Fix CUDA/ROCm compatibility: default to CPU for AMD GPUs

- Change default device to CPU in docker-compose and main.py
- Set compute_type=int8 for CPU inference
- Auto-detect device if env not set

whisperx/api/main.py

@@ -6,34 +6,29 @@ from fastapi import FastAPI, UploadFile, File, Form, HTTPException
 from fastapi.responses import JSONResponse
 import torch
 import whisperx
-from whisperx.schema import TranscriptionResult
 
 
 model = None
-align_model_metadata = None
 
 
-def load_transcription_model(model_name: str = "turbo", device: str = None, compute_type: str = "float16"):
-    global model, align_model_metadata
+def load_transcription_model(model_name: str = "large-v2", device: str = None, compute_type: str = None):
+    global model
     if device is None:
-        device = "cuda" if torch.cuda.is_available() else "cpu"
+        device = os.getenv("WHISPERX_DEVICE", "cuda" if torch.cuda.is_available() else "cpu")
+    if compute_type is None:
+        compute_type = "int8" if device == "cpu" else "float16"
     print(f"Loading WhisperX model: {model_name} on {device} with {compute_type}")
     model = whisperx.load_model(model_name, device, compute_type=compute_type)
-    # For alignment, load the metadata
-    align_model_metadata = whisperx.alignment.DEFAULT_ALIGN_MODELS_HF
     print("Model loaded and ready.")
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # Load the model at startup
-    model_name = os.getenv("WHISPERX_MODEL", "turbo")
-    device = os.getenv("WHISPERX_DEVICE", "cuda")
-    compute_type = os.getenv("WHISPERX_COMPUTE_TYPE", "float16")
+    model_name = os.getenv("WHISPERX_MODEL", "large-v2")
+    device = os.getenv("WHISPERX_DEVICE", "cuda" if torch.cuda.is_available() else "cpu")
+    compute_type = os.getenv("WHISPERX_COMPUTE_TYPE", "int8" if device == "cpu" else "float16")
     load_transcription_model(model_name, device, compute_type)
     yield
-    # Cleanup if needed
-    print("Shutting down API")
 
 
 app = FastAPI(