again

requirements.txt

@@ -1,29 +1,24 @@
-# Pull CUDA 12.1 wheels for Torch
---extra-index-url https://download.pytorch.org/whl/cu121
+# 1) Do NOT pin torch/torchaudio here — use the CUDA builds that come with the image
 
-######## Core DL stack (Torch must match torchaudio)
-torch==2.5.1+cu121
-torchaudio==2.5.1+cu121
-# Provide NVIDIA runtime libs expected by recent Torch wheels
-nvidia-cudnn-cu12==9.1.0.70
-nvidia-cublas-cu12>=12.1.3
+transformers==4.48.0
+# Removed flash-attention since faster-whisper handles this internally
+# https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.0.8/flash_attn-2.7.4.post1+cu126torch2.4-cp310-cp310-linux_x86_64.whl
+pydantic==2.10.6
 
-######## ASR stacks
-# Newer WhisperX works with newer faster-whisper and requires CT2>=4.5
+# 2) Main whisper stack — WhisperX (with compatible FW/CT2 for cuDNN 8 era)
 whisperx==3.7.0
 faster-whisper==1.2.0
 ctranslate2==4.5.0
+nvidia-cudnn-cu12==9.1.0.70
+# (no 'torch' line — keep the image’s CUDA wheel)
 
-######## Transformers & validation
-transformers==4.48.0
-pydantic==2.10.6
-
-######## Diarization (optional)
-pyannote.audio>=3.3.1
+# Optional: stable-ts for regroup/edit post-processing
+stable-ts>=2.13.3
 
-######## App + utils
+# 3) Extra libs your app really needs
 gradio==5.0.1
 spaces>=0.19.0
+pyannote.audio>=3.3.1
 pandas>=1.5.0
 numpy>=1.24.0
 librosa>=0.10.0
@@ -31,4 +26,7 @@ soundfile>=0.12.0
 ffmpeg-python>=0.2.0
 requests>=2.28.0
 webrtcvad>=2.0.10
-boto3
+boto3
+
+# Provide NVIDIA runtime libs expected by the image’s Torch (cuDNN 8)
+nvidia-cublas-cu12>=12.1.3