A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 19:45 UTC
[Refactor] Remove unused param (#39750)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Apr 21, 2026 at 17:59 UTC
1842447c09224d9161857a03e1cfac33b7701c50
5 files changed
-11
vllm/model_executor/models/qwen2_5_omni_thinker.py
-4
@@ -952,8 +952,6 @@ class Qwen2_5OmniConditionalGenerationMixin:
952
def _process_audio_input(
953
self,
954
audio_input: Qwen2_5OmniAudioFeatureInputs,
955
- audio_hashes: list[str] | None = None,
956
- cached_audio_features: torch.Tensor | None = None,
955
) -> torch.Tensor:
956
input_features = audio_input["input_features"]
957
audio_feature_lengths = audio_input["audio_feature_lengths"]
@@ -990,8 +988,6 @@ class Qwen2_5OmniConditionalGenerationMixin:
988
def _process_video_input(
989
self,
990
video_input: Qwen2_5_VLVideoInputs,
993
- video_hashes: list[str] = None,
994
- cached_video_embeds: torch.Tensor = None,
991
) -> torch.Tensor:
992
if video_input["type"] == "video_embeds":
993
return video_input["video_embeds"].type(self.visual.dtype)
vllm/model_executor/models/qwen3_asr.py
-2
@@ -363,8 +363,6 @@ class Qwen3ASRForConditionalGeneration(
363
def _process_audio_input(
364
self,
365
audio_input: Qwen2_5OmniAudioFeatureInputs,
366
- audio_hashes: list[str] | None = None,
367
- cached_audio_features: torch.Tensor | None = None,
366
) -> torch.Tensor:
367
input_features = audio_input["input_features"]
368
audio_feature_lengths = audio_input["audio_feature_lengths"]
vllm/model_executor/models/qwen3_omni_moe_thinker.py
-2
@@ -1646,8 +1646,6 @@ class Qwen3OmniMoeConditionalGenerationMixin(Qwen2_5OmniConditionalGenerationMix
1646
def _process_audio_input(
1647
self,
1648
audio_input: Qwen2_5OmniAudioFeatureInputs,
1649
- audio_hashes: list[str] | None = None,
1650
- cached_audio_features: torch.Tensor | None = None,
1649
) -> tuple[torch.Tensor, ...]:
1650
input_features = audio_input["input_features"]
1651
audio_feature_lengths = audio_input["audio_feature_lengths"]
vllm/v1/engine/async_llm.py
-2
@@ -77,7 +77,6 @@ class AsyncLLM(EngineClient):
77
log_stats: bool,
78
usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
79
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
80
- use_cached_outputs: bool = False,
80
log_requests: bool = True,
81
start_engine_loop: bool = True,
82
stat_loggers: list[StatLoggerFactory] | None = None,
@@ -95,7 +94,6 @@ class AsyncLLM(EngineClient):
94
log_stats: Whether to log stats.
95
usage_context: Usage context of the LLM.
96
mm_registry: Multi-modal registry.
98
- use_cached_outputs: Whether to use cached outputs.
97
log_requests: Whether to log requests.
98
start_engine_loop: Whether to start the engine loop.
99
stat_loggers: customized stat loggers for the engine.
vllm/v1/engine/llm_engine.py
-1
@@ -56,7 +56,6 @@ class LLMEngine:
56
usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
57
stat_loggers: list[StatLoggerFactory] | None = None,
58
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
59
- use_cached_outputs: bool = False,
59
multiprocess_mode: bool = False,
60
) -> None:
61
self.vllm_config = vllm_config