Upload app.py with huggingface_hub

app.py

@@ -6,15 +6,15 @@ import torch
 import tempfile
 import traceback
 import gc
-from fastapi import FastAPI, Request, Response
+from fastapi import FastAPI, Request
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import HTMLResponse
 import uvicorn
 from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
 from TTS.api import TTS
 
-# --- [v138] 🚀 ZEROGPU LEGACY-FREE ENGINE ---
-print(f"--- [v138] 📡 BOOTING API ENGINE ---")
+# --- [v139] 🚀 H200 SAFEST MODE (FP32 + Standard Attention) ---
+print(f"--- [v139] 📡 BOOTING SAFEST ENGINE ---")
 
 try:
     import spaces
@@ -27,12 +27,14 @@ except ImportError:
             if f is None: return lambda x: x
             return f
 
-# --- System Config ---
+# --- Strict Stability Config ---
 os.environ["COQUI_TOS_AGREED"] = "1"
 os.environ["PYTHONWARNINGS"] = "ignore"
+# Disable all hardware acceleration that might cause kernel alignment errors
 os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
 torch.backends.cuda.matmul.allow_tf32 = False
 torch.backends.cudnn.allow_tf32 = False
+torch.backends.cudnn.deterministic = True # Extra safety
 
 app = FastAPI()
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
@@ -40,15 +42,16 @@ app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], all
 MODELS = {"stt": None, "tts": None}
 
 def load_gpu_models():
-    """Persistent loading into GPU VRAM. Only runs once per worker."""
+    """Persistent loading into GPU VRAM (FP32 focus)."""
     global MODELS
     device = "cuda"
     
     if MODELS.get("stt") is None:
-        print("--- [v138] 📥 LOADING NATIVE WHISPER (Large-v3-Turbo) ---")
+        print("--- [v139] 📥 LOADING NATIVE WHISPER (FP32 / No-SDPA) ---")
         model_id = "openai/whisper-large-v3-turbo"
+        # Force float32 to avoid CUBLAS_STATUS_INVALID_VALUE on H200 MIG
         model = AutoModelForSpeechSeq2Seq.from_pretrained(
-            model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True, use_safetensors=True
+            model_id, torch_dtype=torch.float32, low_cpu_mem_usage=True, use_safetensors=True
         ).to(device)
         processor = AutoProcessor.from_pretrained(model_id)
         
@@ -57,22 +60,24 @@ def load_gpu_models():
             model=model,
             tokenizer=processor.tokenizer,
             feature_extractor=processor.feature_extractor,
-            torch_dtype=torch.float16,
+            torch_dtype=torch.float32,
             device=device,
-            model_kwargs={"attn_implementation": "sdpa"}
+            # Explicitly avoid SDPA/Flash Attention to dodge kernel bugs
+            model_kwargs={"attn_implementation": "eager"} 
         )
-        print("--- [v138] ✅ WHISPER LOADED ---")
+        print("--- [v139] ✅ WHISPER LOADED (FP32) ---")
 
     if MODELS.get("tts") is None:
-        print("--- [v138] 📥 LOADING XTTS (VRAM STABLE) ---")
+        print("--- [v139] 📥 LOADING XTTS (SINGLETON) ---")
+        # XTTS is generally stable if in VRAM
         MODELS["tts"] = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
-        print("--- [v138] ✅ XTTS LOADED ---")
+        print("--- [v139] ✅ XTTS LOADED ---")
 
 @spaces.GPU(duration=120)
 def core_process(request_dict):
     global MODELS
     action = request_dict.get("action")
-    print(f"--- [v138] 🛠️ HOPPER ENGINE: {action} ---")
+    print(f"--- [v139] 🛠️ SAFE ENGINE: {action} ---")
     t1 = time.time()
     
     try:
@@ -86,7 +91,8 @@ def core_process(request_dict):
             
             try:
                 lang = request_dict.get("lang")
-                result = MODELS["stt"](temp_path, generate_kwargs={"language": lang if lang and len(lang) <= 3 else None})
+                # batch_size=1 for maximum stability
+                result = MODELS["stt"](temp_path, batch_size=1, generate_kwargs={"language": lang if lang and len(lang) <= 3 else None})
                 stt_text = result["text"].strip()
             finally:
                 if os.path.exists(temp_path): os.unlink(temp_path)
@@ -129,60 +135,30 @@ def core_process(request_dict):
             return {"text": stt_text, "translated": trans_text, "audio": audio_b64}
 
     except Exception as e:
-        print(f"❌ [v138] ERROR: {traceback.format_exc()}")
+        print(f"❌ [v139] ERROR: {traceback.format_exc()}")
         return {"error": str(e)}
     finally:
-        print(f"--- [v138] ✨ DONE ({time.time()-t1:.1f}s) ---")
+        print(f"--- [v139] ✨ DONE ({time.time()-t1:.1f}s) ---")
         torch.cuda.empty_cache() 
 
-# --- API Endpoints ---
 @app.post("/process")
 async def api_process(request: Request):
     try:
         data = await request.json()
-        if data.get("action") == "health": return {"status": "awake", "v": "138"}
+        if data.get("action") == "health": return {"status": "awake", "v": "139"}
         return core_process(data)
     except Exception as e: return {"error": str(e)}
 
 @app.get("/health")
-def health(): return {"status": "ok", "v": "138", "gpu": HAS_SPACES}
+def health(): return {"status": "ok", "v": "139", "gpu": HAS_SPACES}
 
-# --- Minimal UI ---
 @app.get("/", response_class=HTMLResponse)
 def root():
     return """
-    <html>
-        <head>
-            <title>S2ST v138</title>
-            <style>
-                body { font-family: sans-serif; background: #111; color: #eee; text-align: center; padding-top: 50px; }
-                .card { background: #222; border: 1px solid #444; padding: 20px; border-radius: 10px; display: inline-block; }
-                button { background: #007bff; color: #fff; border: none; padding: 10px 20px; border-radius: 5px; cursor: pointer; }
-                #log { margin-top: 20px; color: #aaa; font-family: monospace; }
-            </style>
-        </head>
-        <body>
-            <div class="card">
-                <h1>🚀 AI Engine v138</h1>
-                <p>HOPPER NATIVE - FASTAPI ONLY</p>
-                <button onclick="checkHealth()">Test API Connectivity</button>
-                <div id="log">Status: Awaiting test...</div>
-            </div>
-            <script>
-                async function checkHealth() {
-                    const log = document.getElementById('log');
-                    log.innerText = 'Checking...';
-                    try {
-                        const res = await fetch('/health');
-                        const data = await res.json();
-                        log.innerText = 'Response: ' + JSON.stringify(data);
-                    } catch (e) {
-                        log.innerText = 'Error: ' + e;
-                    }
-                }
-            </script>
-        </body>
-    </html>
+    <html><head><title>S2ST v139</title><style>body { font-family: sans-serif; background: #111; color: #eee; text-align: center; padding-top: 50px; }</style></head>
+    <body><h1>🚀 AI Engine v139 (FP32 SAFE)</h1><p>H200 Native Stability Test</p><div id="log">Awaiting test...</div>
+    <script>fetch('/health').then(r=>r.json()).then(d=>document.getElementById('log').innerText=JSON.stringify(d));</script>
+    </body></html>
     """
 
 if __name__ == "__main__":