Update app.py

app.py

@@ -1,8 +1,10 @@
 import os
 
-# Ensure environment variables are set before loading transformers
+# Set environment variables VERY early, before HF or Transformers are imported:
 os.environ["HF_HOME"] = "/tmp/hf_cache"
 os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
+os.environ["HF_DATASETS_CACHE"] = "/tmp/hf_cache"
+os.environ["XDG_CACHE_HOME"] = "/tmp"
 
 from flask import Flask, request, jsonify
 import requests
@@ -12,7 +14,7 @@ from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
 app = Flask(__name__)
 
-# Use a smaller model for faster CPU loading
+# Use a smaller model for CPU
 model_id = "openai/whisper-base"
 processor = WhisperProcessor.from_pretrained(model_id)
 model = WhisperForConditionalGeneration.from_pretrained(model_id)
@@ -23,16 +25,16 @@ model.to(device)
 forced_decoder_ids = processor.get_decoder_prompt_ids(language="he", task="transcribe")
 
 def transcribe_audio(audio_url):
-    # 1) Download the file to /tmp
+    # 1) Download audio file to /tmp
     response = requests.get(audio_url)
     audio_path = "/tmp/temp_audio.wav"
     with open(audio_path, "wb") as f:
         f.write(response.content)
 
-    # 2) Load audio with librosa
+    # 2) Load with librosa
     waveform, sr = librosa.load(audio_path, sr=16000)
 
-    # 3) Truncate to 1 hour max
+    # 3) Truncate to 1 hour
     max_duration_sec = 3600
     waveform = waveform[:sr * max_duration_sec]
 
@@ -43,18 +45,15 @@ def transcribe_audio(audio_url):
 
     partial_text = ""
     for chunk in chunks:
-        # Preprocess chunk
         inputs = processor(chunk, sampling_rate=16000, return_tensors="pt", padding=True)
         input_features = inputs.input_features.to(device)
 
-        # Transcribe chunk
         with torch.no_grad():
             predicted_ids = model.generate(
-                input_features, 
+                input_features,
                 forced_decoder_ids=forced_decoder_ids
             )
 
-        # Convert IDs back to text
         transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
         partial_text += transcription + "\n"
 
@@ -67,10 +66,8 @@ def transcribe_endpoint():
     if not audio_url:
         return jsonify({"error": "Missing 'audio_url' in request"}), 400
 
-    # Perform transcription
     transcription = transcribe_audio(audio_url)
     return jsonify({"transcription": transcription})
 
 if __name__ == '__main__':
-    # Run the Flask app on port 7860
     app.run(host="0.0.0.0", port=7860)