phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-16 14:29:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into dead_code_removal

Browse source
This commit is contained in:
Omar Abdelwahab 2025-10-06 13:21:36 -07:00 • committed by GitHub
commit 9886520b40
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
927 changed files with 171924 additions and 102933 deletions
Download patch file Download diff file Expand all files Collapse all files

2
llama_stack/providers/inline/agents/meta_reference/agents.py
View file

@ -329,6 +329,7 @@ class MetaReferenceAgentsImpl(Agents):
tools: list[OpenAIResponseInputTool] | None = None,
include: list[str] | None = None,
max_infer_iters: int | None = 10,
shields: list | None = None,
) -> OpenAIResponseObject:
return await self.openai_responses_impl.create_openai_response(
input,
@ -342,6 +343,7 @@ class MetaReferenceAgentsImpl(Agents):
tools,
include,
max_infer_iters,
shields,
)
async def list_openai_responses(

75
llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py
View file

@ -8,7 +8,7 @@ import time
import uuid
from collections.abc import AsyncIterator
from pydantic import BaseModel
from pydantic import BaseModel, TypeAdapter
from llama_stack.apis.agents import Order
from llama_stack.apis.agents.openai_responses import (
@ -26,12 +26,16 @@ from llama_stack.apis.agents.openai_responses import (
)
from llama_stack.apis.inference import (
Inference,
OpenAIMessageParam,
OpenAISystemMessageParam,
)
from llama_stack.apis.tools import ToolGroups, ToolRuntime
from llama_stack.apis.vector_io import VectorIO
from llama_stack.log import get_logger
from llama_stack.providers.utils.responses.responses_store import ResponsesStore
from llama_stack.providers.utils.responses.responses_store import (
ResponsesStore,
_OpenAIResponseObjectWithInputAndMessages,
)
from .streaming import StreamingResponseOrchestrator
from .tool_executor import ToolExecutor
@ -72,26 +76,48 @@ class OpenAIResponsesImpl:
async def _prepend_previous_response(
self,
input: str | list[OpenAIResponseInput],
previous_response_id: str | None = None,
previous_response: _OpenAIResponseObjectWithInputAndMessages,
):
new_input_items = previous_response.input.copy()
new_input_items.extend(previous_response.output)
if isinstance(input, str):
new_input_items.append(OpenAIResponseMessage(content=input, role="user"))
else:
new_input_items.extend(input)
return new_input_items
async def _process_input_with_previous_response(
self,
input: str | list[OpenAIResponseInput],
previous_response_id: str | None,
) -> tuple[str | list[OpenAIResponseInput], list[OpenAIMessageParam]]:
"""Process input with optional previous response context.
Returns:
tuple: (all_input for storage, messages for chat completion)
"""
if previous_response_id:
previous_response_with_input = await self.responses_store.get_response_object(previous_response_id)
previous_response: _OpenAIResponseObjectWithInputAndMessages = (
await self.responses_store.get_response_object(previous_response_id)
)
all_input = await self._prepend_previous_response(input, previous_response)
# previous response input items
new_input_items = previous_response_with_input.input
# previous response output items
new_input_items.extend(previous_response_with_input.output)
# new input items from the current request
if isinstance(input, str):
new_input_items.append(OpenAIResponseMessage(content=input, role="user"))
if previous_response.messages:
# Use stored messages directly and convert only new input
message_adapter = TypeAdapter(list[OpenAIMessageParam])
messages = message_adapter.validate_python(previous_response.messages)
new_messages = await convert_response_input_to_chat_messages(input)
messages.extend(new_messages)
else:
new_input_items.extend(input)
# Backward compatibility: reconstruct from inputs
messages = await convert_response_input_to_chat_messages(all_input)
else:
all_input = input
messages = await convert_response_input_to_chat_messages(input)
input = new_input_items
return input
return all_input, messages
async def _prepend_instructions(self, messages, instructions):
if instructions:
@ -102,7 +128,7 @@ class OpenAIResponsesImpl:
response_id: str,
) -> OpenAIResponseObject:
response_with_input = await self.responses_store.get_response_object(response_id)
return OpenAIResponseObject(**{k: v for k, v in response_with_input.model_dump().items() if k != "input"})
return response_with_input.to_response_object()
async def list_openai_responses(
self,
@ -138,6 +164,7 @@ class OpenAIResponsesImpl:
self,
response: OpenAIResponseObject,
input: str | list[OpenAIResponseInput],
messages: list[OpenAIMessageParam],
) -> None:
new_input_id = f"msg_{uuid.uuid4()}"
if isinstance(input, str):
@ -165,6 +192,7 @@ class OpenAIResponsesImpl:
await self.responses_store.store_response_object(
response_object=response,
input=input_items_data,
messages=messages,
)
async def create_openai_response(
@ -180,10 +208,15 @@ class OpenAIResponsesImpl:
tools: list[OpenAIResponseInputTool] | None = None,
include: list[str] | None = None,
max_infer_iters: int | None = 10,
shields: list | None = None,
):
stream = bool(stream)
text = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")) if text is None else text
# Shields parameter received via extra_body - not yet implemented
if shields is not None:
raise NotImplementedError("Shields parameter is not yet implemented in the meta-reference provider")
stream_gen = self._create_streaming_response(
input=input,
model=model,
@ -224,8 +257,7 @@ class OpenAIResponsesImpl:
max_infer_iters: int | None = 10,
) -> AsyncIterator[OpenAIResponseObjectStream]:
# Input preprocessing
input = await self._prepend_previous_response(input, previous_response_id)
messages = await convert_response_input_to_chat_messages(input)
all_input, messages = await self._process_input_with_previous_response(input, previous_response_id)
await self._prepend_instructions(messages, instructions)
# Structured outputs
@ -265,7 +297,8 @@ class OpenAIResponsesImpl:
if store and final_response:
await self._store_response(
response=final_response,
input=input,
input=all_input,
messages=orchestrator.final_messages,
)
async def delete_openai_response(self, response_id: str) -> OpenAIDeleteResponseObject:

5
llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
View file

@ -43,6 +43,7 @@ from llama_stack.apis.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionToolCall,
OpenAIChoice,
OpenAIMessageParam,
)
from llama_stack.log import get_logger
@ -94,6 +95,8 @@ class StreamingResponseOrchestrator:
self.sequence_number = 0
# Store MCP tool mapping that gets built during tool processing
self.mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP] = {}
# Track final messages after all tool executions
self.final_messages: list[OpenAIMessageParam] = []
async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
# Initialize output messages
@ -183,6 +186,8 @@ class StreamingResponseOrchestrator:
messages = next_turn_messages
self.final_messages = messages.copy() + [current_response.choices[0].message]
# Create final response
final_response = OpenAIResponseObject(
created_at=self.created_at,

316
llama_stack/providers/inline/inference/meta_reference/inference.py
View file

@ -5,37 +5,17 @@
# the root directory of this source tree.
import asyncio
import os
import sys
from collections.abc import AsyncGenerator
from collections.abc import AsyncIterator
from typing import Any
from pydantic import BaseModel
from termcolor import cprint
from llama_stack.apis.common.content_types import (
TextDelta,
ToolCallDelta,
ToolCallParseStatus,
)
from llama_stack.apis.inference import (
ChatCompletionRequest,
ChatCompletionResponse,
ChatCompletionResponseEvent,
ChatCompletionResponseEventType,
ChatCompletionResponseStreamChunk,
CompletionMessage,
InferenceProvider,
LogProbConfig,
Message,
ResponseFormat,
SamplingParams,
StopReason,
TokenLogProbs,
ToolChoice,
ToolConfig,
ToolDefinition,
ToolPromptFormat,
UserMessage,
)
from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionChunk,
OpenAIMessageParam,
OpenAIResponseFormatParam,
)
from llama_stack.apis.models import Model, ModelType
from llama_stack.log import get_logger
@ -53,13 +33,6 @@ from llama_stack.providers.utils.inference.model_registry import (
ModelRegistryHelper,
build_hf_repo_model_entry,
)
from llama_stack.providers.utils.inference.openai_compat import (
OpenAIChatCompletionToLlamaStackMixin,
)
from llama_stack.providers.utils.inference.prompt_adapter import (
chat_completion_request_to_messages,
convert_request_to_raw,
)
from .config import MetaReferenceInferenceConfig
from .generators import LlamaGenerator
@ -76,7 +49,6 @@ def llama_builder_fn(config: MetaReferenceInferenceConfig, model_id: str, llama_
class MetaReferenceInferenceImpl(
OpenAIChatCompletionToLlamaStackMixin,
SentenceTransformerEmbeddingMixin,
InferenceProvider,
ModelsProtocolPrivate,
@ -161,10 +133,10 @@ class MetaReferenceInferenceImpl(
self.llama_model = llama_model
log.info("Warming up...")
await self.chat_completion(
model_id=model_id,
messages=[UserMessage(content="Hi how are you?")],
sampling_params=SamplingParams(max_tokens=20),
await self.openai_chat_completion(
model=model_id,
messages=[{"role": "user", "content": "Hi how are you?"}],
max_tokens=20,
)
log.info("Warmed up!")
@ -176,242 +148,30 @@ class MetaReferenceInferenceImpl(
elif request.model != self.model_id:
raise RuntimeError(f"Model mismatch: request model: {request.model} != loaded model: {self.model_id}")
async def chat_completion(
async def openai_chat_completion(
self,
model_id: str,
messages: list[Message],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
tools: list[ToolDefinition] | None = None,
tool_choice: ToolChoice | None = ToolChoice.auto,
tool_prompt_format: ToolPromptFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
tool_config: ToolConfig | None = None,
) -> AsyncGenerator:
if sampling_params is None:
sampling_params = SamplingParams()
if logprobs:
assert logprobs.top_k == 1, f"Unexpected top_k={logprobs.top_k}"
# wrapper request to make it easier to pass around (internal only, not exposed to API)
request = ChatCompletionRequest(
model=model_id,
messages=messages,
sampling_params=sampling_params,
tools=tools or [],
response_format=response_format,
stream=stream,
logprobs=logprobs,
tool_config=tool_config or ToolConfig(),
)
self.check_model(request)
# augment and rewrite messages depending on the model
request.messages = chat_completion_request_to_messages(request, self.llama_model.core_model_id.value)
# download media and convert to raw content so we can send it to the model
request = await convert_request_to_raw(request)
if self.config.create_distributed_process_group:
if SEMAPHORE.locked():
raise RuntimeError("Only one concurrent request is supported")
if request.stream:
return self._stream_chat_completion(request)
else:
results = await self._nonstream_chat_completion([request])
return results[0]
async def _nonstream_chat_completion(
self, request_batch: list[ChatCompletionRequest]
) -> list[ChatCompletionResponse]:
tokenizer = self.generator.formatter.tokenizer
first_request = request_batch[0]
class ItemState(BaseModel):
tokens: list[int] = []
logprobs: list[TokenLogProbs] = []
stop_reason: StopReason | None = None
finished: bool = False
def impl():
states = [ItemState() for _ in request_batch]
for token_results in self.generator.chat_completion(request_batch):
first = token_results[0]
if not first.finished and not first.ignore_token:
if os.environ.get("LLAMA_MODELS_DEBUG", "0") in ("1", "2"):
cprint(first.text, color="cyan", end="", file=sys.stderr)
if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "2":
cprint(f"<{first.token}>", color="magenta", end="", file=sys.stderr)
for result in token_results:
idx = result.batch_idx
state = states[idx]
if state.finished or result.ignore_token:
continue
state.finished = result.finished
if first_request.logprobs:
state.logprobs.append(TokenLogProbs(logprobs_by_token={result.text: result.logprobs[0]}))
state.tokens.append(result.token)
if result.token == tokenizer.eot_id:
state.stop_reason = StopReason.end_of_turn
elif result.token == tokenizer.eom_id:
state.stop_reason = StopReason.end_of_message
results = []
for state in states:
if state.stop_reason is None:
state.stop_reason = StopReason.out_of_tokens
raw_message = self.generator.formatter.decode_assistant_message(state.tokens, state.stop_reason)
results.append(
ChatCompletionResponse(
completion_message=CompletionMessage(
content=raw_message.content,
stop_reason=raw_message.stop_reason,
tool_calls=raw_message.tool_calls,
),
logprobs=state.logprobs if first_request.logprobs else None,
)
)
return results
if self.config.create_distributed_process_group:
async with SEMAPHORE:
return impl()
else:
return impl()
async def _stream_chat_completion(self, request: ChatCompletionRequest) -> AsyncGenerator:
tokenizer = self.generator.formatter.tokenizer
def impl():
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.start,
delta=TextDelta(text=""),
)
)
tokens = []
logprobs = []
stop_reason = None
ipython = False
for token_results in self.generator.chat_completion([request]):
token_result = token_results[0]
if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "1":
cprint(token_result.text, color="cyan", end="", file=sys.stderr)
if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "2":
cprint(f"<{token_result.token}>", color="magenta", end="", file=sys.stderr)
if token_result.token == tokenizer.eot_id:
stop_reason = StopReason.end_of_turn
text = ""
elif token_result.token == tokenizer.eom_id:
stop_reason = StopReason.end_of_message
text = ""
else:
text = token_result.text
if request.logprobs:
assert len(token_result.logprobs) == 1
logprobs.append(TokenLogProbs(logprobs_by_token={token_result.text: token_result.logprobs[0]}))
tokens.append(token_result.token)
if not ipython and token_result.text.startswith("<|python_tag|>"):
ipython = True
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
tool_call="",
parse_status=ToolCallParseStatus.started,
),
)
)
continue
if token_result.token == tokenizer.eot_id:
stop_reason = StopReason.end_of_turn
text = ""
elif token_result.token == tokenizer.eom_id:
stop_reason = StopReason.end_of_message
text = ""
else:
text = token_result.text
if ipython:
delta = ToolCallDelta(
tool_call=text,
parse_status=ToolCallParseStatus.in_progress,
)
else:
delta = TextDelta(text=text)
if stop_reason is None:
if request.logprobs:
assert len(token_result.logprobs) == 1
logprobs.append(TokenLogProbs(logprobs_by_token={token_result.text: token_result.logprobs[0]}))
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=delta,
stop_reason=stop_reason,
logprobs=logprobs if request.logprobs else None,
)
)
if stop_reason is None:
stop_reason = StopReason.out_of_tokens
message = self.generator.formatter.decode_assistant_message(tokens, stop_reason)
parsed_tool_calls = len(message.tool_calls) > 0
if ipython and not parsed_tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
tool_call="",
parse_status=ToolCallParseStatus.failed,
),
stop_reason=stop_reason,
)
)
for tool_call in message.tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
tool_call=tool_call,
parse_status=ToolCallParseStatus.succeeded,
),
stop_reason=stop_reason,
)
)
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.complete,
delta=TextDelta(text=""),
stop_reason=stop_reason,
)
)
if self.config.create_distributed_process_group:
async with SEMAPHORE:
for x in impl():
yield x
else:
for x in impl():
yield x
model: str,
messages: list[OpenAIMessageParam],
frequency_penalty: float | None = None,
function_call: str | dict[str, Any] | None = None,
functions: list[dict[str, Any]] | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_completion_tokens: int | None = None,
max_tokens: int | None = None,
n: int | None = None,
parallel_tool_calls: bool | None = None,
presence_penalty: float | None = None,
response_format: OpenAIResponseFormatParam | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
tool_choice: str | dict[str, Any] | None = None,
tools: list[dict[str, Any]] | None = None,
top_logprobs: int | None = None,
top_p: float | None = None,
user: str | None = None,
) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
raise NotImplementedError("OpenAI chat completion not supported by meta-reference inference provider")

65
llama_stack/providers/inline/inference/sentence_transformers/sentence_transformers.py
View file

@ -4,21 +4,19 @@
# 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 AsyncGenerator
from collections.abc import AsyncIterator
from typing import Any
from llama_stack.apis.inference import (
InferenceProvider,
LogProbConfig,
Message,
ResponseFormat,
SamplingParams,
ToolChoice,
ToolConfig,
ToolDefinition,
ToolPromptFormat,
)
from llama_stack.apis.inference.inference import OpenAICompletion
from llama_stack.apis.inference.inference import (
OpenAIChatCompletion,
OpenAIChatCompletionChunk,
OpenAICompletion,
OpenAIMessageParam,
OpenAIResponseFormatParam,
)
from llama_stack.apis.models import ModelType
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import Model, ModelsProtocolPrivate
@ -69,21 +67,6 @@ class SentenceTransformersInferenceImpl(
async def unregister_model(self, model_id: str) -> None:
pass
async def chat_completion(
self,
model_id: str,
messages: list[Message],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
tools: list[ToolDefinition] | None = None,
tool_choice: ToolChoice | None = ToolChoice.auto,
tool_prompt_format: ToolPromptFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
tool_config: ToolConfig | None = None,
) -> AsyncGenerator:
raise ValueError("Sentence transformers don't support chat completion")
async def openai_completion(
self,
# Standard OpenAI completion parameters
@ -110,6 +93,32 @@ class SentenceTransformersInferenceImpl(
# for fill-in-the-middle type completion
suffix: str | None = None,
) -> OpenAICompletion:
raise NotImplementedError(
"OpenAI completion not supported by sentence transformers provider"
)
raise NotImplementedError("OpenAI completion not supported by sentence transformers provider")
async def openai_chat_completion(
self,
model: str,
messages: list[OpenAIMessageParam],
frequency_penalty: float | None = None,
function_call: str | dict[str, Any] | None = None,
functions: list[dict[str, Any]] | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_completion_tokens: int | None = None,
max_tokens: int | None = None,
n: int | None = None,
parallel_tool_calls: bool | None = None,
presence_penalty: float | None = None,
response_format: OpenAIResponseFormatParam | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
tool_choice: str | dict[str, Any] | None = None,
tools: list[dict[str, Any]] | None = None,
top_logprobs: int | None = None,
top_p: float | None = None,
user: str | None = None,
) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
raise NotImplementedError("OpenAI chat completion not supported by sentence transformers provider")