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 7, 07:45 UTC
[Bugfix] Add missing dynamic_arg_dims for Qwen3-ASR torch.compile (#35869)
Signed-off-by: Nathan Price <nathan@abridge.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Nathan Price committed
Mar 4, 2026 at 02:29 UTC
36bf2131816eb6d75fcf22adb3734850e90cfb98
1 file changed
-2
vllm/model_executor/models/qwen3_asr_realtime.py
-2
@@ -22,7 +22,6 @@ from collections.abc import AsyncGenerator, Mapping
22
import numpy as np
23
import torch
24
25
-from vllm.compilation.decorators import support_torch_compile
25
from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig
26
from vllm.inputs.data import PromptType, TokensPrompt
27
from vllm.logger import init_logger
@@ -177,7 +176,6 @@ class Qwen3ASRRealtimeMultiModalProcessor(Qwen3ASRMultiModalProcessor):
176
info=Qwen3ASRProcessingInfo,
177
dummy_inputs=Qwen3ASRDummyInputsBuilder,
178
)
180
-@support_torch_compile
179
class Qwen3ASRRealtimeGeneration(Qwen3ASRForConditionalGeneration, SupportsRealtime):
180
realtime_max_tokens = 64
181