app.py

import os
import json
import asyncio
import numpy as np
import torch
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from dotenv import load_dotenv
from snac import SNAC
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import login, snapshot_download

# — ENV & HF‑AUTH —
load_dotenv()
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
    login(token=HF_TOKEN)

# — Device —
device = "cuda" if torch.cuda.is_available() else "cpu"

# — Modelle laden —
print("Loading SNAC model...")
snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device)

model_name = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
print("Downloading model weights (config + safetensors)...")
snapshot_download(
    repo_id=model_name,
    allow_patterns=["config.json", "*.safetensors", "model.safetensors.index.json"],
    ignore_patterns=[
        "optimizer.pt", "pytorch_model.bin", "training_args.bin",
        "scheduler.pt", "tokenizer.json", "tokenizer_config.json",
        "special_tokens_map.json", "vocab.json", "merges.txt", "tokenizer.*"
    ]
)

print("Loading Orpheus model...")
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16
).to(device)
model.config.pad_token_id = model.config.eos_token_id

tokenizer = AutoTokenizer.from_pretrained(model_name)

# — Hilfsfunktionen —

def process_prompt(text: str, voice: str):
    """Bereitet input_ids und attention_mask für das Modell vor."""
    prompt = f"{voice}: {text}"
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    start = torch.tensor([[128259]], dtype=torch.int64)
    end   = torch.tensor([[128009, 128260]], dtype=torch.int64)
    ids  = torch.cat([start, input_ids, end], dim=1).to(device)
    mask = torch.ones_like(ids).to(device)
    return ids, mask

def parse_output(generated_ids: torch.LongTensor):
    """Extrahiere rohe Tokenliste nach dem letzten 128257-Start-Token."""
    token_to_find   = 128257
    token_to_remove = 128258

    idxs = (generated_ids == token_to_find).nonzero(as_tuple=True)[1]
    if idxs.numel() > 0:
        cut = idxs[-1].item() + 1
        cropped = generated_ids[:, cut:]
    else:
        cropped = generated_ids

    # Entferne EOS‑Token
    row = cropped[0]
    return row[row != token_to_remove].tolist()

def redistribute_codes(code_list: list[int], snac_model: SNAC):
    """
    Verteilt die Token nur in kompletten 7er‑Blöcken auf die drei SNAC‑Layer
    und dekodiert in Audio. Unvollständige Reste (<7 Tokens) werden verworfen.
    """
    n_blocks = len(code_list) // 7
    layer1, layer2, layer3 = [], [], []

    for i in range(n_blocks):
        base = code_list[7*i : 7*i + 7]
        layer1.append(base[0])
        layer2.append(base[1] -   4096)
        layer3.append(base[2] - 2*4096)
        layer3.append(base[3] - 3*4096)
        layer2.append(base[4] - 4*4096)
        layer3.append(base[5] - 5*4096)
        layer3.append(base[6] - 6*4096)

    if not layer1:
        # kein kompletter Block → leeres Audio
        return np.zeros(0, dtype=np.float32)

    dev = next(snac_model.parameters()).device
    codes = [
        torch.tensor(layer1, device=dev).unsqueeze(0),
        torch.tensor(layer2, device=dev).unsqueeze(0),
        torch.tensor(layer3, device=dev).unsqueeze(0),
    ]
    audio = snac_model.decode(codes)
    return audio.detach().squeeze().cpu().numpy()

# — FastAPI Setup —

app = FastAPI()

@app.get("/")
def greet_json():
    return {"Hello": "World!"}

@app.websocket("/ws/tts")
async def tts_ws(ws: WebSocket):
    await ws.accept()
    try:
        while True:
            # Erwartet JSON: {"text": "...", "voice": "Jakob"}
            data = json.loads(await ws.receive_text())
            text  = data.get("text", "")
            voice = data.get("voice", "Jakob")

            # 1) Tokens vorbereiten
            ids, mask = process_prompt(text, voice)

            # 2) Generierung
            gen_ids = model.generate(
                input_ids=ids,
                attention_mask=mask,
                max_new_tokens=2000,    # hier nach Bedarf anpassen
                do_sample=True,
                temperature=0.7,
                top_p=0.95,
                repetition_penalty=1.1,
                eos_token_id=model.config.eos_token_id,
            )

            # 3) Tokens → Code-Liste → Audio
            codes   = parse_output(gen_ids)
            audio_np = redistribute_codes(codes, snac)

            # 4) in 0.1s‑Stücken PCM16 streamen
            pcm16 = (audio_np * 32767).astype("int16").tobytes()
            chunk = 2400 * 2  # 2400 samples @24kHz = 0.1s * 2 bytes
            for i in range(0, len(pcm16), chunk):
                await ws.send_bytes(pcm16[i : i+chunk])
                await asyncio.sleep(0.1)

        # Ende der while‐Schleife
    except WebSocketDisconnect:
        print("Client disconnected")
    except Exception as e:
        print("Error in /ws/tts:", e)
        await ws.close(code=1011)

if __name__ == "__main__":
    import uvicorn
    uvicorn.run("app:app", host="0.0.0.0", port=7860)