昇腾910b部署qwen-7b-chat进行流式输出【pytorch框架】NPU推理

准备阶段

参考我上一篇文章910b上跑Chatglm3-6b进行流式输出【pytorch框架】

避坑阶段

• 我在mindspore框架下运行了qwen-7b-base、qwen-7b-chat输出都有大大的问题，参考官方文档

qwen-7b-base的输出

qwen-7b-chat的输出

• 像之前给ChatCLM3一样添加代码，会报错 'torch_npu._C._NPUDeviceProperties' object has no attribute 'major'

Traceback (most recent call last):
  File "/home/HwHiAiUser/Qwen/cli_demo.py", line 228, in <module>
    main()
  File "/home/HwHiAiUser/Qwen/cli_demo.py", line 130, in main
    model, tokenizer, config = _load_model_tokenizer(args)
  File "/home/HwHiAiUser/Qwen/cli_demo.py", line 68, in _load_model_tokenizer
    model = AutoModelForCausalLM.from_pretrained(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/models/auto/auto_factory.py", line 553, in from_pretrained
    model_class = get_class_from_dynamic_module(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/dynamic_module_utils.py", line 500, in get_class_from_dynamic_module
    return get_class_in_module(class_name, final_module.replace(".py", ""))
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/dynamic_module_utils.py", line 200, in get_class_in_module
    module = importlib.import_module(module_path)
  File "/home/anaconda3/envs/sakura/lib/python3.9/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1030, in _gcd_import
  File "<frozen importlib._bootstrap>", line 1007, in _find_and_load
  File "<frozen importlib._bootstrap>", line 986, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 680, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 790, in exec_module
  File "<frozen importlib._bootstrap>", line 228, in _call_with_frames_removed
  File "/root/.cache/huggingface/modules/transformers_modules/Qwen-7B-Chat/modeling_qwen.py", line 38, in <module>
    SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/cuda/__init__.py", line 154, in is_bf16_supported
    torch.cuda.get_device_properties(torch.cuda.current_device()).major >= 8
AttributeError: 'torch_npu._C._NPUDeviceProperties' object has no attribute 'major'

解决这个报错能搜到是下面这些

华为官网上的回答，一般都在和稀泥，要么没人理

gitee这个，虽然报错结果一样，但是过程不一样，没有参考价值

Github上看了，推理都是在mindspore框架下进行的，没有参考价值

解决方案

一、modeling_qwen.py

找到第38，39行，替换为

SUPPORT_BF16 = False
SUPPORT_FP16 = True

尝试过SUPPORT_BF16 = True，会报错

Traceback (most recent call last):
  File "/home/HwHiAiUser/Qwen/cli_demo.py", line 228, in <module>
    main()
  File "/home/HwHiAiUser/Qwen/cli_demo.py", line 212, in main
    for response in model.chat_stream(tokenizer, query, history=history, generation_config=config):
  File "/root/.cache/huggingface/modules/transformers_modules/Qwen-7B-Chat/modeling_qwen.py", line 1216, in stream_generator
    for token in self.generate_stream(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 35, in generator_context
    response = gen.send(None)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers_stream_generator/main.py", line 931, in sample_stream
    outputs = self(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/root/.cache/huggingface/modules/transformers_modules/Qwen-7B-Chat/modeling_qwen.py", line 1045, in forward
    transformer_outputs = self.transformer(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/root/.cache/huggingface/modules/transformers_modules/Qwen-7B-Chat/modeling_qwen.py", line 824, in forward
    inputs_embeds = self.wte(input_ids)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/modules/sparse.py", line 162, in forward
    return F.embedding(
  File "/home/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/nn/functional.py", line 2233, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: call aclnnEmbedding failed, detail:EZ1001: weight not implemented for DT_BFLOAT16, should be in dtype support list [DT_DOUBLE,DT_FLOAT,DT_FLOAT16,DT_INT64,DT_INT32,DT_INT16,DT_INT8,DT_UINT8,DT_BOOL,DT_COMPLEX128,DT_COMPLEX64,].

二、cli_demo.py

注释加载模型的resume_download

model = AutoModelForCausalLM.from_pretrained(
    args.checkpoint_path,
    device_map=device_map,
    trust_remote_code=True,
    # resume_download=True,
).npu().eval()

结果展示

• 速度很慢，比mindspore慢了一倍，但是好歹能正常输出了，虽然还输出错了。
• 搞笑的是mindspore框架推理ChatGLM3背诵出师表只背诵前半部分，Qwen这里背诵了后半部分，Torch框架下ChatGLM3可以背出全文。