Update app.py

app.py

@@ -4,31 +4,17 @@ import numpy as np
 from transformers import AutoTokenizer
 import onnxruntime
 from huggingface_hub import hf_hub_download
-import os  # Import the 'os' module
+import os
 
 # --- Configuration ---
-repo_id = "Athspi/Gg"  # Your Hugging Face Hub repository ID
-onnx_filename = "mms_tts_eng.onnx"  # Name of the ONNX file
+repo_id = "Athspi/Gg"
+onnx_filename = "mms_tts_eng.onnx"
 sampling_rate = 16000
 
-# --- Download ONNX Model (and handle location) ---
-
-# Option 1: Use the cached path (Recommended)
+# --- Download ONNX Model ---
 onnx_model_path = hf_hub_download(repo_id=repo_id, filename=onnx_filename)
 print(f"ONNX model downloaded to (cache): {onnx_model_path}")
 
-# Option 2: Download to a specific directory (e.g., the current working directory)
-# output_dir = "."  # Current directory
-# onnx_model_path = hf_hub_download(repo_id=repo_id, filename=onnx_filename, cache_dir=output_dir)
-# print(f"ONNX model downloaded to: {onnx_model_path}")
-
-# Option 3:  Download to a custom directory:
-# output_dir = "models"  # Or any directory you want
-# os.makedirs(output_dir, exist_ok=True) # Create directory if it doesn't exist
-# onnx_model_path = hf_hub_download(repo_id=repo_id, filename=onnx_filename, cache_dir=output_dir)
-# print(f"ONNX model downloaded to: {onnx_model_path}")
-
-
 # --- Load Tokenizer ---
 tokenizer = AutoTokenizer.from_pretrained(repo_id)
 
@@ -64,6 +50,7 @@ max_output_length = input_shape[1] * 10
 output_shape = (1, 1, max_output_length)
 output_tensor = torch.empty(output_shape, dtype=torch.float32, device="cpu").contiguous()
 
+# Bind BEFORE clear_binding_outputs
 io_binding.bind_input(
     name=input_meta.name, device_type="cpu", device_id=0,
     element_type=input_type, shape=input_shape, buffer_ptr=input_tensor.data_ptr(),
@@ -72,7 +59,6 @@ io_binding.bind_output(
     name=output_meta.name, device_type="cpu", device_id=0,
     element_type=np.float32, shape=output_shape, buffer_ptr=output_tensor.data_ptr(),
 )
-
 # --- Inference Function ---
 
 def tts_inference_io_binding(text: str):
@@ -85,7 +71,7 @@ def tts_inference_io_binding(text: str):
 
     if current_input_shape[1] > input_tensor.shape[1]:
         input_tensor = torch.empty(current_input_shape, dtype=torch.int64, device="cpu").contiguous()
-        io_binding.bind_input(
+        io_binding.bind_input(  # Re-bind input
             name=input_meta.name, device_type="cpu", device_id=0,
             element_type=input_type, shape=current_input_shape,
             buffer_ptr=input_tensor.data_ptr(),
@@ -97,16 +83,16 @@ def tts_inference_io_binding(text: str):
     if required_output_length > output_tensor.shape[2]:
         output_shape = (1, 1, required_output_length)
         output_tensor = torch.empty(output_shape, dtype=torch.float32, device="cpu").contiguous()
-        io_binding.bind_output(
+        io_binding.bind_output(  # Re-bind output
             name=output_meta.name, device_type="cpu", device_id=0,
             element_type=np.float32, shape=output_shape,
             buffer_ptr=output_tensor.data_ptr(),
         )
 
-    io_binding.clear_binding_outputs()
+    io_binding.clear_binding_outputs()  # Clear outputs *after* binding
     ort_session.run_with_iobinding(io_binding)
     ort_outputs = io_binding.get_outputs()
-    output_data = ort_outputs[0].numpy()
+    output_data = ort_outputs[0].numpy()  # Directly use the bound output
     return (sampling_rate, output_data.squeeze())
 
 # --- Gradio Interface ---