f32

app.py

@@ -230,13 +230,13 @@ def run_edit(audio_file, caption, num_steps, guidance_scale, guidance_rescale, s
         
         # 这一步将模型送入显卡
         def safe_move_model(m, dev):
-            logger.info("🛡️ Moving model layer by layer to avoid RAM spike...")
+            logger.info("🛡️ Moving model to GPU in FP32...")
             for name, child in m.named_children():
-                # 逐层搬运：CPU内存释放一点 -> GPU显存增加一点
-                child.to(dev, dtype=torch.float16)
-                logger.info(f"Moving {name} to GPU...")
-            m.to(dev, dtype=torch.float16)
+                child.to(dev, dtype=torch.float32)
+                logger.info(f"Moving {name} to GPU (fp32)...")
+            m.to(dev, dtype=torch.float32)
             return m
+
             
         model = safe_move_model(model, device)
         model.eval()
@@ -257,7 +257,7 @@ def run_edit(audio_file, caption, num_steps, guidance_scale, guidance_rescale, s
         torch.manual_seed(int(seed))
         np.random.seed(int(seed))
         
-        wav = load_and_process_audio(audio_file, target_sr).to(device, dtype=torch.float16)
+        wav = load_and_process_audio(audio_file, target_sr).to(device, dtype=torch.float32)
         
         batch = {
             "audio_id": [Path(audio_file).stem],
@@ -272,7 +272,7 @@ def run_edit(audio_file, caption, num_steps, guidance_scale, guidance_rescale, s
         
         logger.info("Inference running...")
         t0 = time.time()
-        with torch.no_grad(), torch.autocast("cuda", dtype=torch.float16):
+        with torch.no_grad():
             out = model.inference(scheduler=scheduler, **batch)