app.py

# app.py ──────────────────────────────────────────────────────────────
import os
import json
import torch
import asyncio
import traceback # Import traceback for better error logging

from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM, LogitsProcessor, StoppingCriteria, StoppingCriteriaList
# Import BaseStreamer for the interface
from transformers.generation.streamers import BaseStreamer
from snac import SNAC # Ensure you have 'pip install snac'

# --- Globals (populated in load_models) ---
tok = None
model = None
snac = None
masker = None
stopping_criteria = None
device = "cuda" if torch.cuda.is_available() else "cpu"

# 0) Login + Device ---------------------------------------------------
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
    print("🔑 Logging in to Hugging Face Hub...")
    login(HF_TOKEN)

# torch.backends.cuda.enable_flash_sdp(False) # Uncomment if needed for PyTorch‑2.2‑Bug

# 1) Konstanten -------------------------------------------------------
REPO = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
# CHUNK_TOKENS = 50 # Not directly used by us with the streamer approach
START_TOKEN = 128259
NEW_BLOCK = 128257
EOS_TOKEN = 128258
AUDIO_BASE = 128266
AUDIO_SPAN = 4096 * 7  # 28672 Codes
# Create AUDIO_IDS on the correct device later in load_models
AUDIO_IDS_CPU = torch.arange(AUDIO_BASE, AUDIO_BASE + AUDIO_SPAN)

# 2) Logit‑Mask -------------------------------------------------------
class AudioMask(LogitsProcessor):
    def __init__(self, audio_ids: torch.Tensor, new_block_token_id: int, eos_token_id: int):
        super().__init__()
        # Allow NEW_BLOCK and all valid audio tokens initially
        self.allow = torch.cat([
            torch.tensor([new_block_token_id], device=audio_ids.device), # Add NEW_BLOCK token ID
            audio_ids
        ], dim=0)
        self.eos = torch.tensor([eos_token_id], device=audio_ids.device) # Store EOS token ID as tensor
        self.allow_with_eos = torch.cat([self.allow, self.eos], dim=0) # Precompute combined tensor
        self.sent_blocks = 0 # State: Number of audio blocks sent

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
        # Determine which tokens are allowed based on whether blocks have been sent
        current_allow = self.allow_with_eos if self.sent_blocks > 0 else self.allow

        # Create a mask initialized to negative infinity
        mask = torch.full_like(scores, float("-inf"))
        # Set allowed token scores to 0 (effectively allowing them)
        mask[:, current_allow] = 0
        # Apply the mask to the scores
        return scores + mask

    def reset(self):
        """Resets the state for a new generation request."""
        self.sent_blocks = 0

# 3) StoppingCriteria für EOS ---------------------------------------
# generate() needs explicit stopping criteria when using a streamer
class EosStoppingCriteria(StoppingCriteria):
    def __init__(self, eos_token_id: int):
        self.eos_token_id = eos_token_id

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        # Check if the *last* generated token is the EOS token
        if input_ids.shape[1] > 0 and input_ids[:, -1] == self.eos_token_id:
            # print("StoppingCriteria: EOS detected.")
            return True
        return False

# 4) Benutzerdefinierter AudioStreamer -------------------------------
class AudioStreamer(BaseStreamer):
    def __init__(self, ws: WebSocket, snac_decoder: SNAC, audio_mask: AudioMask, loop: asyncio.AbstractEventLoop):
        self.ws = ws
        self.snac = snac_decoder
        self.masker = audio_mask # Reference to the mask to update sent_blocks
        self.loop = loop # Event loop of the main thread for run_coroutine_threadsafe
        self.device = snac_decoder.device # Get device from the decoder
        self.buf: list[int] = [] # Buffer for audio token values (AUDIO_BASE subtracted)
        self.tasks = set() # Keep track of pending send tasks

    def _decode_block(self, block7: list[int]) -> bytes:
        """
        Decodes a block of 7 audio token values (AUDIO_BASE subtracted) into audio bytes.
        NOTE: The mapping from the 7 tokens to the 3 SNAC codebooks (l1, l2, l3)
              is based on a common interleaving hypothesis. Verify if model docs specify otherwise.
        """
        if len(block7) != 7:
            print(f"Streamer Warning: _decode_block received {len(block7)} tokens, expected 7. Skipping.")
            return b"" # Return empty bytes if block is incomplete

        # --- Hypothesis: Interleaving mapping ---
        # Assumes 7 tokens map to 3 codebooks like this:
        # Codebook 1 (l1) uses tokens at indices 0, 3, 6
        # Codebook 2 (l2) uses tokens at indices 1, 4
        # Codebook 3 (l3) uses tokens at indices 2, 5
        try:
            l1 = [block7[0], block7[3], block7[6]]
            l2 = [block7[1], block7[4]]
            l3 = [block7[2], block7[5]]
        except IndexError:
            print(f"Streamer Error: Index out of bounds during token mapping. Block: {block7}")
            return b""

        # Convert lists to tensors on the correct device
        codes_l1 = torch.tensor(l1, dtype=torch.long, device=self.device).unsqueeze(0)
        codes_l2 = torch.tensor(l2, dtype=torch.long, device=self.device).unsqueeze(0)
        codes_l3 = torch.tensor(l3, dtype=torch.long, device=self.device).unsqueeze(0)
        codes = [codes_l1, codes_l2, codes_l3] # List of tensors for SNAC

        # Decode using SNAC
        with torch.no_grad():
            audio = self.snac.decode(codes)[0] # Decode expects list of tensors, result might have batch dim

        # Squeeze, move to CPU, convert to numpy
        audio_np = audio.squeeze().detach().cpu().numpy()

        # Convert to 16-bit PCM bytes
        audio_bytes = (audio_np * 32767).astype("int16").tobytes()
        return audio_bytes

    async def _send_audio_bytes(self, data: bytes):
        """Coroutine to send bytes over WebSocket."""
        if not data: # Don't send empty bytes
            return
        try:
            await self.ws.send_bytes(data)
            # print(f"Streamer: Sent {len(data)} audio bytes.")
        except WebSocketDisconnect:
            print("Streamer: WebSocket disconnected during send.")
        except Exception as e:
            print(f"Streamer: Error sending bytes: {e}")

    def put(self, value: torch.LongTensor):
        """
        Receives new token IDs (Tensor) from generate() (runs in worker thread).
        Processes tokens, decodes full blocks, and schedules sending via run_coroutine_threadsafe.
        """
        # Ensure value is on CPU and flatten to a list of ints
        if value.numel() == 0:
            return
        new_token_ids = value.squeeze().tolist()
        if isinstance(new_token_ids, int): # Handle single token case
            new_token_ids = [new_token_ids]

        for t in new_token_ids:
            if t == EOS_TOKEN:
                # print("Streamer: EOS token encountered.")
                # EOS is handled by StoppingCriteria, no action needed here except maybe logging.
                break # Stop processing this batch if EOS is found

            if t == NEW_BLOCK:
                # print("Streamer: NEW_BLOCK token encountered.")
                # NEW_BLOCK indicates the start of audio, might reset buffer if needed
                self.buf.clear()
                continue # Move to the next token

            # Check if token is within the expected audio range
            if AUDIO_BASE <= t < AUDIO_BASE + AUDIO_SPAN:
                self.buf.append(t - AUDIO_BASE) # Store value relative to base
            else:
                # Log unexpected tokens (like START_TOKEN or others if generation goes wrong)
                # print(f"Streamer Warning: Ignoring unexpected token {t}")
                pass # Ignore tokens outside the audio range

            # If buffer has 7 tokens, decode and send
            if len(self.buf) == 7:
                audio_bytes = self._decode_block(self.buf)
                self.buf.clear() # Clear buffer after processing

                if audio_bytes: # Only send if decoding was successful
                    # Schedule the async send function to run on the main event loop
                    future = asyncio.run_coroutine_threadsafe(self._send_audio_bytes(audio_bytes), self.loop)
                    self.tasks.add(future)
                    # Optional: Remove completed tasks to prevent memory leak if generation is very long
                    future.add_done_callback(self.tasks.discard)


                    # Allow EOS only after the first full block has been processed and scheduled for sending
                    if self.masker.sent_blocks == 0:
                        # print("Streamer: First audio block processed, allowing EOS.")
                        self.masker.sent_blocks = 1 # Update state in the mask

        # Note: No need to explicitly wait for tasks here. put() should return quickly.

    def end(self):
        """Called by generate() when generation finishes."""
        # Handle any remaining tokens in the buffer (optional, here we discard them)
        if len(self.buf) > 0:
            print(f"Streamer: End of generation with incomplete block ({len(self.buf)} tokens). Discarding.")
            self.buf.clear()

        # Optional: Wait briefly for any outstanding send tasks to complete?
        # This is tricky because end() is sync. A robust solution might involve
        # signaling the WebSocket handler to wait before closing.
        # For simplicity, we rely on FastAPI/Uvicorn's graceful shutdown handling.
        # print(f"Streamer: Generation finished. Pending send tasks: {len(self.tasks)}")
        pass

# 5) FastAPI App ------------------------------------------------------
app = FastAPI()

@app.on_event("startup")
async def load_models_startup(): # Make startup async if needed for future async loads
    global tok, model, snac, masker, stopping_criteria, device, AUDIO_IDS_CPU

    print(f"🚀 Starting up on device: {device}")
    print("⏳ Lade Modelle …", flush=True)

    tok = AutoTokenizer.from_pretrained(REPO)
    print("Tokenizer loaded.")

    # Load SNAC first (usually smaller)
    snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device)
    print(f"SNAC loaded to {snac.device}.")

    # Load the main model
    model = AutoModelForCausalLM.from_pretrained(
        REPO,
        device_map={"": 0} if device == "cuda" else None, # Assign to GPU 0 if cuda
        torch_dtype=torch.bfloat16 if device == "cuda" and torch.cuda.is_bf16_supported() else torch.float32, # Use bfloat16 if supported
        low_cpu_mem_usage=True, # Good practice for large models
    )
    model.config.pad_token_id = model.config.eos_token_id # Set pad token
    print(f"Model loaded to {model.device}.")

    # Ensure model is in evaluation mode
    model.eval()

    # Initialize AudioMask (needs AUDIO_IDS on the correct device)
    audio_ids_device = AUDIO_IDS_CPU.to(device)
    masker = AudioMask(audio_ids_device, NEW_BLOCK, EOS_TOKEN)
    print("AudioMask initialized.")

    # Initialize StoppingCriteria
    # IMPORTANT: Create the list and add the criteria instance
    stopping_criteria = StoppingCriteriaList([EosStoppingCriteria(EOS_TOKEN)])
    print("StoppingCriteria initialized.")

    print("✅ Modelle geladen und bereit!", flush=True)

@app.get("/")
def hello():
    return {"status": "ok", "message": "TTS Service is running"}

# 6) Helper zum Prompt Bauen -------------------------------------------
def build_prompt(text: str, voice: str) -> tuple[torch.Tensor, torch.Tensor]:
    """Builds the input_ids and attention_mask for the model."""
    # Format: <START> <VOICE>: <TEXT> <NEW_BLOCK>
    prompt_text = f"{voice}: {text}"
    prompt_ids = tok(prompt_text, return_tensors="pt").input_ids.to(device)

    # Construct input_ids tensor
    input_ids = torch.cat([
        torch.tensor([[START_TOKEN]], device=device), # Start token
        prompt_ids,                                  # Encoded prompt
        torch.tensor([[NEW_BLOCK]], device=device)   # New block token to trigger audio
    ], dim=1)

    # Create attention mask (all ones)
    attention_mask = torch.ones_like(input_ids)
    return input_ids, attention_mask

# 7) WebSocket‑Endpoint (vereinfacht mit Streamer) ---------------------
@app.websocket("/ws/tts")
async def tts(ws: WebSocket):
    await ws.accept()
    print(" клиент подключился") # Client connected
    streamer = None # Initialize for finally block
    main_loop = asyncio.get_running_loop() # Get the current event loop

    try:
        # Receive configuration
        req_text = await ws.receive_text()
        print(f"Received request: {req_text}")
        req = json.loads(req_text)
        text = req.get("text", "Hallo Welt, wie geht es dir heute?") # Default text
        voice = req.get("voice", "Jakob") # Default voice

        if not text:
            await ws.close(code=1003, reason="Text cannot be empty")
            return

        print(f"Generating audio for: '{text}' with voice '{voice}'")

        # Prepare prompt
        ids, attn = build_prompt(text, voice)

        # --- Reset stateful components ---
        masker.reset() # CRITICAL: Reset the mask state for the new request

        # --- Create Streamer Instance ---
        streamer = AudioStreamer(ws, snac, masker, main_loop)

        # --- Run model.generate in a separate thread ---
        # This prevents blocking the main FastAPI event loop
        print("Starting generation...")
        await asyncio.to_thread(
            model.generate,
            input_ids=ids,
            attention_mask=attn,
            max_new_tokens=1500, # Limit generation length (adjust as needed)
            logits_processor=[masker],
            stopping_criteria=stopping_criteria,
            do_sample=False, # Use greedy decoding for potentially more stable audio
            # do_sample=True, temperature=0.7, top_p=0.95, # Or use sampling
            use_cache=True,
            streamer=streamer # Pass the custom streamer
            # No need to manage past_key_values manually
        )
        print("Generation finished.")

    except WebSocketDisconnect:
        print("Client disconnected.")
    except json.JSONDecodeError:
        print("❌ Invalid JSON received.")
        await ws.close(code=1003, reason="Invalid JSON format") # 1003 = Cannot accept data type
    except Exception as e:
        error_details = traceback.format_exc()
        print(f"❌ WS‑Error: {e}\n{error_details}", flush=True)
        # Try to send an error message before closing, if possible
        error_payload = json.dumps({"error": str(e)})
        try:
            if ws.client_state.name == "CONNECTED":
                 await ws.send_text(error_payload) # Send error as text/json
        except Exception:
            pass # Ignore error during error reporting
        # Close with internal server error code
        if ws.client_state.name == "CONNECTED":
            await ws.close(code=1011) # 1011 = Internal Server Error
    finally:
        # Ensure streamer's end method is called if it exists
        if streamer:
            try:
                streamer.end()
            except Exception as e_end:
                 print(f"Error during streamer.end(): {e_end}")

        # Ensure WebSocket is closed
        print("Closing connection.")
        if ws.client_state.name != "DISCONNECTED":
            try:
                await ws.close(code=1000) # 1000 = Normal Closure
            except RuntimeError as e_close:
                 # Can happen if connection is already closing/closed
                 print(f"Runtime error closing websocket: {e_close}")
            except Exception as e_close_final:
                 print(f"Error closing websocket: {e_close_final}")
        print("Connection closed.")

# 8) Dev‑Start --------------------------------------------------------
if __name__ == "__main__":
    import uvicorn
    print("Starting Uvicorn server...")
    # Use reload=True only for development, remove for production
    uvicorn.run("app:app", host="0.0.0.0", port=7860, log_level="info") #, reload=True)