bartowski

/

Llama-3.2-3B-Instruct-uncensored-GGUF

like 35

I am trying to load this in llama-cpp-python. Any help is appreciated. I downloaded the Q4_K_M.

class LLMHandler:
def init(self):
try:
model_path = str(config.models.LLAMA_PATH)
if not os.path.exists(model_path):
raise FileNotFoundError(f"LLaMA model not found at: {model_path}")

        if torch.cuda.is_available():
            logger.info(f"CUDA available: {torch.cuda.get_device_name()}")
            logger.info(f"CUDA memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB")
            torch.cuda.empty_cache()
            gc.collect()
        
        logger.info("Initializing LLaMA with Q4_K_M specific settings")
        
        # Settings specifically for Q4_K_M
        self.model = Llama(
            model_path=model_path,
            n_ctx=2048,
            n_batch=512,
            n_threads=4,
            n_gpu_layers=20,     # Partial GPU offload
            seed=42,
            use_mmap=True,
            use_mlock=False,
            main_gpu=0,
            tensor_split=None,
            vocab_only=False,
            use_float16=True,    # Enable float16 for Q4_K
            rope_freq_base=500000,  # From model metadata
            rope_freq_scale=1.0,
            n_gqa=8,            # From model metadata
            rms_norm_eps=1e-5,  # From model metadata
            verbose=True
        )
        
        logger.info("Model initialized, testing...")
        test = self.model.create_completion("Test", max_tokens=1)
        logger.info("Model test successful")
        
    except Exception as e:
        import traceback
        detailed_error = f"Failed to initialize LLaMA: {str(e)}\n"
        detailed_error += f"Traceback: {traceback.format_exc()}"
        logger.error(detailed_error)
        raise RuntimeError(detailed_error)

async def generate(self, prompt: str) -> str:
    try:
        response = self.model.create_completion(
            prompt,
            max_tokens=config.system.max_tokens,
            temperature=config.system.temperature,
            top_p=config.system.top_p,
            stop=["Human:", "Assistant:"],
            stream=True
        )
        
        full_response = ""
        for chunk in response:
            if chunk.choices[0].text:
                full_response += chunk.choices[0].text
                
        return full_response.strip()
    except Exception as e:
        logger.error(f"Error generating response: {e}")
        raise