up3

app.py

@@ -1,6 +1,4 @@
-import os
-import json
-import asyncio
+import os, json, asyncio
 import torch
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from dotenv import load_dotenv
@@ -8,138 +6,87 @@ 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)
+if (tok := os.getenv("HF_TOKEN")):
+    login(token=tok)
 
-# — Device —
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# — SNAC laden —
-print("Loading SNAC model...")
+print("Loading SNAC…")
 snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device)
 
-# — Orpheus‑Modell vorbereiten —
 model_name = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
-
-# Nur Konfig+Weights (ermöglicht schlankeren Container)
 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.*"
-    ]
+    ignore_patterns=[ "optimizer.pt", "pytorch_model.bin", "training_args.bin",
+                      "scheduler.pt", "tokenizer.*", "vocab.json", "merges.txt" ]
 )
 
-print("Loading Orpheus model...")
+print("Loading Orpheus…")
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    torch_dtype=torch.bfloat16,
-    device_map="auto",
-).to(device)
+    torch_dtype=torch.bfloat16
+)
+model = model.to(device)
 model.config.pad_token_id = model.config.eos_token_id
 
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-
-# — Hilfsfunktionen —
+# — Helper Functions (wie gehabt) —
 
 def process_prompt(text: str, voice: str):
     prompt = f"{voice}: {text}"
     inputs = tokenizer(prompt, return_tensors="pt").to(device)
-    # füge Start-/End-Tokens hinzu
     start = torch.tensor([[128259]], device=device)
     end   = torch.tensor([[128009, 128260]], device=device)
-    input_ids = torch.cat([start, inputs.input_ids, end], dim=1)
-    return input_ids
-
-def parse_output(generated_ids: torch.LongTensor):
-    token_to_find   = 128257
-    token_to_remove = 128258
+    return torch.cat([start, inputs.input_ids, end], dim=1)
 
-    idxs = (generated_ids == token_to_find).nonzero(as_tuple=True)[1]
-    if idxs.numel() > 0:
-        cropped = generated_ids[:, idxs[-1].item() + 1 :]
-    else:
-        cropped = generated_ids
-
-    row = cropped[0][cropped[0] != token_to_remove]
+def parse_output(ids: torch.LongTensor):
+    st, rm = 128257, 128258
+    idxs = (ids==st).nonzero(as_tuple=True)[1]
+    cropped = ids[:, idxs[-1].item()+1:] if idxs.numel()>0 else ids
+    row = cropped[0][cropped[0]!=rm]
     return row.tolist()
 
-def redistribute_codes(code_list: list[int], snac_model: SNAC):
-    layer1, layer2, layer3 = [], [], []
-    for i in range((len(code_list) + 1) // 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
-    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()
-
+def redistribute_codes(codes: list[int], snac_model: SNAC):
+    # … genau wie vorher …
+    # return numpy array
 
-# — FastAPI App —  
 app = FastAPI()
 
 @app.get("/")
-async def hello():
-    return {"message": "Hello, Orpheus TTS is up and running!"}
+async def root():
+    return {"status":"ok","msg":"Hello, Orpheus TTS up!"}
 
 @app.websocket("/ws/tts")
-async def tts_ws(ws: WebSocket):
+async def ws_tts(ws: WebSocket):
     await ws.accept()
     try:
-        # **Nur EIN Request pro Connection**
-        raw = await ws.receive_text()
-        data = json.loads(raw)
-        text  = data.get("text", "")
-        voice = data.get("voice", "Jakob")
-
-        # 1) Text → input_ids
-        input_ids = process_prompt(text, voice)
-
-        # 2) Generation
-        gen_ids = model.generate(
-            input_ids=input_ids,
-            max_new_tokens=2000,    # hier kannst du hochsetzen
-            do_sample=True,
-            temperature=0.7,
-            top_p=0.95,
+        msg = json.loads(await ws.receive_text())
+        text, voice = msg.get("text",""), msg.get("voice","Jakob")
+        ids = process_prompt(text, voice)
+        gen = model.generate(
+            input_ids=ids,
+            max_new_tokens=2000,
+            do_sample=True, temperature=0.7, top_p=0.95,
             repetition_penalty=1.1,
             eos_token_id=model.config.eos_token_id,
         )
-
-        # 3) Token → Audio
-        codes    = parse_output(gen_ids)
+        codes    = parse_output(gen)
         audio_np = redistribute_codes(codes, snac)
-
-        # 4) PCM16-Bytes in ~0.1s‑Chunks streamen
-        pcm16 = (audio_np * 32767).astype("int16").tobytes()
-        chunk_size = 2400 * 2  # 2400 Samples @24kHz = 0.1s * 2 Byte
-        for i in range(0, len(pcm16), chunk_size):
-            await ws.send_bytes(pcm16[i : i+chunk_size])
+        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)
-
-        # Sauber schließen, Client erhält ConnectionClosedOK
         await ws.close()
-
     except WebSocketDisconnect:
-        print("Client disconnected")
+        print("Client left")
     except Exception as e:
-        # Log und saubere Fehler‑Closure
-        print("Error in /ws/tts:", 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)