app.py

import os
import json
import asyncio
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"
# Nur die Konfig + Safetensors, alles andere wird ignoriert
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)

# — Konstanten für Audio‑Token →
# (muss übereinstimmen mit Deinem Training; hier 128266)
AUDIO_TOKEN_OFFSET = 128266

# — Hilfsfunktionen —

def process_prompt(text: str, voice: str):
    prompt = f"{voice}: {text}"
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    # Laut Spezifikation:
    #  start_token=128259, end_tokens=(128009,128260)
    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):
    """
    Croppt nach dem letzten 128257-Start-Token, entfernt Padding (128258)
    und zieht dann den Audio‑Offset ab, um echte Code‑IDs zu bekommen.
    """
    # finde letztes Audio‑Start‑Token
    token_to_start  = 128257
    token_to_remove = model.config.eos_token_id  # 128258

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

    # flatten & remove PAD, dann Offset abziehen
    flat = cropped[0][cropped[0] != token_to_remove]
    codes = [(int(t) - AUDIO_TOKEN_OFFSET) for t in flat]
    return codes

def redistribute_codes(code_list: list[int], snac_model: SNAC):
    """
    Verteilt die flache Code‑Liste in 3 Layers und dekodiert mit SNAC.
    """
    layer1, layer2, layer3 = [], [], []
    for i in range(len(code_list) // 7):
        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)

    dev = next(snac_model.parameters()).device
    c1 = torch.tensor(layer1, device=dev).unsqueeze(0)
    c2 = torch.tensor(layer2, device=dev).unsqueeze(0)
    c3 = torch.tensor(layer3, device=dev).unsqueeze(0)
    audio = snac_model.decode([c1, c2, c3])
    return audio.detach().squeeze().cpu().numpy()

# — FastAPI + WebSocket-Endpoint —
app = FastAPI()

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

            # 1) Prompt → Token‑Tensoren
            ids, mask = process_prompt(text, voice)

            # 2) Generation
            gen_ids = model.generate(
                input_ids=ids,
                attention_mask=mask,
                max_new_tokens=200,  # zum Debug
                do_sample=True,
                temperature=0.7,
                top_p=0.95,
                repetition_penalty=1.1,
                eos_token_id=model.config.eos_token_id,
            )

            # 3) Token → Code‑Liste → Audio (Float32 @24 kHz)
            code_list = parse_output(gen_ids)
            audio_np  = redistribute_codes(code_list, snac)

            # 4) In 0.1 s‑Chunks (2400 Samples) als PCM16 streamen
            pcm16 = (audio_np * 32767).astype("int16").tobytes()
            chunk = 2400 * 2
            for i in range(0, len(pcm16), chunk):
                await ws.send_bytes(pcm16[i : i+chunk])
                await asyncio.sleep(0.1)

    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)