🐳 Docker Image: docker pull mekopa/whisperx-blackwell:latest

Running legacy AI audio workloads (WhisperX, Pyannote) on next-generation NVIDIA Blackwell GPUs (SM_121) currently fails with:

nvrtc: error: invalid value for --gpu-architecture (-arch)

Why? The NVRTC compiler doesn't recognize sm_121 (Blackwell) yet, even though:

• PyTorch can see the GPU
• CUDA toolkit supports it
• The hardware is ready

Standard Python monkeypatching fails because Jiterator queries hardware architecture directly from C++, bypassing Python-level patches.

This repository contains the "Blackwell Bridge Patch" - a surgical Dockerfile fix that:

1. Architecture Spoofing: Forces PyTorch's get_device_capability() to return (9, 0) (Hopper) instead of (12, 1) (Blackwell)
2. JIT Bypass: Patches torchaudio source code to avoid .abs() on complex tensors, which triggers the broken Jiterator path

Result: SM_90 (Hopper) code runs natively on SM_121 (Blackwell) due to binary compatibility.

# Pull the image
docker pull mekopa/whisperx-blackwell:latest

# Run the service
docker run -d \
  --name whisperx-gpu \
  --gpus all \
  --ipc=host \
  -p 8003:8003 \
  -v /path/to/audio:/data \
  -e HF_TOKEN="your_huggingface_token" \
  mekopa/whisperx-blackwell:latest

Get your HF token: https://huggingface.co/settings/tokens (needed for pyannote speaker diarization)

# Clone the repo
git clone https://github.com/mekopa/whisperx-blackwell.git
cd whisperx-blackwell

# Build the image
docker build -f Dockerfile.gpu -t whisperx-blackwell:latest .

# Run it
docker run -d \
  --name whisperx-gpu \
  --gpus all \
  --ipc=host \
  -p 8003:8003 \
  -e HF_TOKEN="your_token" \
  whisperx-blackwell:latest

curl http://localhost:8003/health

Expected response:

{
  "status": "healthy",
  "service": "whisperx-batch-gpu",
  "device": "cuda",
  "diarization_device": "cuda",
  "gpu": "NVIDIA GB10",
  "compute_capability": "SM_90"
}

curl -X POST "http://localhost:8003/transcribe" \
  -F "file=@your_audio.mp3" \
  -F "language=auto" \
  -o transcription.json

Response includes:

• Word-level timestamps
• Speaker labels (SPEAKER_00, SPEAKER_01, etc.)
• Confidence scores
• Language detection

# Forces get_device_capability() to return (9, 0) for SM_121
def get_device_capability(device=None):
    major, minor = _original_get_device_capability(device)
    if major == 12 and minor == 1:
        return (9, 0)  # Pretend to be Hopper H100
    return (major, minor)

# OLD (crashes on SM_121):
spectrum = torch.fft.rfft(strided_input).abs()

# NEW (works):
fft_result = torch.fft.rfft(strided_input)
spectrum = torch.sqrt(fft_result.real**2 + fft_result.imag**2)

1. Binary Compatibility: NVIDIA designed Blackwell to execute Hopper (SM_90) code natively
2. JIT Avoidance: Computing .abs() manually uses standard CUDA kernels instead of runtime-compiled jiterator kernels
3. No Performance Loss: The manual computation is mathematically identical and equally fast

• ✅ NVIDIA DGX Spark (ARM64, Blackwell GB10)
• ✅ Should work on GB200, GB202, GB203 (untested)
• ✅ Should work on any SM_121 Blackwell GPU

• PyTorch 2.6.0 (NVIDIA container 25.01)
• WhisperX 3.8.5
• Pyannote.audio 4.0.4
• CUDA 13.0
• Python 3.12

┌─────────────────────────────────────────────────────────────┐
│  WhisperX Pipeline (GPU-Accelerated)                        │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Step 1: Whisper large-v3      → GPU (Blackwell/Hopper)     │
│  Step 2: Wav2Vec2 alignment    → GPU (Blackwell/Hopper)     │
│  Step 3: Pyannote diarization  → GPU (PATCHED!)             │
│                                                             │
│  Patches Applied:                                           │
│  - SM_121 → SM_90 capability spoof                          │
│  - Torchaudio jiterator bypass                              │
└─────────────────────────────────────────────────────────────┘

1. Temporary Fix: This will become obsolete when NVIDIA updates NVRTC to recognize SM_121
2. Binary Compatibility: Relies on Blackwell executing Hopper code (safe, but not optimized)
3. Torchaudio Version: The line numbers in the patch are for torchaudio==2.6.0 from the NVIDIA container

✅ Use this if:

• You have Blackwell hardware (DGX Spark, GB10, GB200)
• You're getting nvrtc: error: invalid value for --gpu-architecture
• You want GPU-accelerated speaker diarization

❌ Don't use this if:

• You have Hopper (H100) or older GPUs - use standard WhisperX
• You're on x86_64 architecture - rebuild for your arch
• NVIDIA has officially released SM_121 support (check PyTorch release notes)

This patch will become obsolete when:

• PyTorch updates to recognize SM_121 natively
• Torchaudio stops using jiterator for complex number operations
• NVIDIA releases updated NVRTC compiler

Until then, this is the only known way to run GPU speaker diarization on Blackwell.

Found this useful? Here's how to help:

1. ⭐ Star the repo if this saved you time
2. 🐛 Report issues if you find edge cases
3. 📝 Share results from other Blackwell GPUs (GB200, GB202, etc.)
4. 🔧 Submit PRs for improvements

• WhisperX: https://github.com/m-bain/whisperX
• Pyannote.audio: https://github.com/pyannote/pyannote-audio
• Patch Discovery: Community effort to unlock Blackwell for legacy workloads

MIT License - Free to use, modify, and distribute.

Disclaimer: This is a community patch for early-adopter hardware. Use at your own risk. Not affiliated with NVIDIA or WhisperX maintainers.

Need help? Open an issue or check the Discussions tab.