#7895 ASR Error [Qwen-ASR(Local)] CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasGemmEx( handle, opa, opb, m,

202.150**5 Posted at: 1 hour ago

ASR Error [Qwen-ASR(Local)] CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasGemmEx( handle, opa, opb, m, n, k, alpha_ptr, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, beta_ptr, c, CUDA_R_16F, ldc, compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)
Traceback (most recent call last):
File "videotrans\task\job.py", line 106, in run
File "videotrans\task\_speech2text.py", line 153, in recogn
File "videotrans\recognition\__init__.py", line 240, in run
File "videotrans\recognition\_base.py", line 142, in run
File "videotrans\recognition\_qwenasrlocal.py", line 52, in _exec
File "videotrans\configure\_base.py", line 290, in _new_process
RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasGemmEx( handle, opa, opb, m, n, k, alpha_ptr, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, beta_ptr, c, CUDA_R_16F, ldc, compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)
TaskCfgSTT(is_cuda=True, uuid='0abe6da7e9', cache_folder='E:/ytrans/tmp/22132/0abe6da7e9', target_dir='E:/ytrans/output/recogn', source_language=None, source_language_code=None, source_sub=None, source_wav=None, source_wav_output=None, target_language=None, target_language_code=None, target_sub='E:/ytrans/output/recogn/16.srt', target_wav=None, target_wav_output=None, name='E:/2023/personal/inspo2/Redhong/new subs/Redhong-S4/16.m4a', noextname='16', basename='16.m4a', ext='m4a', dirname='E:/2023/personal/inspo2/Redhong/new subs/Redhong-S4', shound_del_name=None, detect_language='ko', recogn_type=2, model_name='0.6B', shibie_audio='E:/ytrans/tmp/22132/0abe6da7e9/16-1791500971.9803245.wav', remove_noise=False, enable_diariz=False, nums_diariz=0, rephrase=1, fix_punc=False)

=
system:Windows-10-10.0.26200-SP0
version:v3.97
frozen:True
language:en
root_dir:E:/ytrans

1 Replies

#1 DeepSeek reply at 1 hour ago

This is a CUDA/cuBLAS runtime failure inside the Qwen-ASR local model when running in fp16 (CUDA_R_16F). It is NOT a problem with the input audio file. Common causes: GPU driver / CUDA / cuDNN version mismatch or corruption, insufficient VRAM for fp16 weights, or leftover/corrupted model files.

Note: your log shows version:v3.97, which is far older than the current v4.15. Qwen-ASR in v4.12+ was switched to the official transformers -hf models, so the old build is much more likely to be incompatible with the current CUDA stack. Upgrading is the most reliable fix.

Recommended actions (in order):

Upgrade first (highest priority)
- Download the v4.15 full package, extract to a clean non-system, pure-English path (e.g. D:\pyVideoTrans), then overwrite the v4.15 patch as well. Do NOT extract inside C:\Program Files.
- Restart and retry Qwen-ASR.
Verify CUDA stack
- Requirement: NVIDIA GPU + CUDA Toolkit 12.8+ and cuDNN 9.11+.
- In CMD run nvcc -V and nvidia-smi to check versions. Update the NVIDIA driver to the latest.
- Ensure the CUDA bin and lib folders (and the cuDNN bin/lib folders) are present in the system Path.
Check VRAM
- Qwen-ASR 0.6B in fp16 needs a few GB of free VRAM. Close other GPU-consuming apps, or try the smaller model.
- In Menu -> Tools/Options -> Advanced Options -> ASR Settings, set Compute data type to int8 (or float16 on a healthy GPU) to reduce memory pressure.
Re-download the model (in case the existing files are corrupted/incomplete)
- Delete the local Qwen-ASR model folder, then let it re-download, or download manually:
  - 0.6B: https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf/tree/main -> 软件目录/models/models--Qwen--Qwen3-ASR-0.6B-hf
  - (also needs Qwen3-ForcedAligner-0.6B-hf for segmentation)
Workarounds if the error persists
- Switch to a different ASR channel, e.g. faster-whisper(内置) with large-v3-turbo, which is the default recommended free local option.
- Or disable CUDA (run on CPU) to confirm the issue is GPU-specific.

请查阅相关文档：