Request: Update to CUDA 12.8 / PyTorch 2.7+ for Blackwell GPU support (RTX 50 series)

[maarteek]

Summary

RTX 50 series GPUs (Blackwell architecture, sm_120 compute capability) are incompatible with the current Whishper Docker image, which ships with PyTorch 2.4.1 + CUDA 11.8.

Error

When attempting GPU transcription on RTX 5070 Ti:

RuntimeError: cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED

The underlying issue is that CUDA 11.8 only supports up to sm_90 (Ampere/Ada), while Blackwell requires CUDA 12.8+.

Upstream Fix Available

The WhisperX library has already addressed this in PR #1182, merged October 2025:

• PyTorch upgraded to 2.7.1
• CUDA upgraded to 12.8 (cu128 index)
• CTranslate2 upgraded to 4.5.0+
• Confirmed working with sm_120 (Blackwell)

Requested Changes

Update the Whishper Docker image to use:

• Base image: nvidia/cuda:12.8.0-cudnn9-runtime-ubuntu22.04 (or similar)
• PyTorch 2.7.1+ with CUDA 12.8: pip install torch --index-url https://download.pytorch.org/whl/cu128
• CTranslate2 4.5.0+

Environment

• GPU: NVIDIA RTX 5070 Ti (Blackwell, sm_120)
• Host CUDA: 12.9
• OS: WSL2 on Windows 11
• Whishper version: latest-gpu

Workaround

Currently using PUBLIC_WHISHPER_PROFILE=cpu which works but is significantly slower.

References

• PyTorch Blackwell tracking issue
• WhisperX CUDA 12.8 PR
• WhisperX Blackwell issue