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)

# — Gerät wählen —
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"
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 Token‑Mapping —
AUDIO_TOKEN_OFFSET = 128266
START_TOKEN  = 128259
SOS_TOKEN    = 128257
EOS_TOKEN    = 128258

# — Hilfsfunktionen —
def process_prompt(text: str, voice: str):
    prompt = f"{voice}: {text}"
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
    start = torch.tensor([[START_TOKEN]], dtype=torch.int64, device=device)
    end   = torch.tensor([[128009, 128260]], dtype=torch.int64, device=device)
    ids  = torch.cat([start, input_ids, end], dim=1)
    mask = torch.ones_like(ids, dtype=torch.int64, device=device)
    return ids, mask

def redistribute_codes(block: list[int], snac_model: SNAC):
    # exakt wie vorher: 7 Codes → 3 Layer → SNAC.decode → NumPy float32 @24 kHz
    l1, l2, l3 = [], [], []
    for i in range(len(block)//7):
        b = block[7*i:7*i+7]
        l1.append(b[0])
        l2.append(b[1] -   4096)
        l3.append(b[2] - 2*4096)
        l3.append(b[3] - 3*4096)
        l2.append(b[4] - 4*4096)
        l3.append(b[5] - 5*4096)
        l3.append(b[6] - 6*4096)
    dev = next(snac_model.parameters()).device
    codes = [
        torch.tensor(l1, device=dev).unsqueeze(0),
        torch.tensor(l2, device=dev).unsqueeze(0),
        torch.tensor(l3, device=dev).unsqueeze(0),
    ]
    audio = snac_model.decode(codes)  # → Tensor[1, T]
    return audio.squeeze().cpu().numpy()

# — FastAPI Setup —
app = FastAPI()

# 1) Hello‑World Endpoint
@app.get("/")
async def root():
    return {"message": "Hallo Welt"}

# 2) WebSocket Token‑für‑Token TTS
@app.websocket("/ws/tts")
async def tts_ws(ws: WebSocket):
    await ws.accept()
    try:
        while True:
            # JSON mit Text & Voice empfangen
            raw = await ws.receive_text()
            req = json.loads(raw)
            text, voice = req.get("text", ""), req.get("voice", "Jakob")
            ids, mask = process_prompt(text, voice)

            past_kv = None
            collected = []

            # im Sampling‑Loop Token für Token generieren
            with torch.no_grad():
                for _ in range(2000):  # max 200 Tokens
                    out = model(
                        input_ids=ids if past_kv is None else None,
                        attention_mask=mask if past_kv is None else None,
                        past_key_values=past_kv,
                        use_cache=True,
                    )
                    logits = out.logits[:, -1, :]
                    next_id = torch.multinomial(torch.softmax(logits, dim=-1), num_samples=1)
                    past_kv = out.past_key_values

                    token = next_id.item()
                    # Ende
                    if token == EOS_TOKEN:
                        break
                    # Reset bei SOS
                    if token == SOS_TOKEN:
                        collected = []
                        continue

                    # in Audio‑Code konvertieren
                    collected.append(token - AUDIO_TOKEN_OFFSET)

                    # sobald 7 Codes → direkt dekodieren & streamen
                    if len(collected) >= 7:
                        block = collected[:7]
                        collected = collected[7:]
                        audio_np = redistribute_codes(block, snac)
                        pcm16 = (audio_np * 32767).astype("int16").tobytes()
                        await ws.send_bytes(pcm16)

                    # ab jetzt nur noch past_kv verwenden
                    ids  = None
                    mask = None

            # zum Schluss End‑Of‑Stream signalisieren
            await ws.send_text(json.dumps({"event": "eos"}))

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

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