Update app.py

app.py

@@ -2,28 +2,28 @@ import os
 import gradio as gr
 import torch
 import numpy as np
-from transformers import pipeline
+from transformers import pipeline, AutoTokenizer
 from diffusers import DiffusionPipeline
 from pyannote.audio import Pipeline as PyannotePipeline
-from dia.model import Dia
+from dia.model import DiaConfig, DiaModel, Dia
 from dac.utils import load_model as load_dac_model
+from accelerate import init_empty_weights, load_checkpoint_and_dispatch
 
-# Load HF token and configure multi-GPU sharding
 HF_TOKEN   = os.environ["HF_TOKEN"]
 device_map = "auto"
 
-# 1. Descript Audio Codec (RVQ)
+# RVQ Codec
 rvq = load_dac_model(tag="latest", model_type="44khz")
 rvq.eval()
 if torch.cuda.is_available(): rvq = rvq.to("cuda")
 
-# 2. Voice Activity Detection via Pyannote
+# VAD Pipeline
 vad_pipe = PyannotePipeline.from_pretrained(
     "pyannote/voice-activity-detection",
     use_auth_token=HF_TOKEN
 )
 
-# 3. Ultravox ASR+LLM (generic pipeline)
+# Ultravox Pipeline
 ultravox_pipe = pipeline(
     model="fixie-ai/ultravox-v0_4",
     trust_remote_code=True,
@@ -31,50 +31,51 @@ ultravox_pipe = pipeline(
     torch_dtype=torch.float16
 )
 
-# 4. Audio Diffusion (Diffusers loader)
+# Audio Diffusion
 diff_pipe = DiffusionPipeline.from_pretrained(
     "teticio/audio-diffusion-instrumental-hiphop-256",
     torch_dtype=torch.float16
 ).to("cuda")
 
-# 5. Dia TTS with device sharding
-dia = Dia.from_pretrained(
+# Dia TTS Loading
+config = DiaConfig.from_pretrained("nari-labs/Dia-1.6B")
+with init_empty_weights():
+    base_model = DiaModel(config)
+base_model = load_checkpoint_and_dispatch(
+    base_model,
     "nari-labs/Dia-1.6B",
     device_map=device_map,
-    torch_dtype=torch.float16,
-    trust_remote_code=True
+    dtype=torch.float16
 )
+dia = Dia(base_model, config)
+
+# Save tokenizer for Dia text processing
+tokenizer = AutoTokenizer.from_pretrained("nari-labs/Dia-1.6B")
 
 def process_audio(audio):
     sr, array = audio
     array = array.numpy() if torch.is_tensor(array) else array
 
-    # VAD segmentation
-    _ = vad_pipe({"waveform": torch.tensor(array).unsqueeze(0), "sample_rate": sr})
-
-    # RVQ encode/decode
-    x       = torch.tensor(array).unsqueeze(0).to("cuda")
-    codes   = rvq.encode(x)
-    decoded = rvq.decode(codes).squeeze().cpu().numpy()
+    vad_pipe({"waveform": torch.tensor(array).unsqueeze(0), "sample_rate": sr})
+    x = torch.tensor(array).unsqueeze(0).to("cuda")
+    codes = rvq.encode(x); decoded = rvq.decode(codes).squeeze().cpu().numpy()
 
-    # Ultravox: speech → text
-    out  = ultravox_pipe({"array": decoded, "sampling_rate": sr})
-    text = out.get("text", "")
+    ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
+    text = ultra_out.get("text", "")
 
-    # Diffusion-based prosody
     pros = diff_pipe(raw_audio=decoded)["audios"][0]
 
-    # Dia TTS synthesis
-    tts   = dia.generate(f"[emotion:neutral] {text}")
-    tts_np = tts.squeeze().cpu().numpy()
+    inputs = tokenizer(f"[emotion:neutral] {text}", return_tensors="pt").to("cuda")
+    tts_tensors = dia.generate(**inputs)
+    tts_np = tts_tensors.squeeze().cpu().numpy()
     tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95 if tts_np.size else tts_np
 
     return (sr, tts_np), text
 
 with gr.Blocks(title="Maya AI 📈") as demo:
     gr.Markdown("## Maya-AI: Supernatural Conversational Agent")
-    audio_in  = gr.Audio(source="microphone", type="numpy", label="Your Voice")
-    send_btn  = gr.Button("Send")
+    audio_in = gr.Audio(source="microphone", type="numpy", label="Your Voice")
+    send_btn = gr.Button("Send")
     audio_out = gr.Audio(label="AI Response")
     text_out  = gr.Textbox(label="Generated Text")
     send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])