[V0 Deprecation] Remove async_output_proc, preemption mode, delay factor (#25334)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-09-21 08:52:32 -07:00
committed by GitHub
parent 26e673fe93
commit 0ff8ebb2d7
15 changed files with 12 additions and 210 deletions
-4
View File
@@ -454,9 +454,6 @@ class VllmConfig:
self.try_verify_and_update_config()
if self.model_config is not None:
self.model_config.verify_async_output_proc(self.parallel_config,
self.speculative_config,
self.device_config)
self.model_config.verify_with_parallel_config(self.parallel_config)
self.model_config.verify_dual_chunk_attention_config(
self.load_config)
@@ -877,7 +874,6 @@ class VllmConfig:
f"served_model_name={self.model_config.served_model_name}, "
f"enable_prefix_caching={self.cache_config.enable_prefix_caching}, "
f"chunked_prefill_enabled={self.scheduler_config.chunked_prefill_enabled}, " # noqa
f"use_async_output_proc={self.model_config.use_async_output_proc}, "
f"pooler_config={self.model_config.pooler_config!r}, "
f"compilation_config={self.compilation_config!r}")
+6 -42
View File
@@ -223,8 +223,6 @@ class ModelConfig:
that this name(s) will also be used in `model_name` tag content of
prometheus metrics, if multiple names provided, metrics tag will take the
first one."""
use_async_output_proc: bool = True
"""Whether to use async output processor."""
config_format: Union[str, ConfigFormat] = "auto"
"""The format of the model config to load:\n
- "auto" will try to load the config in hf format if available else it
@@ -1119,37 +1117,6 @@ class ModelConfig:
raise ValueError("please set VLLM_ATTENTION_BACKEND to "
f"{STR_DUAL_CHUNK_FLASH_ATTN_VAL}")
def verify_async_output_proc(self, parallel_config, speculative_config,
device_config) -> None:
if not self.use_async_output_proc:
# Nothing to check
return
if parallel_config.pipeline_parallel_size > 1:
self.use_async_output_proc = False
return
# Reminder: Please update docs/features/compatibility_matrix.md
# If the feature combo become valid
from vllm.platforms import current_platform
if not current_platform.is_async_output_supported(self.enforce_eager):
self.use_async_output_proc = False
return
if envs.VLLM_USE_RAY_SPMD_WORKER:
self.use_async_output_proc = False
return
# Async postprocessor is not necessary for pooling models
# since there is no token generation
if self.runner_type == "pooling":
self.use_async_output_proc = False
# Reminder: Please update docs/features/compatibility_matrix.md
# If the feature combo become valid
if speculative_config:
self.use_async_output_proc = False
def verify_with_parallel_config(
self,
parallel_config: ParallelConfig,
@@ -1173,15 +1140,12 @@ class ModelConfig:
self._verify_with_expert_parallelism()
pipeline_parallel_size = parallel_config.pipeline_parallel_size
if pipeline_parallel_size > 1:
if not self.registry.is_pp_supported_model(self.architectures,
self):
raise NotImplementedError(
"Pipeline parallelism is not supported for this model. "
"Supported models implement the `SupportsPP` interface.")
if self.use_async_output_proc:
self.use_async_output_proc = False
if (pipeline_parallel_size > 1
and not self.registry.is_pp_supported_model(
self.architectures, self)):
raise NotImplementedError(
"Pipeline parallelism is not supported for this model. "
"Supported models implement the `SupportsPP` interface.")
def get_sliding_window(self) -> Optional[int]:
"""Get the sliding window size from the HF text config if present."""
+1 -14
View File
@@ -3,7 +3,7 @@
import hashlib
from dataclasses import field
from typing import Any, Literal, Optional, Union
from typing import Any, Literal, Union
from pydantic import SkipValidation, model_validator
from pydantic.dataclasses import dataclass
@@ -18,7 +18,6 @@ from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS,
logger = init_logger(__name__)
RunnerType = Literal["generate", "pooling", "draft"]
PreemptionMode = Literal["swap", "recompute"]
SchedulerPolicy = Literal["fcfs", "priority"]
@@ -78,10 +77,6 @@ class SchedulerConfig:
3. more than one value (e.g. 1 2 128) is provided, then the capture list
will follow the provided list."""
delay_factor: float = 0.0
"""Apply a delay (of delay factor multiplied by previous
prompt latency) before scheduling next prompt."""
enable_chunked_prefill: SkipValidation[bool] = None # type: ignore
"""If True, prefill requests can be chunked based
on the remaining max_num_batched_tokens."""
@@ -103,14 +98,6 @@ class SchedulerConfig:
NOTE: This is not currently configurable. It will be overridden by
max_num_batched_tokens in case max multimodal embedding size is larger."""
preemption_mode: Optional[PreemptionMode] = None
"""Whether to perform preemption by swapping or
recomputation. If not specified, we determine the mode as follows:
We use recomputation by default since it incurs lower overhead than
swapping. However, when the sequence group has multiple sequences
(e.g., beam search), recomputation is not currently supported. In
such a case, we use swapping instead."""
send_delta_data: bool = False
"""Private API. If used, scheduler sends delta data to
workers instead of an entire data. It should be enabled only