[Bug] Fix Failure in /v1/chat/completions/render for Multimodal Requests (https://github.com/vllm-project/vllm/issues/35665) (#35684)
This commit is contained in:
@@ -24,14 +24,29 @@ from vllm.entrypoints.openai.parser.harmony_utils import (
|
||||
parse_chat_inputs_to_harmony_messages,
|
||||
render_for_completion,
|
||||
)
|
||||
from vllm.entrypoints.utils import create_error_response
|
||||
from vllm.entrypoints.serve.disagg.protocol import (
|
||||
GenerateRequest,
|
||||
MultiModalFeatures,
|
||||
PlaceholderRangeInfo,
|
||||
)
|
||||
from vllm.entrypoints.utils import (
|
||||
create_error_response,
|
||||
get_max_tokens,
|
||||
)
|
||||
from vllm.inputs.data import ProcessorInputs, PromptType, SingletonPrompt, TokensPrompt
|
||||
from vllm.logger import init_logger
|
||||
from vllm.multimodal.inputs import MultiModalHashes, MultiModalPlaceholderDict
|
||||
from vllm.parser import ParserManager
|
||||
from vllm.renderers import BaseRenderer, merge_kwargs
|
||||
from vllm.renderers.inputs.preprocess import parse_model_prompt, prompt_to_seq
|
||||
from vllm.renderers.inputs.preprocess import (
|
||||
extract_prompt_components,
|
||||
extract_prompt_len,
|
||||
parse_model_prompt,
|
||||
prompt_to_seq,
|
||||
)
|
||||
from vllm.tokenizers import TokenizerLike
|
||||
from vllm.tool_parsers import ToolParser
|
||||
from vllm.utils import random_uuid
|
||||
from vllm.utils.mistral import is_mistral_tokenizer
|
||||
from vllm.utils.mistral import mt as _mt
|
||||
|
||||
@@ -83,10 +98,18 @@ class OpenAIServingRender:
|
||||
self.supports_browsing = False
|
||||
self.supports_code_interpreter = False
|
||||
|
||||
self.default_sampling_params = model_config.get_diff_sampling_param()
|
||||
mc = model_config
|
||||
self.override_max_tokens = (
|
||||
self.default_sampling_params.get("max_tokens")
|
||||
if mc.generation_config not in ("auto", "vllm")
|
||||
else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
|
||||
)
|
||||
|
||||
async def render_chat_request(
|
||||
self,
|
||||
request: ChatCompletionRequest,
|
||||
) -> tuple[list[ConversationMessage], list[ProcessorInputs]] | ErrorResponse:
|
||||
) -> GenerateRequest | ErrorResponse:
|
||||
"""Validate the model and preprocess a chat completion request.
|
||||
|
||||
This is the authoritative implementation used directly by the
|
||||
@@ -96,7 +119,56 @@ class OpenAIServingRender:
|
||||
if error_check_ret is not None:
|
||||
logger.error("Error with model %s", error_check_ret)
|
||||
return error_check_ret
|
||||
return await self.render_chat(request)
|
||||
|
||||
if request.use_beam_search:
|
||||
return self.create_error_response(
|
||||
"Beam search is not supported by the render endpoint"
|
||||
)
|
||||
|
||||
result = await self.render_chat(request)
|
||||
if isinstance(result, ErrorResponse):
|
||||
return result
|
||||
|
||||
_, engine_prompts = result
|
||||
|
||||
if len(engine_prompts) != 1:
|
||||
return self.create_error_response(
|
||||
f"Expected exactly 1 engine prompt, got {len(engine_prompts)}"
|
||||
)
|
||||
|
||||
engine_prompt = engine_prompts[0]
|
||||
|
||||
prompt_components = extract_prompt_components(self.model_config, engine_prompt)
|
||||
token_ids = prompt_components.token_ids
|
||||
if not token_ids:
|
||||
return self.create_error_response("No token_ids rendered")
|
||||
token_ids = list(token_ids)
|
||||
|
||||
input_length = extract_prompt_len(self.model_config, engine_prompt)
|
||||
max_tokens = get_max_tokens(
|
||||
self.model_config.max_model_len,
|
||||
request.max_completion_tokens
|
||||
if request.max_completion_tokens is not None
|
||||
else request.max_tokens,
|
||||
input_length,
|
||||
self.default_sampling_params,
|
||||
self.override_max_tokens,
|
||||
)
|
||||
params = request.to_sampling_params(max_tokens, self.default_sampling_params)
|
||||
|
||||
request_id = f"chatcmpl-{random_uuid()}"
|
||||
|
||||
return GenerateRequest(
|
||||
request_id=request_id,
|
||||
token_ids=token_ids,
|
||||
features=self._extract_mm_features(engine_prompt),
|
||||
sampling_params=params,
|
||||
model=request.model,
|
||||
stream=bool(request.stream),
|
||||
stream_options=(request.stream_options if request.stream else None),
|
||||
cache_salt=request.cache_salt,
|
||||
priority=request.priority,
|
||||
)
|
||||
|
||||
async def render_chat(
|
||||
self,
|
||||
@@ -183,7 +255,7 @@ class OpenAIServingRender:
|
||||
async def render_completion_request(
|
||||
self,
|
||||
request: CompletionRequest,
|
||||
) -> list[ProcessorInputs] | ErrorResponse:
|
||||
) -> list[GenerateRequest] | ErrorResponse:
|
||||
"""Validate the model and preprocess a completion request.
|
||||
|
||||
This is the authoritative implementation used directly by the
|
||||
@@ -192,7 +264,48 @@ class OpenAIServingRender:
|
||||
error_check_ret = await self._check_model(request)
|
||||
if error_check_ret is not None:
|
||||
return error_check_ret
|
||||
return await self.render_completion(request)
|
||||
result = await self.render_completion(request)
|
||||
if isinstance(result, ErrorResponse):
|
||||
return result
|
||||
generate_requests: list[GenerateRequest] = []
|
||||
for engine_prompt in result:
|
||||
prompt_components = extract_prompt_components(
|
||||
self.model_config, engine_prompt
|
||||
)
|
||||
token_ids = prompt_components.token_ids
|
||||
if not token_ids:
|
||||
return self.create_error_response("No token_ids rendered")
|
||||
token_ids = list(token_ids)
|
||||
|
||||
input_length = extract_prompt_len(self.model_config, engine_prompt)
|
||||
max_tokens = get_max_tokens(
|
||||
self.model_config.max_model_len,
|
||||
request.max_tokens,
|
||||
input_length,
|
||||
self.default_sampling_params,
|
||||
self.override_max_tokens,
|
||||
)
|
||||
params = request.to_sampling_params(
|
||||
max_tokens, self.default_sampling_params
|
||||
)
|
||||
|
||||
request_id = f"cmpl-{random_uuid()}"
|
||||
|
||||
generate_requests.append(
|
||||
GenerateRequest(
|
||||
request_id=request_id,
|
||||
token_ids=token_ids,
|
||||
features=self._extract_mm_features(engine_prompt),
|
||||
sampling_params=params,
|
||||
model=request.model,
|
||||
stream=bool(request.stream),
|
||||
stream_options=(request.stream_options if request.stream else None),
|
||||
cache_salt=request.cache_salt,
|
||||
priority=request.priority,
|
||||
)
|
||||
)
|
||||
|
||||
return generate_requests
|
||||
|
||||
async def render_completion(
|
||||
self,
|
||||
@@ -223,6 +336,33 @@ class OpenAIServingRender:
|
||||
|
||||
return engine_prompts
|
||||
|
||||
@staticmethod
|
||||
def _extract_mm_features(
|
||||
engine_prompt: ProcessorInputs,
|
||||
) -> MultiModalFeatures | None:
|
||||
"""Extract multimodal metadata from a rendered engine prompt.
|
||||
|
||||
Returns ``None`` for text-only prompts.
|
||||
"""
|
||||
if engine_prompt.get("type") != "multimodal":
|
||||
return None
|
||||
|
||||
# At this point engine_prompt is a MultiModalInputs TypedDict.
|
||||
mm_hashes: MultiModalHashes = engine_prompt["mm_hashes"] # type: ignore[typeddict-item]
|
||||
raw_placeholders: MultiModalPlaceholderDict = engine_prompt["mm_placeholders"] # type: ignore[typeddict-item]
|
||||
|
||||
mm_placeholders = {
|
||||
modality: [
|
||||
PlaceholderRangeInfo(offset=p.offset, length=p.length) for p in ranges
|
||||
]
|
||||
for modality, ranges in raw_placeholders.items()
|
||||
}
|
||||
|
||||
return MultiModalFeatures(
|
||||
mm_hashes=mm_hashes,
|
||||
mm_placeholders=mm_placeholders,
|
||||
)
|
||||
|
||||
def _make_request_with_harmony(
|
||||
self,
|
||||
request: ChatCompletionRequest,
|
||||
|
||||
Reference in New Issue
Block a user