Upload app.py with huggingface_hub

app.py

@@ -14,13 +14,45 @@ subprocess.run(
 )
 print(f"After pip install: cuda_available={torch.cuda.is_available()}")
 
+# Test: just import chatterbox, don't load model
 from chatterbox_flash import ChatterboxFlashTTS
 print(f"After chatterbox import: cuda_available={torch.cuda.is_available()}")
 
-# Load model on CPU
-tts = ChatterboxFlashTTS.from_pretrained("ResembleAI/chatterbox-flash", device="cpu", dtype=torch.bfloat16)
-print(f"After model load on CPU: cuda_available={torch.cuda.is_available()}")
-print("Model loaded successfully.")
+# Test: load just the voice encoder (small model)
+from chatterbox.models.ve.model import VoiceEncoder
+from safetensors.torch import load_file
+import tempfile, os as _os
+from huggingface_hub import hf_hub_download
+
+print("Downloading ve.safetensors...")
+ve_path = hf_hub_download("ResembleAI/chatterbox-flash", "ve.safetensors")
+ve = VoiceEncoder()
+ve.load_state_dict(load_file(ve_path))
+ve.to("cpu").eval()
+print(f"After ve load: cuda_available={torch.cuda.is_available()}")
+
+# Test: load just the t3 model
+from chatterbox_flash.t3 import ChatterboxFlashT3
+print("Downloading t3_flash.safetensors...")
+t3_path = hf_hub_download("ResembleAI/chatterbox-flash", "t3_flash.safetensors")
+t3 = ChatterboxFlashT3(drf_block_size=16)
+t3_state = load_file(t3_path)
+if "model" in t3_state:
+    t3_state = t3_state["model"][0]
+t3.load_state_dict(t3_state)
+t3.to(device="cpu", dtype=torch.bfloat16).eval()
+print(f"After t3 load: cuda_available={torch.cuda.is_available()}")
+
+# Test: load s3gen
+from chatterbox.models.s3gen.s3gen import S3Gen
+print("Downloading s3gen.safetensors...")
+s3gen_path = hf_hub_download("ResembleAI/chatterbox-flash", "s3gen.safetensors")
+s3gen = S3Gen(meanflow=True)
+s3gen.load_state_dict(load_file(s3gen_path), strict=False)
+s3gen.to("cpu").eval()
+print(f"After s3gen load: cuda_available={torch.cuda.is_available()}")
+
+print("All models loaded on CPU. Testing GPU...")
 
 import gradio as gr