Upload app.py with huggingface_hub

app.py

@@ -29,20 +29,28 @@ def load_model():
         model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False, optimize=False)
         
         # CRITICAL FIX: Force float32 on CPU
-        # VoxCPM2 uses bfloat16 by default, which causes "Dimension out of range" errors
-        # in MiniCPM4's scaled_dot_product_attention on CPU
+        # VoxCPM2 uses bfloat16 by default, which causes dimension/dtype errors on CPU:
+        # 1. "Dimension out of range" in MiniCPM4's scaled_dot_product_attention (bfloat16 SDPA bug)
+        # 2. "mat1 and mat2 must have the same dtype" when model is float32 but inputs are bfloat16
+        # Fix: change config.dtype BEFORE anything creates tensors, then convert model
         if device == "cpu":
             print("Converting model to float32 for CPU compatibility...")
+            # Step 1: Change config dtype so _inference creates float32 tensors
+            if hasattr(model.tts_model, 'config'):
+                old_dtype = model.tts_model.config.dtype
+                model.tts_model.config.dtype = "float32"
+                print(f"  config.dtype: {old_dtype} -> float32")
+            # Step 2: Convert all model parameters and buffers to float32
             model.tts_model = model.tts_model.to(torch.float32)
-            # Also fix KV caches (they are created with config dtype = bfloat16)
+            # Step 3: Fix KV caches (created in __init__ with old dtype)
             if hasattr(model.tts_model, 'base_lm') and hasattr(model.tts_model.base_lm, 'kv_cache'):
                 if model.tts_model.base_lm.kv_cache is not None:
                     model.tts_model.base_lm.kv_cache.kv_cache = model.tts_model.base_lm.kv_cache.kv_cache.to(torch.float32)
-                    print("  base_lm KV cache converted to float32")
+                    print("  base_lm KV cache -> float32")
             if hasattr(model.tts_model, 'residual_lm') and hasattr(model.tts_model.residual_lm, 'kv_cache'):
                 if model.tts_model.residual_lm.kv_cache is not None:
                     model.tts_model.residual_lm.kv_cache.kv_cache = model.tts_model.residual_lm.kv_cache.kv_cache.to(torch.float32)
-                    print("  residual_lm KV cache converted to float32")
+                    print("  residual_lm KV cache -> float32")
             print("Model conversion to float32 complete!")
         
         print("Model loaded successfully!")