phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-06-28 02:53:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat: updated inline vllm inference provider (#880)

Browse source 
# What does this PR do?

This PR updates the inline vLLM inference provider in several
significant ways:
* Models are now attached at run time to instances of the provider via
the `.../models` API instead of hard-coding the model's full name into
the provider's YAML configuration.
* The provider supports models that are not Meta Llama models. Any model
that vLLM supports can be loaded by passing Huggingface coordinates in
the "provider_model_id" field. Custom fine-tuned versions of Meta Llama
models can be loaded by specifying a path on local disk in the
"provider_model_id".
* To implement full chat completions support, including tool calling and
constrained decoding, the provider now routes the `chat_completions` API
to a captive (i.e. called directly in-process, not via HTTPS) instance
of vLLM's OpenAI-compatible server .
* The `logprobs` parameter and completions API are also working.

## Test Plan

Existing tests in
`llama_stack/providers/tests/inference/test_text_inference.py` have good
coverage of the new functionality. These tests can be invoked as
follows:

```
cd llama-stack && pytest \
    -vvv \
    llama_stack/providers/tests/inference/test_text_inference.py \
    --providers inference=vllm \
    --inference-model meta-llama/Llama-3.2-3B-Instruct
====================================== test session starts ======================================
platform linux -- Python 3.12.8, pytest-8.3.4, pluggy-1.5.0 -- /mnt/datadisk1/freiss/llama/env/bin/python3.12
cachedir: .pytest_cache
metadata: {'Python': '3.12.8', 'Platform': 'Linux-6.8.0-1016-ibm-x86_64-with-glibc2.39', 'Packages': {'pytest': '8.3.4', 'pluggy': '1.5.0'}, 'Plugins': {'anyio': '4.8.0', 'html': '4.1.1', 'metadata': '3.1.1', 'asyncio': '0.25.2'}, 'JAVA_HOME': '/usr/lib/jvm/java-8-openjdk-amd64'}
rootdir: /mnt/datadisk1/freiss/llama/llama-stack
configfile: pyproject.toml
plugins: anyio-4.8.0, html-4.1.1, metadata-3.1.1, asyncio-0.25.2
asyncio: mode=Mode.STRICT, asyncio_default_fixture_loop_scope=None
collected 9 items                                                                               

llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_model_list[-vllm] PASSED [ 11%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_completion[-vllm] PASSED [ 22%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_completion_logprobs[-vllm] PASSED [ 33%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_completion_structured_output[-vllm] PASSED [ 44%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_chat_completion_non_streaming[-vllm] PASSED [ 55%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_structured_output[-vllm] PASSED [ 66%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_chat_completion_streaming[-vllm] PASSED [ 77%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_chat_completion_with_tool_calling[-vllm] PASSED [ 88%]
llama_stack/providers/tests/inference/test_text_inference.py::TestInference::test_chat_completion_with_tool_calling_streaming[-vllm] PASSED [100%]

=========================== 9 passed, 13 warnings in 97.18s (0:01:37) ===========================

```

## Sources


## Before submitting

- [X] Ran pre-commit to handle lint / formatting issues.
- [X] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
      Pull Request section?
- [ ] Updated relevant documentation.
- [ ] Wrote necessary unit or integration tests.

---------

Co-authored-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
This commit is contained in:
Fred Reiss 2025-03-07 13:38:23 -08:00 • committed by GitHub
parent acbae66b9d
commit a8d0cdaf37
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 887 additions and 176 deletions
Download patch file Download diff file Expand all files Collapse all files

170
llama_stack/providers/inline/inference/vllm/openai_utils.py Normal file
View file

@ -0,0 +1,170 @@
# 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 List, Optional
import vllm
from llama_stack.apis.inference import (
ChatCompletionRequest,
GrammarResponseFormat,
JsonSchemaResponseFormat,
Message,
ToolChoice,
UserMessage,
)
from llama_stack.models.llama.datatypes import BuiltinTool, ToolDefinition
from llama_stack.providers.utils.inference.openai_compat import (
convert_message_to_openai_dict,
get_sampling_options,
)
###############################################################################
# This file contains OpenAI compatibility code that is currently only used
# by the inline vLLM connector. Some or all of this code may be moved to a
# central location at a later date.
def _merge_context_into_content(message: Message) -> Message: # type: ignore
"""
Merge the ``context`` field of a Llama Stack ``Message`` object into
the content field for compabilitiy with OpenAI-style APIs.
Generates a content string that emulates the current behavior
of ``llama_models.llama3.api.chat_format.encode_message()``.
:param message: Message that may include ``context`` field
:returns: A version of ``message`` with any context merged into the
``content`` field.
"""
if not isinstance(message, UserMessage): # Separate type check for linter
return message
if message.context is None:
return message
return UserMessage(
role=message.role,
# Emumate llama_models.llama3.api.chat_format.encode_message()
content=message.content + "\n\n" + message.context,
context=None,
)
def _llama_stack_tools_to_openai_tools(
tools: Optional[List[ToolDefinition]] = None,
) -> List[vllm.entrypoints.openai.protocol.ChatCompletionToolsParam]:
"""
Convert the list of available tools from Llama Stack's format to vLLM's
version of OpenAI's format.
"""
if tools is None:
return []
result = []
for t in tools:
if isinstance(t.tool_name, BuiltinTool):
raise NotImplementedError("Built-in tools not yet implemented")
if t.parameters is None:
parameters = None
else: # if t.parameters is not None
# Convert the "required" flags to a list of required params
required_params = [k for k, v in t.parameters.items() if v.required]
parameters = {
"type": "object", # Mystery value that shows up in OpenAI docs
"properties": {
k: {"type": v.param_type, "description": v.description} for k, v in t.parameters.items()
},
"required": required_params,
}
function_def = vllm.entrypoints.openai.protocol.FunctionDefinition(
name=t.tool_name, description=t.description, parameters=parameters
)
# Every tool definition is double-boxed in a ChatCompletionToolsParam
result.append(vllm.entrypoints.openai.protocol.ChatCompletionToolsParam(function=function_def))
return result
async def llama_stack_chat_completion_to_openai_chat_completion_dict(
request: ChatCompletionRequest,
) -> dict:
"""
Convert a chat completion request in Llama Stack format into an
equivalent set of arguments to pass to an OpenAI-compatible
chat completions API.
:param request: Bundled request parameters in Llama Stack format.
:returns: Dictionary of key-value pairs to use as an initializer
for a dataclass or to be converted directly to JSON and sent
over the wire.
"""
converted_messages = [
# This mystery async call makes the parent function also be async
await convert_message_to_openai_dict(_merge_context_into_content(m), download=True)
for m in request.messages
]
converted_tools = _llama_stack_tools_to_openai_tools(request.tools)
# Llama will try to use built-in tools with no tool catalog, so don't enable
# tool choice unless at least one tool is enabled.
converted_tool_choice = "none"
if (
request.tool_config is not None
and request.tool_config.tool_choice == ToolChoice.auto
and request.tools is not None
and len(request.tools) > 0
):
converted_tool_choice = "auto"
# TODO: Figure out what to do with the tool_prompt_format argument.
# Other connectors appear to drop it quietly.
# Use Llama Stack shared code to translate sampling parameters.
sampling_options = get_sampling_options(request.sampling_params)
# get_sampling_options() translates repetition penalties to an option that
# OpenAI's APIs don't know about.
# vLLM's OpenAI-compatible API also handles repetition penalties wrong.
# For now, translate repetition penalties into a format that vLLM's broken
# API will handle correctly. Two wrongs make a right...
if "repeat_penalty" in sampling_options:
del sampling_options["repeat_penalty"]
if request.sampling_params.repetition_penalty is not None and request.sampling_params.repetition_penalty != 1.0:
sampling_options["repetition_penalty"] = request.sampling_params.repetition_penalty
# Convert a single response format into four different parameters, per
# the OpenAI spec
guided_decoding_options = dict()
if request.response_format is None:
# Use defaults
pass
elif isinstance(request.response_format, JsonSchemaResponseFormat):
guided_decoding_options["guided_json"] = request.response_format.json_schema
elif isinstance(request.response_format, GrammarResponseFormat):
guided_decoding_options["guided_grammar"] = request.response_format.bnf
else:
raise TypeError(f"ResponseFormat object is of unexpected subtype '{type(request.response_format)}'")
logprob_options = dict()
if request.logprobs is not None:
logprob_options["logprobs"] = request.logprobs.top_k
# Marshall together all the arguments for a ChatCompletionRequest
request_options = {
"model": request.model,
"messages": converted_messages,
"tools": converted_tools,
"tool_choice": converted_tool_choice,
"stream": request.stream,
**sampling_options,
**guided_decoding_options,
**logprob_options,
}
return request_options