:05 +09:00
commit 3f66f55771
No known key found for this signature in database
GPG key ID: 13B032C99CBD373A
137 changed files with 35682 additions and 1800 deletions
Download patch file Download diff file Expand all files Collapse all files

7
llama_stack/providers/remote/inference/bedrock/bedrock.py
View file

@ -14,6 +14,7 @@ from llama_stack.apis.inference import (
Inference,
OpenAIChatCompletionRequestWithExtraBody,
OpenAICompletionRequestWithExtraBody,
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.apis.inference.inference import (
@ -124,11 +125,7 @@ class BedrockInferenceAdapter(
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()

11
llama_stack/providers/remote/inference/cerebras/cerebras.py
View file

@ -6,7 +6,10 @@
from urllib.parse import urljoin
from llama_stack.apis.inference import OpenAIEmbeddingsResponse
from llama_stack.apis.inference import (
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.providers.utils.inference.openai_mixin import OpenAIMixin
from .config import CerebrasImplConfig
@ -20,10 +23,6 @@ class CerebrasInferenceAdapter(OpenAIMixin):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()

7
llama_stack/providers/remote/inference/llama_openai_compat/llama.py
View file

@ -7,6 +7,7 @@
from llama_stack.apis.inference.inference import (
OpenAICompletion,
OpenAICompletionRequestWithExtraBody,
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.log import get_logger
@ -40,10 +41,6 @@ class LlamaCompatInferenceAdapter(OpenAIMixin):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()

17
llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -9,6 +9,7 @@ from openai import NOT_GIVEN
from llama_stack.apis.inference import (
OpenAIEmbeddingData,
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
OpenAIEmbeddingUsage,
)
@ -78,11 +79,7 @@ class NVIDIAInferenceAdapter(OpenAIMixin):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
"""
OpenAI-compatible embeddings for NVIDIA NIM.
@ -99,11 +96,11 @@ class NVIDIAInferenceAdapter(OpenAIMixin):
)
response = await self.client.embeddings.create(
model=await self._get_provider_model_id(model),
input=input,
encoding_format=encoding_format if encoding_format is not None else NOT_GIVEN,
dimensions=dimensions if dimensions is not None else NOT_GIVEN,
user=user if user is not None else NOT_GIVEN,
model=await self._get_provider_model_id(params.model),
input=params.input,
encoding_format=params.encoding_format if params.encoding_format is not None else NOT_GIVEN,
dimensions=params.dimensions if params.dimensions is not None else NOT_GIVEN,
user=params.user if params.user is not None else NOT_GIVEN,
extra_body=extra_body,
)

7
llama_stack/providers/remote/inference/passthrough/passthrough.py
View file

@ -16,6 +16,7 @@ from llama_stack.apis.inference import (
OpenAIChatCompletionRequestWithExtraBody,
OpenAICompletion,
OpenAICompletionRequestWithExtraBody,
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.apis.models import Model
@ -69,11 +70,7 @@ class PassthroughInferenceAdapter(Inference):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()

11
llama_stack/providers/remote/inference/tgi/tgi.py
View file

@ -10,7 +10,10 @@ from collections.abc import Iterable
from huggingface_hub import AsyncInferenceClient, HfApi
from pydantic import SecretStr
from llama_stack.apis.inference import OpenAIEmbeddingsResponse
from llama_stack.apis.inference import (
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.log import get_logger
from llama_stack.providers.utils.inference.openai_mixin import OpenAIMixin
@ -40,11 +43,7 @@ class _HfAdapter(OpenAIMixin):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
raise NotImplementedError()

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

@ -11,6 +11,7 @@ from together import AsyncTogether
from together.constants import BASE_URL
from llama_stack.apis.inference import (
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.apis.inference.inference import OpenAIEmbeddingUsage
@ -62,11 +63,7 @@ class TogetherInferenceAdapter(OpenAIMixin, NeedsRequestProviderData):
async def openai_embeddings(
self,
model: str,
input: str | list[str],
encoding_format: str | None = "float",
dimensions: int | None = None,
user: str | None = None,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
"""
Together's OpenAI-compatible embeddings endpoint is not compatible with
@ -78,25 +75,27 @@ class TogetherInferenceAdapter(OpenAIMixin, NeedsRequestProviderData):
- does not support dimensions param, returns 400 Unrecognized request arguments supplied: dimensions
"""
# Together support ticket #13332 -> will not fix
if user is not None:
if params.user is not None:
raise ValueError("Together's embeddings endpoint does not support user param.")
# Together support ticket #13333 -> escalated
if dimensions is not None:
if params.dimensions is not None:
raise ValueError("Together's embeddings endpoint does not support dimensions param.")
response = await self.client.embeddings.create(
model=await self._get_provider_model_id(model),
input=input,
encoding_format=encoding_format,
model=await self._get_provider_model_id(params.model),
input=params.input,
encoding_format=params.encoding_format,
)
response.model = model # return the user the same model id they provided, avoid exposing the provider model id
response.model = (
params.model
) # return the user the same model id they provided, avoid exposing the provider model id
# Together support ticket #13330 -> escalated
# - togethercomputer/m2-bert-80M-32k-retrieval *does not* return usage information
if not hasattr(response, "usage") or response.usage is None:
logger.warning(
f"Together's embedding endpoint for {model} did not return usage information, substituting -1s."
f"Together's embedding endpoint for {params.model} did not return usage information, substituting -1s."
)
response.usage = OpenAIEmbeddingUsage(prompt_tokens=-1, total_tokens=-1)

10
llama_stack/providers/remote/inference/watsonx/config.py
View file

@ -7,18 +7,18 @@
import os
from typing import Any
from pydantic import BaseModel, ConfigDict, Field
from pydantic import BaseModel, Field
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack.schema_utils import json_schema_type
class WatsonXProviderDataValidator(BaseModel):
model_config = ConfigDict(
from_attributes=True,
extra="forbid",
watsonx_project_id: str | None = Field(
default=None,
description="IBM WatsonX project ID",
)
watsonx_api_key: str | None
watsonx_api_key: str | None = None
@json_schema_type

243
llama_stack/providers/remote/inference/watsonx/watsonx.py
View file

@ -4,42 +4,259 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from collections.abc import AsyncIterator
from typing import Any
import litellm
import requests
from llama_stack.apis.inference import ChatCompletionRequest
from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionChunk,
OpenAIChatCompletionRequestWithExtraBody,
OpenAIChatCompletionUsage,
OpenAICompletion,
OpenAICompletionRequestWithExtraBody,
OpenAIEmbeddingsRequestWithExtraBody,
OpenAIEmbeddingsResponse,
)
from llama_stack.apis.models import Model
from llama_stack.apis.models.models import ModelType
from llama_stack.log import get_logger
from llama_stack.providers.remote.inference.watsonx.config import WatsonXConfig
from llama_stack.providers.utils.inference.litellm_openai_mixin import LiteLLMOpenAIMixin
from llama_stack.providers.utils.inference.openai_compat import prepare_openai_completion_params
from llama_stack.providers.utils.telemetry.tracing import get_current_span
logger = get_logger(name=__name__, category="providers::remote::watsonx")
class WatsonXInferenceAdapter(LiteLLMOpenAIMixin):
_model_cache: dict[str, Model] = {}
provider_data_api_key_field: str = "watsonx_api_key"
def __init__(self, config: WatsonXConfig):
self.available_models = None
self.config = config
api_key = config.auth_credential.get_secret_value() if config.auth_credential else None
LiteLLMOpenAIMixin.__init__(
self,
litellm_provider_name="watsonx",
api_key_from_config=config.auth_credential.get_secret_value() if config.auth_credential else None,
api_key_from_config=api_key,
provider_data_api_key_field="watsonx_api_key",
openai_compat_api_base=self.get_base_url(),
)
async def openai_chat_completion(
self,
params: OpenAIChatCompletionRequestWithExtraBody,
) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
"""
Override parent method to add timeout and inject usage object when missing.
This works around a LiteLLM defect where usage block is sometimes dropped.
"""
# Add usage tracking for streaming when telemetry is active
stream_options = params.stream_options
if params.stream and get_current_span() is not None:
if stream_options is None:
stream_options = {"include_usage": True}
elif "include_usage" not in stream_options:
stream_options = {**stream_options, "include_usage": True}
model_obj = await self.model_store.get_model(params.model)
request_params = await prepare_openai_completion_params(
model=self.get_litellm_model_name(model_obj.provider_resource_id),
messages=params.messages,
frequency_penalty=params.frequency_penalty,
function_call=params.function_call,
functions=params.functions,
logit_bias=params.logit_bias,
logprobs=params.logprobs,
max_completion_tokens=params.max_completion_tokens,
max_tokens=params.max_tokens,
n=params.n,
parallel_tool_calls=params.parallel_tool_calls,
presence_penalty=params.presence_penalty,
response_format=params.response_format,
seed=params.seed,
stop=params.stop,
stream=params.stream,
stream_options=stream_options,
temperature=params.temperature,
tool_choice=params.tool_choice,
tools=params.tools,
top_logprobs=params.top_logprobs,
top_p=params.top_p,
user=params.user,
api_key=self.get_api_key(),
api_base=self.api_base,
# These are watsonx-specific parameters
timeout=self.config.timeout,
project_id=self.config.project_id,
)
result = await litellm.acompletion(**request_params)
# If not streaming, check and inject usage if missing
if not params.stream:
# Use getattr to safely handle cases where usage attribute might not exist
if getattr(result, "usage", None) is None:
# Create usage object with zeros
usage_obj = OpenAIChatCompletionUsage(
prompt_tokens=0,
completion_tokens=0,
total_tokens=0,
)
# Use model_copy to create a new response with the usage injected
result = result.model_copy(update={"usage": usage_obj})
return result
# For streaming, wrap the iterator to normalize chunks
return self._normalize_stream(result)
def _normalize_chunk(self, chunk: OpenAIChatCompletionChunk) -> OpenAIChatCompletionChunk:
"""
Normalize a chunk to ensure it has all expected attributes.
This works around LiteLLM not always including all expected attributes.
"""
# Ensure chunk has usage attribute with zeros if missing
if not hasattr(chunk, "usage") or chunk.usage is None:
usage_obj = OpenAIChatCompletionUsage(
prompt_tokens=0,
completion_tokens=0,
total_tokens=0,
)
chunk = chunk.model_copy(update={"usage": usage_obj})
# Ensure all delta objects in choices have expected attributes
if hasattr(chunk, "choices") and chunk.choices:
normalized_choices = []
for choice in chunk.choices:
if hasattr(choice, "delta") and choice.delta:
delta = choice.delta
# Build update dict for missing attributes
delta_updates = {}
if not hasattr(delta, "refusal"):
delta_updates["refusal"] = None
if not hasattr(delta, "reasoning_content"):
delta_updates["reasoning_content"] = None
# If we need to update delta, create a new choice with updated delta
if delta_updates:
new_delta = delta.model_copy(update=delta_updates)
new_choice = choice.model_copy(update={"delta": new_delta})
normalized_choices.append(new_choice)
else:
normalized_choices.append(choice)
else:
normalized_choices.append(choice)
# If we modified any choices, create a new chunk with updated choices
if any(normalized_choices[i] is not chunk.choices[i] for i in range(len(chunk.choices))):
chunk = chunk.model_copy(update={"choices": normalized_choices})
return chunk
async def _normalize_stream(
self, stream: AsyncIterator[OpenAIChatCompletionChunk]
) -> AsyncIterator[OpenAIChatCompletionChunk]:
"""
Normalize all chunks in the stream to ensure they have expected attributes.
This works around LiteLLM sometimes not including expected attributes.
"""
try:
async for chunk in stream:
# Normalize and yield each chunk immediately
yield self._normalize_chunk(chunk)
except Exception as e:
logger.error(f"Error normalizing stream: {e}", exc_info=True)
raise
async def openai_completion(
self,
params: OpenAICompletionRequestWithExtraBody,
) -> OpenAICompletion:
"""
Override parent method to add watsonx-specific parameters.
"""
from llama_stack.providers.utils.inference.openai_compat import prepare_openai_completion_params
model_obj = await self.model_store.get_model(params.model)
request_params = await prepare_openai_completion_params(
model=self.get_litellm_model_name(model_obj.provider_resource_id),
prompt=params.prompt,
best_of=params.best_of,
echo=params.echo,
frequency_penalty=params.frequency_penalty,
logit_bias=params.logit_bias,
logprobs=params.logprobs,
max_tokens=params.max_tokens,
n=params.n,
presence_penalty=params.presence_penalty,
seed=params.seed,
stop=params.stop,
stream=params.stream,
stream_options=params.stream_options,
temperature=params.temperature,
top_p=params.top_p,
user=params.user,
suffix=params.suffix,
api_key=self.get_api_key(),
api_base=self.api_base,
# These are watsonx-specific parameters
timeout=self.config.timeout,
project_id=self.config.project_id,
)
return await litellm.atext_completion(**request_params)
async def openai_embeddings(
self,
params: OpenAIEmbeddingsRequestWithExtraBody,
) -> OpenAIEmbeddingsResponse:
"""
Override parent method to add watsonx-specific parameters.
"""
model_obj = await self.model_store.get_model(params.model)
# Convert input to list if it's a string
input_list = [params.input] if isinstance(params.input, str) else params.input
# Call litellm embedding function with watsonx-specific parameters
response = litellm.embedding(
model=self.get_litellm_model_name(model_obj.provider_resource_id),
input=input_list,
api_key=self.get_api_key(),
api_base=self.api_base,
dimensions=params.dimensions,
# These are watsonx-specific parameters
timeout=self.config.timeout,
project_id=self.config.project_id,
)
# Convert response to OpenAI format
from llama_stack.apis.inference import OpenAIEmbeddingUsage
from llama_stack.providers.utils.inference.litellm_openai_mixin import b64_encode_openai_embeddings_response
data = b64_encode_openai_embeddings_response(response.data, params.encoding_format)
usage = OpenAIEmbeddingUsage(
prompt_tokens=response["usage"]["prompt_tokens"],
total_tokens=response["usage"]["total_tokens"],
)
return OpenAIEmbeddingsResponse(
data=data,
model=model_obj.provider_resource_id,
usage=usage,
)
self.available_models = None
self.config = config
def get_base_url(self) -> str:
return self.config.url
async def _get_params(self, request: ChatCompletionRequest) -> dict[str, Any]:
# Get base parameters from parent
params = await super()._get_params(request)
# Add watsonx.ai specific parameters
params["project_id"] = self.config.project_id
params["time_limit"] = self.config.timeout
return params
# Copied from OpenAIMixin
async def check_model_availability(self, model: str) -> bool:
"""

4
llama_stack/providers/remote/safety/bedrock/bedrock.py
View file

@ -7,7 +7,7 @@
import json
from typing import Any
from llama_stack.apis.inference import Message
from llama_stack.apis.inference import OpenAIMessageParam
from llama_stack.apis.safety import (
RunShieldResponse,
Safety,
@ -56,7 +56,7 @@ class BedrockSafetyAdapter(Safety, ShieldsProtocolPrivate):
pass
async def run_shield(
self, shield_id: str, messages: list[Message], params: dict[str, Any] = None
self, shield_id: str, messages: list[OpenAIMessageParam], params: dict[str, Any] = None
) -> RunShieldResponse:
shield = await self.shield_store.get_shield(shield_id)
if not shield:

10
llama_stack/providers/remote/safety/nvidia/nvidia.py
View file

@ -8,12 +8,11 @@ from typing import Any
import requests
from llama_stack.apis.inference import Message
from llama_stack.apis.inference import OpenAIMessageParam
from llama_stack.apis.safety import ModerationObject, RunShieldResponse, Safety, SafetyViolation, ViolationLevel
from llama_stack.apis.shields import Shield
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import ShieldsProtocolPrivate
from llama_stack.providers.utils.inference.openai_compat import convert_message_to_openai_dict_new
from .config import NVIDIASafetyConfig
@ -44,7 +43,7 @@ class NVIDIASafetyAdapter(Safety, ShieldsProtocolPrivate):
pass
async def run_shield(
self, shield_id: str, messages: list[Message], params: dict[str, Any] | None = None
self, shield_id: str, messages: list[OpenAIMessageParam], params: dict[str, Any] | None = None
) -> RunShieldResponse:
"""
Run a safety shield check against the provided messages.
@ -118,7 +117,7 @@ class NeMoGuardrails:
response.raise_for_status()
return response.json()
async def run(self, messages: list[Message]) -> RunShieldResponse:
async def run(self, messages: list[OpenAIMessageParam]) -> RunShieldResponse:
"""
Queries the /v1/guardrails/checks endpoint of the NeMo guardrails deployed API.
@ -132,10 +131,9 @@ class NeMoGuardrails:
Raises:
requests.HTTPError: If the POST request fails.
"""
request_messages = [await convert_message_to_openai_dict_new(message) for message in messages]
request_data = {
"model": self.model,
"messages": request_messages,
"messages": [{"role": message.role, "content": message.content} for message in messages],
"temperature": self.temperature,
"top_p": 1,
"frequency_penalty": 0,

12
llama_stack/providers/remote/safety/sambanova/sambanova.py
View file

@ -4,13 +4,12 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import json
from typing import Any
import litellm
import requests
from llama_stack.apis.inference import Message
from llama_stack.apis.inference import OpenAIMessageParam
from llama_stack.apis.safety import (
RunShieldResponse,
Safety,
@ -21,7 +20,6 @@ from llama_stack.apis.shields import Shield
from llama_stack.core.request_headers import NeedsRequestProviderData
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import ShieldsProtocolPrivate
from llama_stack.providers.utils.inference.openai_compat import convert_message_to_openai_dict_new
from .config import SambaNovaSafetyConfig
@ -72,7 +70,7 @@ class SambaNovaSafetyAdapter(Safety, ShieldsProtocolPrivate, NeedsRequestProvide
pass
async def run_shield(
self, shield_id: str, messages: list[Message], params: dict[str, Any] | None = None
self, shield_id: str, messages: list[OpenAIMessageParam], params: dict[str, Any] | None = None
) -> RunShieldResponse:
shield = await self.shield_store.get_shield(shield_id)
if not shield:
@ -80,12 +78,8 @@ class SambaNovaSafetyAdapter(Safety, ShieldsProtocolPrivate, NeedsRequestProvide
shield_params = shield.params
logger.debug(f"run_shield::{shield_params}::messages={messages}")
content_messages = [await convert_message_to_openai_dict_new(m) for m in messages]
logger.debug(f"run_shield::final:messages::{json.dumps(content_messages, indent=2)}:")
response = litellm.completion(
model=shield.provider_resource_id, messages=content_messages, api_key=self._get_api_key()
)
response = litellm.completion(model=shield.provider_resource_id, messages=messages, api_key=self._get_api_key())
shield_message = response.choices[0].message.content
if "unsafe" in shield_message.lower():