[Bugfix] Fix SHM cache initialization (#26427)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-10-09 02:48:04 -07:00
committed by GitHub
parent dc7976dd9f
commit 4bdf7ac593
30 changed files with 357 additions and 417 deletions
+199 -17
View File
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import asyncio
import json
import sys
import time
@@ -15,17 +16,13 @@ from pydantic import BaseModel, ConfigDict, Field
from starlette.datastructures import Headers
from typing_extensions import TypeIs
from vllm.entrypoints.utils import _validate_truncation_size
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.processor import Processor
if sys.version_info >= (3, 12):
from typing import TypedDict
else:
from typing_extensions import TypedDict
import vllm.envs as envs
from vllm.config import ModelConfig
from vllm.beam_search import BeamSearchSequence, create_sort_beams_key_function
from vllm.engine.protocol import EngineClient
from vllm.entrypoints.chat_utils import (
ChatCompletionMessageParam,
@@ -68,9 +65,14 @@ from vllm.entrypoints.openai.protocol import (
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
from vllm.entrypoints.openai.tool_parsers import ToolParser
from vllm.entrypoints.renderer import BaseRenderer, CompletionRenderer, RenderConfig
from vllm.entrypoints.utils import _validate_truncation_size
from vllm.inputs.data import PromptType
from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
from vllm.inputs.parse import PromptComponents, get_prompt_components
from vllm.inputs.parse import (
PromptComponents,
get_prompt_components,
is_explicit_encoder_decoder_prompt,
)
from vllm.logger import init_logger
from vllm.logprobs import Logprob, PromptLogprobs
from vllm.lora.request import LoRARequest
@@ -78,7 +80,7 @@ from vllm.multimodal import ( # noqa: F401 - Required to resolve Pydantic error
MultiModalDataDict,
MultiModalUUIDDict,
)
from vllm.outputs import PoolingRequestOutput, RequestOutput
from vllm.outputs import CompletionOutput, PoolingRequestOutput, RequestOutput
from vllm.pooling_params import PoolingParams
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.tracing import (
@@ -89,11 +91,13 @@ from vllm.tracing import (
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
from vllm.utils import (
AsyncMicrobatchTokenizer,
collect_from_async_generator,
is_list_of,
make_async,
merge_async_iterators,
random_uuid,
)
from vllm.v1.engine import EngineCoreRequest
logger = init_logger(__name__)
@@ -240,7 +244,6 @@ class OpenAIServing:
def __init__(
self,
engine_client: EngineClient,
model_config: ModelConfig,
models: OpenAIServingModels,
*,
request_logger: Optional[RequestLogger],
@@ -251,8 +254,6 @@ class OpenAIServing:
super().__init__()
self.engine_client = engine_client
self.model_config = model_config
self.max_model_len = model_config.max_model_len
self.models = models
@@ -268,12 +269,194 @@ class OpenAIServing:
self._async_tokenizer_pool: dict[AnyTokenizer, AsyncMicrobatchTokenizer] = {}
self.log_error_stack = log_error_stack
async def _get_processor(self) -> Processor:
if not hasattr(self, "_processor"):
vllm_config = await self.engine_client.get_vllm_config()
self._processor = Processor(vllm_config)
self.processor = self.models.processor
self.io_processor = self.models.io_processor
self.model_config = self.models.model_config
self.max_model_len = self.model_config.max_model_len
return self._processor
async def beam_search(
self,
prompt: PromptType,
request_id: str,
params: BeamSearchParams,
lora_request: Optional[LoRARequest] = None,
) -> AsyncGenerator[RequestOutput, None]:
beam_width = params.beam_width
max_tokens = params.max_tokens
ignore_eos = params.ignore_eos
temperature = params.temperature
length_penalty = params.length_penalty
include_stop_str_in_output = params.include_stop_str_in_output
processor = self.processor
tokenizer = processor.tokenizer
if tokenizer is None:
raise ValueError(
"You cannot use beam search when `skip_tokenizer_init` is True"
)
eos_token_id: int = tokenizer.eos_token_id # type: ignore
if is_explicit_encoder_decoder_prompt(prompt):
raise NotImplementedError
else:
processed_inputs = processor.input_preprocessor._prompt_to_llm_inputs(
prompt
)
if processed_inputs["type"] == "embeds":
raise NotImplementedError
# This is a workaround to fix multimodal beam search; this is a
# bandaid fix for 2 small problems:
# 1. Multi_modal_data on the processed_inputs currently resolves to
# `None`.
# 2. preprocessing above expands the multimodal placeholders. However,
# this happens again in generation, so the double expansion causes
# a mismatch.
# TODO - would be ideal to handle this more gracefully.
prompt_text: Optional[str]
prompt_token_ids: list[int]
multi_modal_data: Optional[MultiModalDataDict]
if isinstance(prompt, str):
prompt_text = prompt
prompt_token_ids = []
multi_modal_data = None
else:
prompt_text = prompt.get("prompt") # type: ignore
prompt_token_ids = prompt.get("prompt_token_ids", []) # type: ignore
multi_modal_data = prompt.get("multi_modal_data") # type: ignore
mm_processor_kwargs: Optional[dict[str, Any]] = processed_inputs.get(
"mm_processor_kwargs"
) # type: ignore
tokenized_length = len(prompt_token_ids)
sort_beams_key = create_sort_beams_key_function(eos_token_id, length_penalty)
beam_search_params = SamplingParams(
logprobs=2 * beam_width,
max_tokens=1,
temperature=temperature,
)
all_beams = [
BeamSearchSequence(
tokens=prompt_token_ids,
cum_logprob=0,
logprobs=[],
multi_modal_data=multi_modal_data,
mm_processor_kwargs=mm_processor_kwargs,
lora_request=lora_request,
)
]
completed = []
for _ in range(max_tokens):
prompts_batch, lora_req_batch = zip(
*[
(
EngineTokensPrompt(
prompt_token_ids=beam.tokens,
multi_modal_data=beam.multi_modal_data,
mm_processor_kwargs=beam.mm_processor_kwargs,
),
beam.lora_request,
)
for beam in all_beams
]
)
tasks = []
request_id_batch = f"{request_id}-{random_uuid()}"
for i, (individual_prompt, lora_req) in enumerate(
zip(prompts_batch, lora_req_batch)
):
request_id_item = f"{request_id_batch}-beam-{i}"
task = asyncio.create_task(
collect_from_async_generator(
self.engine_client.generate(
individual_prompt,
beam_search_params,
request_id_item,
lora_request=lora_req,
)
)
)
tasks.append(task)
output = [x[0] for x in await asyncio.gather(*tasks)]
new_beams = []
for i, current_beam in enumerate(all_beams):
result = output[i]
if result.outputs[0].logprobs is not None:
logprobs = result.outputs[0].logprobs[0]
for token_id, logprob_obj in logprobs.items():
if token_id == eos_token_id and not ignore_eos:
completed.append(
BeamSearchSequence(
tokens=current_beam.tokens + [token_id]
if include_stop_str_in_output
else current_beam.tokens,
logprobs=current_beam.logprobs + [logprobs],
cum_logprob=current_beam.cum_logprob
+ logprob_obj.logprob,
finish_reason="stop",
stop_reason=eos_token_id,
)
)
else:
new_beams.append(
BeamSearchSequence(
tokens=current_beam.tokens + [token_id],
logprobs=current_beam.logprobs + [logprobs],
lora_request=current_beam.lora_request,
cum_logprob=current_beam.cum_logprob
+ logprob_obj.logprob,
multi_modal_data=current_beam.multi_modal_data,
mm_processor_kwargs=current_beam.mm_processor_kwargs,
)
)
sorted_beams = sorted(new_beams, key=sort_beams_key, reverse=True)
all_beams = sorted_beams[:beam_width]
completed.extend(all_beams)
sorted_completed = sorted(completed, key=sort_beams_key, reverse=True)
best_beams = sorted_completed[:beam_width]
for beam in best_beams:
if beam.tokens[-1] == eos_token_id and not ignore_eos:
# Skip the eos token in the text.
tokens = beam.tokens[tokenized_length:-1]
else:
tokens = beam.tokens[tokenized_length:]
beam.text = tokenizer.decode(tokens)
yield RequestOutput(
request_id=request_id,
prompt=prompt_text,
outputs=[
CompletionOutput(
text=beam.text, # type: ignore
cumulative_logprob=beam.cum_logprob,
token_ids=beam.tokens[tokenized_length:],
index=i,
logprobs=beam.logprobs,
finish_reason=beam.finish_reason
if beam.finish_reason is not None
else "length",
stop_reason=beam.stop_reason,
)
for (i, beam) in enumerate(best_beams)
],
finished=True,
prompt_token_ids=prompt_token_ids,
prompt_logprobs=None,
)
def _get_renderer(self, tokenizer: Optional[AnyTokenizer]) -> BaseRenderer:
"""
@@ -938,8 +1121,7 @@ class OpenAIServing:
self.max_model_len, params.truncate_prompt_tokens, tokenization_kwargs
)
processor = await self._get_processor()
engine_request = processor.process_inputs(
engine_request = self.processor.process_inputs(
request_id,
engine_prompt,
params,