app.py

# app.py ──────────────────────────────────────────────────────────────
import os, json, torch, asyncio
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM, LogitsProcessor
from snac import SNAC

# 0) Login + Device ---------------------------------------------------
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
    login(HF_TOKEN)

device = "cuda" if torch.cuda.is_available() else "cpu"
torch.backends.cuda.enable_flash_sdp(False)          # PyTorch‑2.2‑Bug

# 1) Konstanten -------------------------------------------------------
REPO           = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
CHUNK_TOKENS   = 50
START_TOKEN    = 128259
NEW_BLOCK      = 128257
EOS_TOKEN      = 128258
AUDIO_BASE     = 128266
AUDIO_IDS      = torch.arange(AUDIO_BASE, AUDIO_BASE + 4096)

# 2) Logit‑Mask (NEW_BLOCK + Audio; EOS erst nach 1. Block) ----------
class AudioMask(LogitsProcessor):
    def __init__(self, audio_ids: torch.Tensor):
        super().__init__()
        self.allow = torch.cat([
            torch.tensor([NEW_BLOCK], device=audio_ids.device),
            audio_ids
        ])
        self.eos   = torch.tensor([EOS_TOKEN], device=audio_ids.device)
        self.sent_blocks = 0

    def __call__(self, input_ids, logits):
        allowed = self.allow
        if self.sent_blocks:                        # ab 1. Block EOS zulassen
            allowed = torch.cat([allowed, self.eos])
        mask = logits.new_full(logits.shape, float("-inf"))
        mask[:, allowed] = 0
        return logits + mask

# 3) FastAPI Grundgerüst ---------------------------------------------
app = FastAPI()

@app.get("/")
def hello():
    return {"status": "ok"}

@app.on_event("startup")
def load_models():
    global tok, model, snac, masker
    print("⏳ Lade Modelle …", flush=True)

    tok   = AutoTokenizer.from_pretrained(REPO)
    snac  = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device)
    model = AutoModelForCausalLM.from_pretrained(
        REPO,
        device_map={"": 0} if device == "cuda" else None,
        torch_dtype=torch.bfloat16 if device == "cuda" else None,
        low_cpu_mem_usage=True,
    )
    model.config.pad_token_id = model.config.eos_token_id
    masker = AudioMask(AUDIO_IDS.to(device))

    print("✅ Modelle geladen", flush=True)

# 4) Helper -----------------------------------------------------------
def build_prompt(text: str, voice: str):
    prompt_ids = tok(f"{voice}: {text}", return_tensors="pt").input_ids.to(device)
    ids   = torch.cat([torch.tensor([[START_TOKEN]], device=device),
                       prompt_ids,
                       torch.tensor([[128009, 128260]], device=device)], 1)
    attn  = torch.ones_like(ids)
    return ids, attn

def decode_block(block7: list[int]) -> bytes:
    l1,l2,l3=[],[],[]
    l1.append(block7[0])
    l2.append(block7[1]-4096)
    l3 += [block7[2]-8192, block7[3]-12288]
    l2.append(block7[4]-16384)
    l3 += [block7[5]-20480, block7[6]-24576]

    with torch.no_grad():
        codes = [torch.tensor(x, device=device).unsqueeze(0)
                 for x in (l1,l2,l3)]
        audio = snac.decode(codes).squeeze().detach().cpu().numpy()

    return (audio*32767).astype("int16").tobytes()

# 5) WebSocket‑Endpoint ----------------------------------------------
@app.websocket("/ws/tts")
async def tts(ws: WebSocket):
    await ws.accept()
    try:
        req   = json.loads(await ws.receive_text())
        text  = req.get("text", "")
        voice = req.get("voice", "Jakob")

        ids, attn  = build_prompt(text, voice)
        past       = None
        offset_len = ids.size(1)          # wie viele Tokens existieren schon
        last_tok   = None
        buf        = []

        while True:
            # --- Mini‑Generate -------------------------------------------
            gen = model.generate(
                input_ids       = ids if past is None else torch.tensor([[last_tok]], device=device),
                attention_mask  = attn if past is None else None,
                past_key_values = past,
                max_new_tokens  = CHUNK_TOKENS,
                logits_processor= [masker],
                do_sample=True, temperature=0.7, top_p=0.95,
                use_cache=True
            )

            # ----- neue Tokens heraus schneiden --------------------------
            new = gen[0, offset_len:].tolist()
            if not new:                         # nichts -> fertig
                break
            offset_len += len(new)

            # ----- weiter mit Cache (letzte PKV steht im Modell) ---------
            past     = model._past_key_values
            last_tok = new[-1]

            print("new tokens:", new[:25], flush=True)

            # ----- Token‑Handling ----------------------------------------
            for t in new:
                if t == EOS_TOKEN:
                    raise StopIteration
                if t == NEW_BLOCK:
                    buf.clear()
                    continue
                buf.append(t - AUDIO_BASE)
                if len(buf) == 7:
                    await ws.send_bytes(decode_block(buf))
                    buf.clear()
                    masker.sent_blocks = 1      # ab jetzt EOS zulässig

    except (StopIteration, WebSocketDisconnect):
        pass
    except Exception as e:
        print("❌ WS‑Error:", e, flush=True)
        if ws.client_state.name != "DISCONNECTED":
            await ws.close(code=1011)
    finally:
        if ws.client_state.name != "DISCONNECTED":
            try:
                await ws.close()
            except RuntimeError:
                pass

# 6) Dev‑Start --------------------------------------------------------
if __name__ == "__main__":
    import uvicorn, sys
    uvicorn.run("app:app", host="0.0.0.0", port=7860, log_level="info")