语调

语言 中文|[English](https://huggingface.co/YCWTG/Qwen3-Coder-Next-int2-mixed-AutoRound) ## 模型详情 该模型是一个 **mixed-bits INT2 量化** 模型,group_size 为 512,并对 [Qwen/Qwen3-Coder-Next](https://huggingface.co/Qwen/Qwen3-Coder-Next) 采用对称量化,由 [intel/auto-round](https://github.com/intel/auto-round) 生成。请遵循原始模型的 license。 ### 量化策略(Intel MoE Recipe) | 层类型 | Bits | 说明 | | ---------------------------------- | ------ | ------------------- | | Expert layers(512 个 experts) | 2-bit | MoE expert MLPs | | Non-expert layers(attention, gate) | 16-bit | 为保证质量使用更高精度 | | shared_expert_gate | 16-bit | 跳过(shape 不能被 32 整除) | | lm_head | 原始精度 | 被 AutoRound 排除 | ### 模型大小 * **原始 BF16**:~160GB * **mixed INT2**:~25GB(**减少 84%↓↓**) ## 快速开始 ### Transformers 使用方法 在已经安装Transformers的前提下,创建文件example.py,将下方示例代码复制到example.py中,并运行命令 ``` python example.py ``` ```python import math import os os.environ.setdefault( "PYTORCH_ALLOC_CONF", # 保持更安全的 allocator 默认设置,并透明迁移已弃用的环境变量 os.environ.pop("PYTORCH_CUDA_ALLOC_CONF", None) or "expandable_segments:True", ) import torch from transformers import AutoModelForCausalLM, AutoTokenizer MODEL_NAME = "YCWTG/Qwen3-Coder-Next-int2-mixed-AutoRound" AUTO_MAX_TOKENS = True MANUAL_MAX_NEW_TOKENS = 128 AUTO_MAX_TOKENS_RATIO = 1.5 HAS_CUDA = torch.cuda.is_available() # 读取一次 VRAM 总量,并据此选择默认加载模式 GPU_TOTAL_MIB = torch.cuda.get_device_properties(0).total_memory // (1024 ** 2) if HAS_CUDA else 0 # 32GB 级别 GPU 默认 False;更小显存的 GPU 默认 True ENABLE_CPU_OFFLOAD = HAS_CUDA and GPU_TOTAL_MIB < 32000 MAX_MEMORY = {0: "18GiB", "cpu": "64GiB"} if ENABLE_CPU_OFFLOAD else {0: "22GiB", "cpu": "16GiB"} def get_input_device(model): # 使用 device_map="auto" 时,第一个可用设备可能不是 model.device device_map = getattr(model, "hf_device_map", None) cpu_device = None if isinstance(device_map, dict): for loc in device_map.values(): if isinstance(loc, int): return torch.device(f"cuda:{loc}") if isinstance(loc, str): if loc.startswith("cuda"): return torch.device(loc) if loc.startswith("cpu"): cpu_device = torch.device("cpu") return cpu_device or next(model.parameters()).device def load_model(): print("正在加载模型...") tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True, use_fast=True) # 使用 EOS 作为 PAD,避免在没有 pad token 的模型上进行 chat generation 时出现警告 tokenizer.pad_token = tokenizer.pad_token or tokenizer.eos_token tokenizer.padding_side = "left" model_kwargs = { "pretrained_model_name_or_path": MODEL_NAME, "dtype": torch.bfloat16, "trust_remote_code": True, "low_cpu_mem_usage": True, "device_map": "auto" if HAS_CUDA else "cpu", } if HAS_CUDA: print(f"GPU 总显存: {GPU_TOTAL_MIB} MiB") model_kwargs["max_memory"] = MAX_MEMORY if ENABLE_CPU_OFFLOAD: model_kwargs["offload_buffers"] = True print("CPU offload: 开启") else: print("CPU offload: 关闭(优先使用 GPU,允许少量 CPU 溢出)") else: print("CUDA 不可用,在 CPU 上运行") try: model = AutoModelForCausalLM.from_pretrained(**model_kwargs) except RuntimeError as e: # 提供对新手更友好的提示,而不是只显示原始报错堆栈 if "out of memory" in str(e).lower(): print("\n加载模型时发生 CUDA OOM") print("请关闭其他 GPU 程序,或设置 ENABLE_CPU_OFFLOAD = True 后重试") raise model.eval() return model, tokenizer def multiline_input(): print('用户(单独一行输入 "END" 发送,输入 "exit" 退出):') lines = [] while True: line = input() text = line.strip() if text.lower() in {"exit", "quit"}: return None if text == "END": break lines.append(line) return "\n".join(lines) def build_input_ids(tokenizer, messages, device): if getattr(tokenizer, "chat_template", None): # chat 模型的优先路径:由 tokenizer 构建 prompt 格式 prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) else: # 没有内置 chat template 时的通用回退方案 prompt = "\n".join( [f"{'User' if m['role'] == 'user' else 'Assistant'}: {m['content']}" for m in messages] + ["Assistant:"] ) return tokenizer(prompt, return_tensors="pt")["input_ids"].to(device) def chat_loop(model, tokenizer): print("\n===== 对话已开始 =====\n") print(f"自动 max_tokens: {'开启' if AUTO_MAX_TOKENS else '关闭'}") if not AUTO_MAX_TOKENS: print(f"手动 max_new_tokens: {MANUAL_MAX_NEW_TOKENS}") print( "提示:设置 ENABLE_CPU_OFFLOAD = False 可尝试更快的全 GPU 模式" if ENABLE_CPU_OFFLOAD else "提示:若 max_tokens 过大导致 CUDA OOM,可设置 ENABLE_CPU_OFFLOAD = True" ) messages = [] device = get_input_device(model) print(f"输入设备: {device}") while True: user_text = multiline_input() if user_text is None: break messages.append({"role": "user", "content": user_text}) input_ids = build_input_ids(tokenizer, messages, device) prompt_tokens = int(input_ids.shape[-1]) # 自动模式:输出长度随 prompt 长度缩放(默认 1.5 倍) max_new_tokens = max(1, math.ceil(prompt_tokens * AUTO_MAX_TOKENS_RATIO)) if AUTO_MAX_TOKENS else int(MANUAL_MAX_NEW_TOKENS) print(f"Prompt tokens: {prompt_tokens}") print(f"max_new_tokens: {max_new_tokens}") try: with torch.inference_mode(): output_ids = model.generate( input_ids=input_ids, max_new_tokens=max_new_tokens, do_sample=True, temperature=1.0, top_p=0.95, top_k=40, use_cache=False, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id, ) except RuntimeError as e: error_text = str(e).lower() if HAS_CUDA and ("cublas_status_alloc_failed" in error_text or "out of memory" in error_text): # 清理 CUDA 缓存块,使下一次尝试在更干净的 CUDA 状态下开始 torch.cuda.empty_cache() print("\n生成过程中发生 CUDA OOM") print("请设置 ENABLE_CPU_OFFLOAD = True,或关闭 AUTO_MAX_TOKENS 并降低 MANUAL_MAX_NEW_TOKENS") messages.pop() continue raise reply_text = tokenizer.decode(output_ids[0, input_ids.shape[-1]:], skip_special_tokens=True) print(f"\nAssistant:\n{reply_text}\n") messages.append({"role": "assistant", "content": reply_text}) if __name__ == "__main__": model, tokenizer = load_model() chat_loop(model, tokenizer) ``` ## 生成模型 ```python from auto_round import AutoRound model_name = "Qwen/Qwen3-Coder-Next" # 为 mixed-bits 构建 layer config(Intel recipe) layer_config = {} for i in range(48): # 48 层 prefix = f"model.layers.{i}" # Attention 层 -> 16-bit if i in [3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47]: # self_attn 层 for proj in ["q_proj", "k_proj", "v_proj", "o_proj"]: layer_config[f"{prefix}.self_attn.{proj}"] = {"bits": 16} else: # linear_attn 层 -> 16-bit for proj in ["in_proj_qkvz", "in_proj_ba", "out_proj"]: layer_config[f"{prefix}.linear_attn.{proj}"] = {"bits": 16} # MLP gate -> 16-bit layer_config[f"{prefix}.mlp.gate"] = {"bits": 16} # shared_expert_gate -> 16-bit(跳过) layer_config[f"{prefix}.mlp.shared_expert_gate"] = {"bits": 16} autoround = AutoRound( model_name, bits=2, # experts 默认 2-bit group_size=128, sym=True, iters=1000, nsamples=512, lr=2e-3, layer_config=layer_config, low_gpu_mem_usage=True, enable_alg_ext=True ) output_dir="~/model/YCWTG--Qwen3-Coder-Next-int2-mixed-AutoRound" autoround.quantize_and_save(output_dir, format="auto_round") ``` ## 伦理考量与局限性 该模型可能会生成事实不准确的输出,因此不应被依赖用于提供绝对准确的信息。由于 pretrained model 以及 finetuning datasets 的局限性,模型可能生成带有低俗(lewd)、偏见(biased)或其他具有冒犯性的内容。 因此,在部署任何基于该模型的应用之前,开发者应进行安全测试。 ## 注意事项与建议 用户(包括直接用户和下游用户)应充分了解模型的风险、偏见和局限性。 以下是一些用于进一步了解 Intel AI software 的有用链接: * [Intel Neural Compressor](https://github.com/intel/neural-compressor) * [AutoRound](https://github.com/intel/auto-round) ## 免责声明 本模型的 license 不构成法律建议。对于第三方使用本模型所产生的行为,我们不承担责任。在将该模型用于商业用途之前,请咨询律师。 ## 引用 @article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} } [arxiv](https://arxiv.org/abs/2309.05516)