本页面只读。您可以查看源文件，但不能更改它。如果您觉得这是系统错误，请联系管理员。 ====== transformer模型信息统计 ====== ===== 统计一些LLM模型的每层的参数shape和dtype ===== <code python llm.py> def print_model_info(model_path): model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True) print("=" * 100) print(f"{'参数名称':<60} {'形状':<25} {'精度':<10} {'大小(MB)':<10}") print("=" * 100) total_size = 0 for name, param in model.named_parameters(): shape_str = str(tuple(param.shape)) dtype_str = str(param.dtype).replace('torch.', '') size_mb = param.numel() * param.element_size() / 1024 / 1024 total_size += size_mb print(f"{name:<60} {shape_str:<25} {dtype_str:<10} {size_mb: >8.2f}") </code>