Switch to Docker SDK to use custom Dockerfile with multi-stage caching

README.md

@@ -3,9 +3,8 @@ title: Caribbean Voices - OWSM v3.1 Platform
 emoji: 🎤
 colorFrom: purple
 colorTo: pink
-sdk: gradio
-sdk_version: 5.49.1
-app_file: app.py
+sdk: docker
+app_port: 7860
 pinned: false
 hardware: gpu-a10g-large
 license: mit

requirements.txt

@@ -1,45 +1,13 @@
-# Complete requirements file for Hugging Face Spaces
-# NOTE: HF Spaces may use their own Dockerfile, so this file must be self-contained
-# IMPORTANT: torch MUST be listed before flash-attn (flash-attn needs torch to build)
-
-# Core dependencies
-numpy>=1.24.0
-pandas>=2.0.0
-scikit-learn>=1.3.0
-
-# Audio processing
-librosa>=0.10.0
-soundfile>=0.12.0
-
-# PyTorch - MUST be installed before flash-attn
-# PyTorch 2.2+ with CUDA support (required for Flash Attention on A10G GPUs)
-torch>=2.2.0
-torchaudio>=2.2.0
-# torchcodec required for audio decoding in datasets library
-torchcodec>=0.1.0,<1.0.0
-
-# ML frameworks
-transformers>=4.30.0
-
-# Hugging Face ecosystem
-huggingface_hub>=0.20.0
-datasets>=2.14.0
-
-# UI
-gradio>=5.49.1
-
-# ESPnet dependencies
-espnet>=202301
-espnet_model_zoo>=0.1.0
-
-# Evaluation
+# Remaining dependencies not in requirements-base.txt or requirements-espnet.txt
+# These are installed after base and ESPnet for optimal caching
+# NOTE: torch is installed from requirements-base.txt, so flash-attn can build successfully
 jiwer>=3.0.0
 
 # Build dependencies for Flash Attention compilation
 packaging>=21.0
 ninja>=1.10.0
 
-# Flash Attention - MUST be installed AFTER torch
-# Optimized for NVIDIA A10G GPUs (Ampere architecture)
+# Flash Attention - optimized for NVIDIA A10G GPUs (Ampere architecture)
 # Required for best performance on HF Spaces with GPU
+# NOTE: torch must be installed first (from requirements-base.txt) for flash-attn to build
 flash-attn>=2.3.0