MiniCPM4-0.5B-QAT-Int4-GPTQ-format · 模型库

from modelscope import AutoTokenizer from vllm import LLM, SamplingParams model_name = "OpenBMB/MiniCPM4-0.5B-QAT-Int4-GPTQ-format" prompt = [{"role": "user", "content": "推荐5个北京的景点。"}] tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) input_text = tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=True) llm = LLM( model=model_name, quantization="gptq_marlin", trust_remote_code=True, max_num_batched_tokens=32768, dtype="bfloat16", gpu_memory_utilization=0.8, ) sampling_params = SamplingParams(top_p=0.7, temperature=0.7, max_tokens=1024, repetition_penalty=1.02) outputs = llm.generate(prompts=input_text, sampling_params=sampling_params) print(outputs[0].outputs[0].text)