Update app.py

app.py

@@ -1,10 +1,13 @@
 import os
 import sys
 import logging
+import warnings
 
-# 1. SUPRESSÃO DE RUÍDO NOS LOGS
+# 1. SUPRESSÃO TOTAL DE RUÍDO NOS LOGS
+warnings.filterwarnings("ignore")
 logging.getLogger("asyncio").setLevel(logging.CRITICAL)
 logging.getLogger("gradio").setLevel(logging.ERROR)
+logging.getLogger("httpx").setLevel(logging.CRITICAL)
 
 # 2. CONFIGURAÇÃO CUDA ZERO-GPU
 cuda_paths = [
@@ -40,11 +43,11 @@ 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}
+    "blazerrhino-3b": {"repo_id": "Davizig10jojo/BlazerRhino-3B-GGUF", "filename": "BlazerRhino-3B-Instruct.Q4_K_M.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},
+    "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},
+    "blazerstandard-4b": {"repo_id": "Davizig10jojo/BlazerStandard-4B-GGUF", "filename": "blazerstandard-4b-Q6_K.gguf", "min_size_mb": 2000}
 }
 
 print("📥 Mapeando caminhos dos arquivos locais...")
@@ -52,7 +55,8 @@ MODEL_PATHS = {}
 for key, meta in MODEL_REGISTRY.items():
     try:
         path = hf_hub_download(repo_id=meta["repo_id"], filename=meta["filename"])
-        if os.path.getsize(path) / (1024 * 1024) < meta["min_size_mb"]: raise Exception("Arquivo corrompido")
+        if os.path.getsize(path) / (1024 * 1024) < meta["min_size_mb"]: 
+            raise Exception("Arquivo corrompido ou incompleto")
         MODEL_PATHS[key] = path
         print(f"📦 Arquivo pronto: {key}")
     except Exception as e:
@@ -73,13 +77,15 @@ class ChatCompletionRequest(BaseModel):
     stream: Optional[bool] = True
 
 @spaces.GPU(duration=1)
-def check_hf(): return True
+def check_hf(): 
+    return True
 
-# 4. STREAMING COM TIMEOUT ESTENDIDO E LIMPEZA DE VRAM
+# 4. STREAMING OTIMIZADO
 @spaces.GPU(duration=120) 
 def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temperature: float):
     from llama_cpp import Llama
     
+    # Fallback inteligente: se o modelo falhar ou não existir, usa o Tiny-1b que é mais estável
     if model_id not in MODEL_PATHS:
         model_id = "blazertiny-1b"
     
@@ -96,15 +102,15 @@ def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temp
             use_mmap=True
         )
         
-        # Template Qwen Oficial + Stop Sequences para evitar vazamento de pensamento
-        stop_sequences = ["<|im_end|>", "</think>", "User:", "Assistant:"]
+        # Stops agressivos para evitar alucinação e vazamento de pensamento
+        stop_sequences = ["<|im_end|>", "</think>", "User:", "Assistant:", "I'm sorry", "I cannot"]
         
         response_stream = llm(
             prompt_formatado,
             max_tokens=max_tokens,
             temperature=temperature,
             top_p=0.95,
-            repeat_penalty=1.1,
+            repeat_penalty=1.15, # Penaliza repetições
             stop=stop_sequences,
             echo=False,
             stream=True
@@ -113,7 +119,7 @@ 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 and "</think>" not in token:
+            if token and "</think>" not in 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}]
@@ -124,22 +130,25 @@ def stream_generator(model_id: str, prompt_formatado: str, max_tokens: int, temp
         yield "data: [DONE]\n"
         
     except Exception as e:
-        print(f"🔴 Erro Crítico: {e}")
+        print(f"🔴 Erro Crítico no Worker: {e}")
         raise RuntimeError(str(e))
     finally:
-        if llm: del llm
+        if llm: 
+            del llm
         gc.collect()
-        # Limpeza extra de VRAM
         try:
             import torch
-            if torch.cuda.is_available(): torch.cuda.empty_cache()
-        except: pass
+            if torch.cuda.is_available(): 
+                torch.cuda.empty_cache()
+        except: 
+            pass
 
 @api_app.post("/v1/chat/completions")
 async def chat_completions(request: ChatCompletionRequest):
     try:
-        # System Prompt direto para evitar recusas ou alucinações
-        system_prompt = "You are Blazer, a helpful assistant. Answer directly and concisely."
+        # SYSTEM PROMPT BLINDADO: Força o português e proíbe recusas
+        system_prompt = "Você é o Blazer, um assistente prestativo. Responda SEMPRE em português brasileiro de forma direta, curta e natural. NUNCA diga que não entende ou peça mais informações. Apenas responda ao cumprimento ou pergunta."
+        
         prompt_formatado = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
         
         for msg in request.messages:
@@ -149,10 +158,16 @@ async def chat_completions(request: ChatCompletionRequest):
         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")
@@ -163,12 +178,15 @@ def run_fastapi():
     uvicorn.run(api_app, host="127.0.0.1", port=8000, log_level="warning")
 
 def iniciar_ngrok():
-    if not NGROK_TOKEN: return
+    if not NGROK_TOKEN: 
+        print("⚠️ NGROK_TOKEN ausente.")
+        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
@@ -176,11 +194,13 @@ def iniciar_ngrok():
             time.sleep(10)
 
 with gr.Blocks() as demo:
-    gr.Markdown("# 🦏 Blazer API Hub [ZeroGPU + Qwen]")
+    gr.Markdown("# 🦏 Blazer API Hub [ZeroGPU Otimizado]")
 
 if __name__ == "__main__":
-    try: check_hf()
-    except: pass
+    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)