Update app.py

app.py

@@ -2,13 +2,11 @@ import os
 import sys
 import logging
 
-# Suprime warnings de asyncio e gradio
+# 1. SUPRESSÃO DE RUÍDO NOS LOGS
 logging.getLogger("asyncio").setLevel(logging.CRITICAL)
 logging.getLogger("gradio").setLevel(logging.ERROR)
 
-# ==============================================================================
-# 1. CONFIGURAÇÃO DE AMBIENTE CUDA (ZeroGPU)
-# ==============================================================================
+# 2. CONFIGURAÇÃO CUDA ZERO-GPU
 cuda_paths = [
     "/usr/local/cuda/lib64",
     "/usr/local/nvidia/lib64",
@@ -16,17 +14,13 @@ cuda_paths = [
     os.path.join(sys.prefix, "lib", f"python{sys.version_info.major}.{sys.version_info.minor}", "site-packages", "nvidia", "cuda_runtime", "lib"),
     os.path.join(sys.prefix, "lib", f"python{sys.version_info.major}.{sys.version_info.minor}", "site-packages", "nvidia", "cublas", "lib"),
 ]
-
 existing_ld = os.environ.get("LD_LIBRARY_PATH", "")
 new_paths = [p for p in cuda_paths if os.path.exists(p)]
 os.environ["LD_LIBRARY_PATH"] = ":".join(new_paths + [existing_ld])
-
 os.environ["GRADIO_SSR_MODE"] = "False"
 os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
 
-# ==============================================================================
-# 2. IMPORTAÇÕES GLOBAIS
-# ==============================================================================
+# 3. IMPORTAÇÕES
 import threading
 import uvicorn
 import time
@@ -36,7 +30,7 @@ import gc
 import gradio as gr
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import StreamingResponse, JSONResponse
+from fastapi.responses import StreamingResponse
 from pydantic import BaseModel
 from typing import List, Optional
 from pyngrok import ngrok
@@ -46,55 +40,26 @@ import spaces
 NGROK_TOKEN = os.getenv("NGROK_TOKEN")
 
 MODEL_REGISTRY = {
-    "qwen2.5-coder-7b": {
-        "repo_id": "Davizig10jojo/Qwen2.5-Coder-Mix-7B-GGUF",
-        "filename": "Qwen2.5-Coder-Mix-7B-Q2_K.gguf",
-        "min_size_mb": 1000
-    },
-    "blazernano-0.6b": {
-        "repo_id": "Davizig10jojo/BlazerNano-0.6b-GGUF",
-        "filename": "blazernano-0.6b-Q6_K.gguf",
-        "min_size_mb": 100
-    },
-    "blazertiny-1b": {
-        "repo_id": "Davizig10jojo/BlazerTiny-1b-GGUF",
-        "filename": "blazertiny-1b-Q6_K.gguf",
-        "min_size_mb": 200
-    },
-    "blazerstandard-4b": {
-        "repo_id": "Davizig10jojo/BlazerStandard-4B-GGUF",
-        "filename": "blazerstandard-4b-Q6_K.gguf",
-        "min_size_mb": 2000
-    },
-    "blazerrhino-3b": {
-        "repo_id": "Davizig10jojo/BlazerRhino-3B-GGUF",
-        "filename": "BlazerRhino-3B-Instruct.Q4_K_M.gguf",
-        "min_size_mb": 1000
-    }
+    "qwen2.5-coder-7b": {"repo_id": "Davizig10jojo/Qwen2.5-Coder-Mix-7B-GGUF", "filename": "Qwen2.5-Coder-Mix-7B-Q2_K.gguf", "min_size_mb": 1000},
+    "blazernano-0.6b": {"repo_id": "Davizig10jojo/BlazerNano-0.6b-GGUF", "filename": "blazernano-0.6b-Q6_K.gguf", "min_size_mb": 100},
+    "blazertiny-1b": {"repo_id": "Davizig10jojo/BlazerTiny-1b-GGUF", "filename": "blazertiny-1b-Q6_K.gguf", "min_size_mb": 200},
+    "blazerstandard-4b": {"repo_id": "Davizig10jojo/BlazerStandard-4B-GGUF", "filename": "blazerstandard-4b-Q6_K.gguf", "min_size_mb": 2000},
+    "blazerrhino-3b": {"repo_id": "Davizig10jojo/BlazerRhino-3B-GGUF", "filename": "BlazerRhino-3B-Instruct.Q4_K_M.gguf", "min_size_mb": 1000}
 }
 
 print("📥 Mapeando caminhos dos arquivos locais...")
 MODEL_PATHS = {}
-
 for key, meta in MODEL_REGISTRY.items():
     try:
         path = hf_hub_download(repo_id=meta["repo_id"], filename=meta["filename"])
-        size_mb = os.path.getsize(path) / (1024 * 1024)
-        if size_mb < meta["min_size_mb"]:
-            raise Exception(f"Arquivo muito pequeno ({size_mb:.2f}MB).")
+        if os.path.getsize(path) / (1024 * 1024) < meta["min_size_mb"]: raise Exception("Arquivo corrompido")
         MODEL_PATHS[key] = path
-        print(f"📦 Arquivo pronto: {key} ({size_mb:.2f} MB)")
+        print(f"📦 Arquivo pronto: {key}")
     except Exception as e:
         print(f"⚠️ Erro ao baixar {key}: {e}")
 
 api_app = FastAPI(title="Blazer GPU Engine")
-api_app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_credentials=True,
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
+api_app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"])
 
 class ChatMessage(BaseModel):
     role: str
@@ -108,28 +73,15 @@ class ChatCompletionRequest(BaseModel):
     stream: Optional[bool] = True
 
 @spaces.GPU(duration=1)
-def check_hf():
-    return True
-
-# Healthcheck para manter o Ngrok vivo
-@api_app.get("/health")
-def healthcheck():
-    return {"status": "ok"}
+def check_hf(): return True
 
-# ==============================================================================
-# 3. FUNÇÃO DE STREAMING OTIMIZADA PARA QWEN/BLAZER
-# ==============================================================================
-@spaces.GPU(duration=120) # Aumentado para 120s para evitar timeout em códigos longos
+# 4. STREAMING COM TIMEOUT ESTENDIDO E LIMPEZA DE VRAM
+@spaces.GPU(duration=120) 
 def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temperature: float):
-    # LAZY LOADING: Importa apenas quando a GPU já está alocada
     from llama_cpp import Llama
     
     if model_id not in MODEL_PATHS:
-        fallback_model = "blazertiny-1b"
-        if fallback_model in MODEL_PATHS:
-            model_id = fallback_model
-        else:
-            raise ValueError("Nenhum modelo disponível.")
+        model_id = "blazertiny-1b"
     
     print(f"💻 Executando inferência via GPU (ZeroGPU) para: {model_id}")
     
@@ -139,13 +91,12 @@ def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temp
             model_path=MODEL_PATHS[model_id],
             n_ctx=1024,
             n_threads=2,
-            n_gpu_layers=-1, # Todas as camadas na GPU
+            n_gpu_layers=-1,
             verbose=False,
-            use_mmap=True, # Otimização de memória
-            use_mlock=False
+            use_mmap=True
         )
         
-        # Paradas específicas para Qwen/CoT
+        # Template Qwen Oficial + Stop Sequences para evitar vazamento de pensamento
         stop_sequences = ["<|im_end|>", "</think>", "User:", "Assistant:"]
         
         response_stream = llm(
@@ -162,98 +113,66 @@ def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temp
         chunk_id = f"chatcmpl-{int(time.time())}"
         for chunk in response_stream:
             token = chunk["choices"][0]["text"]
-            if token:
-                # Filtra tokens de pensamento
-                if "</think>" in token or "<think>" in token:
-                    continue
-                    
+            if token and "</think>" not in token:
                 data = {
-                    "id": chunk_id,
-                    "object": "chat.completion.chunk",
-                    "created": int(time.time()),
-                    "model": model_id,
-                    "choices": [{
-                        "index": 0,
-                        "delta": {"content": token},
-                        "finish_reason": None
-                    }]
+                    "id": chunk_id, "object": "chat.completion.chunk", "created": int(time.time()),
+                    "model": model_id, "choices": [{"index": 0, "delta": {"content": token}, "finish_reason": None}]
                 }
                 yield f"data: {json.dumps(data)}\n"
         
-        fim = {
-            "id": chunk_id,
-            "object": "chat.completion.chunk",
-            "created": int(time.time()),
-            "model": model_id,
-            "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]
-        }
-        yield f"data: {json.dumps(fim)}\n"
+        yield f"data: {json.dumps({'id': chunk_id, 'object': 'chat.completion.chunk', 'created': int(time.time()), 'model': model_id, 'choices': [{'index': 0, 'delta': {}, 'finish_reason': 'stop'}]})}\n"
         yield "data: [DONE]\n"
         
     except Exception as e:
-        print(f"🔴 Erro Crítico no Worker: {e}")
-        print(traceback.format_exc())
+        print(f"🔴 Erro Crítico: {e}")
         raise RuntimeError(str(e))
     finally:
-        if llm: 
-            del llm
+        if llm: del llm
         gc.collect()
-        # Força liberação de memória CUDA
-        import torch
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        # Limpeza extra de VRAM
+        try:
+            import torch
+            if torch.cuda.is_available(): torch.cuda.empty_cache()
+        except: pass
 
 @api_app.post("/v1/chat/completions")
 async def chat_completions(request: ChatCompletionRequest):
     try:
-        # Template Qwen Oficial
-        system_prompt = "You are Blazer, a helpful coding assistant. Answer directly and concisely. Do not show your thinking process."
-        
+        # System Prompt direto para evitar recusas ou alucinações
+        system_prompt = "You are Blazer, a helpful assistant. Answer directly and concisely."
         prompt_formatado = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
         
         for msg in request.messages:
             role = msg.role.lower() if msg.role else "user"
-            content = msg.content.strip()
-            prompt_formatado += f"<|im_start|>{role}\n{content}<|im_end|>\n"
+            prompt_formatado += f"<|im_start|>{role}\n{msg.content.strip()}<|im_end|>\n"
             
         prompt_formatado += "<|im_start|>assistant\n"
         
         return StreamingResponse(
-            stream_generator(
-                model_id=request.model,
-                prompt_formatado=prompt_formatado,
-                max_tokens=request.max_tokens,
-                temperature=request.temperature
-            ),
+            stream_generator(model_id=request.model, prompt_formatado=prompt_formatado, max_tokens=request.max_tokens, temperature=request.temperature),
             media_type="text/event-stream"
         )
     except Exception as e:
-        print(f"🔴 ERRO NA API: {e}")
         raise HTTPException(status_code=500, detail=str(e))
 
 @api_app.get("/v1/models")
 async def listar_modelos():
-    opcoes = [{"id": chave, "object": "model", "created": 1677652288, "owned_by": "user"} for chave in MODEL_PATHS.keys()]
-    return {"object": "list", "data": opcoes}
+    return {"object": "list", "data": [{"id": k, "object": "model", "created": 1677652288, "owned_by": "user"} for k in MODEL_PATHS.keys()]}
 
 def run_fastapi():
     uvicorn.run(api_app, host="127.0.0.1", port=8000, log_level="warning")
 
 def iniciar_ngrok():
-    if not NGROK_TOKEN: 
-        print("⚠️ NGROK_TOKEN ausente.")
-        return
+    if not NGROK_TOKEN: return
     time.sleep(15)
     ngrok.set_auth_token(NGROK_TOKEN)
     for tentativa in range(5):
         try:
-            ngrok.kill()
-            time.sleep(3)
+            ngrok.kill(); time.sleep(3)
             public_url = ngrok.connect(8000, proto="http", bind_tls=True)
             print(f"\n🔗 URL POCKETPAL: {public_url.public_url}\n")
             return
         except Exception as e:
-            print(f"⚠️ Falha Ngrok {tentativa + 1}: {e}")
             time.sleep(10)
 
 with gr.Blocks() as demo:
@@ -262,7 +181,6 @@ with gr.Blocks() as demo:
 if __name__ == "__main__":
     try: check_hf()
     except: pass
-    
     threading.Thread(target=run_fastapi, daemon=True).start()
     threading.Thread(target=iniciar_ngrok, daemon=True).start()
     demo.launch(server_name="0.0.0.0", server_port=7860)