Update app.py

app.py

@@ -1,14 +1,33 @@
 import gradio as gr
-import tempfile
 import os
+from huggingface_hub import snapshot_download
+from omegaconf import OmegaConf
 from nemo.collections.asr.models import ASRModel
 
-# Define the model name (using the original author's repo or your fork)
 MODEL_NAME = "trysem/stt_ml_fastconformer_ctc_med_punct"
 
-print(f"Downloading and loading {MODEL_NAME}...")
-# Because we are on NeMo 1.23, ASRModel will successfully resolve the class
-model = ASRModel.from_pretrained(model_name=MODEL_NAME)
+print("1. Downloading unzipped model files from Hugging Face...")
+# snapshot_download pulls the raw files (model_config.yaml, model_weights.ckpt) into the cache
+model_dir = snapshot_download(repo_id=MODEL_NAME)
+
+print("2. Patching the configuration in memory...")
+config_path = os.path.join(model_dir, "model_config.yaml")
+config = OmegaConf.load(config_path)
+
+# Allow structural modifications to the config object
+OmegaConf.set_struct(config, False)
+
+# Remove the 'future' PyTorch SDPA keys that NeMo 1.23.0 doesn't recognize
+if 'encoder' in config:
+    config.encoder.pop('use_pytorch_sdpa', None)
+    config.encoder.pop('use_pytorch_sdpa_backends', None)
+
+print("3. Restoring model with patched config (this may take a moment)...")
+# Inject our patched config directly into the restore_from method!
+model = ASRModel.restore_from(
+    restore_path=model_dir,
+    override_config_path=config
+)
 model.eval()
 print("Model loaded successfully!")
 
@@ -16,12 +35,9 @@ def transcribe(audio_path):
     if not audio_path:
         return "Please upload or record audio."
     
-    # NeMo expects a list of paths for transcription
     try:
-        # Transcribe returns a tuple, we want the first element (the text)
         transcription = model.transcribe(paths2audio_files=[audio_path])[0]
         
-        # If it's a batch of 1, extract the single string
         if isinstance(transcription, list):
             return transcription[0]
         return transcription
@@ -49,5 +65,4 @@ with gr.Blocks(title="Malayalam FastConformer ASR") as demo:
     )
 
 if __name__ == "__main__":
-    # 0.0.0.0 is required for Docker containers to expose the port correctly
     demo.launch(server_name="0.0.0.0", server_port=7860)