llama2-7b-chat量化完推理报错

[AlexMa0]

Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:42<00:00, 21.33s/it]
Some weights of the model checkpoint at /data/AutoSmoothQuant/quantized_model/llama2-7b-chat/Llama-2-7b-chat-hf-smoothquant were not used when initializing Int8LlamaForCausalLM: ['model.layers.0.mlp.down_proj.quant_scale', 'model.layers.0.self_attn.o_proj.quant_scale', 'model.layers.1.mlp.down_proj.quant_scale', 'model.layers.1.self_attn.o_proj.quant_scale', 'model.layers.10.mlp.down_proj.quant_scale', 'model.layers.10.self_attn.o_proj.quant_scale', 'model.layers.11.mlp.down_proj.quant_scale', 'model.layers.11.self_attn.o_proj.quant_scale', 'model.layers.12.mlp.down_proj.quant_scale', 'model.layers.12.self_attn.o_proj.quant_scale', 'model.layers.13.mlp.down_proj.quant_scale', 'model.layers.13.self_attn.o_proj.quant_scale', 'model.layers.14.mlp.down_proj.quant_scale', 'model.layers.14.self_attn.o_proj.quant_scale', 'model.layers.15.mlp.down_proj.quant_scale', 'model.layers.15.self_attn.o_proj.quant_scale', 'model.layers.16.mlp.down_proj.quant_scale', 'model.layers.16.self_attn.o_proj.quant_scale', 'model.layers.17.mlp.down_proj.quant_scale', 'model.layers.17.self_attn.o_proj.quant_scale', 'model.layers.18.mlp.down_proj.quant_scale', 'model.layers.18.self_attn.o_proj.quant_scale', 'model.layers.19.mlp.down_proj.quant_scale', 'model.layers.19.self_attn.o_proj.quant_scale', 'model.layers.2.mlp.down_proj.quant_scale', 'model.layers.2.self_attn.o_proj.quant_scale', 'model.layers.20.mlp.down_proj.quant_scale', 'model.layers.20.self_attn.o_proj.quant_scale', 'model.layers.21.mlp.down_proj.quant_scale', 'model.layers.21.self_attn.o_proj.quant_scale', 'model.layers.22.mlp.down_proj.quant_scale', 'model.layers.22.self_attn.o_proj.quant_scale', 'model.layers.23.mlp.down_proj.quant_scale', 'model.layers.23.self_attn.o_proj.quant_scale', 'model.layers.24.mlp.down_proj.quant_scale', 'model.layers.24.self_attn.o_proj.quant_scale', 'model.layers.25.mlp.down_proj.quant_scale', 'model.layers.25.self_attn.o_proj.quant_scale', 'model.layers.26.mlp.down_proj.quant_scale', 'model.layers.26.self_attn.o_proj.quant_scale', 'model.layers.27.mlp.down_proj.quant_scale', 'model.layers.27.self_attn.o_proj.quant_scale', 'model.layers.28.mlp.down_proj.quant_scale', 'model.layers.28.self_attn.o_proj.quant_scale', 'model.layers.29.mlp.down_proj.quant_scale', 'model.layers.29.self_attn.o_proj.quant_scale', 'model.layers.3.mlp.down_proj.quant_scale', 'model.layers.3.self_attn.o_proj.quant_scale', 'model.layers.30.mlp.down_proj.quant_scale', 'model.layers.30.self_attn.o_proj.quant_scale', 'model.layers.31.mlp.down_proj.quant_scale', 'model.layers.31.self_attn.o_proj.quant_scale', 'model.layers.4.mlp.down_proj.quant_scale', 'model.layers.4.self_attn.o_proj.quant_scale', 'model.layers.5.mlp.down_proj.quant_scale', 'model.layers.5.self_attn.o_proj.quant_scale', 'model.layers.6.mlp.down_proj.quant_scale', 'model.layers.6.self_attn.o_proj.quant_scale', 'model.layers.7.mlp.down_proj.quant_scale', 'model.layers.7.self_attn.o_proj.quant_scale', 'model.layers.8.mlp.down_proj.quant_scale', 'model.layers.8.self_attn.o_proj.quant_scale', 'model.layers.9.mlp.down_proj.quant_scale', 'model.layers.9.self_attn.o_proj.quant_scale']

• This IS expected if you are initializing Int8LlamaForCausalLM from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
• This IS NOT expected if you are initializing Int8LlamaForCausalLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
  Traceback (most recent call last):
  File "/data/AutoSmoothQuant/autosmoothquant/examples/test_model.py", line 60, in
  main()
  File "/data/conda_env/base/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
  return func(*args, **kwargs)
  File "/data/AutoSmoothQuant/autosmoothquant/examples/test_model.py", line 54, in main
  output_ids = model.generate(**inputs, max_new_tokens=20)
  File "/data/conda_env/base/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
  return func(*args, **kwargs)
  File "/data/conda_env/base/lib/python3.9/site-packages/transformers/generation/utils.py", line 1527, in generate
  result = self._greedy_search(
  File "/data/conda_env/base/lib/python3.9/site-packages/transformers/generation/utils.py", line 2411, in _greedy_search
  outputs = self(
  File "/data/conda_env/base/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
  return forward_call(*args, **kwargs)
  File "/data/conda_env/base/lib/python3.9/site-packages/transformers/models/llama/modeling_llama.py", line 1196, in forward
  outputs = self.model(
  File "/data/conda_env/base/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
  return forward_call(*args, **kwargs)
  File "/data/conda_env/base/lib/python3.9/site-packages/transformers/models/llama/modeling_llama.py", line 990, in forward
  causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
  File "/data/conda_env/base/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1614, in getattr
  raise AttributeError("'{}' object has no attribute '{}'".format(
  AttributeError: 'Int8LlamaModel' object has no attribute '_update_causal_mask'
  推理时报错时什么原因呢？该如何解决？

[AniZpZ]

可能是transformers的版本问题。 建议使用4.36.2版本试试。 或者修改llama.py代码来适配你的transforemrs库

[AlexMa0]

Some weights of the model checkpoint at /data/AutoSmoothQuant/quantized_model/llama2-7b-chat/Llama-2-7b-chat-hf-smoothquant were not used when initializing Int8LlamaForCausalLM: ['model.layers.30.mlp.down_proj.quant_scale', 'model.layers.6.self_attn.o_proj.quant_scale', 'model.layers.5.self_attn.o_proj.quant_scale', 'model.layers.12.mlp.down_proj.quant_scale', 'model.layers.29.mlp.down_proj.quant_scale', 'model.layers.19.self_attn.o_proj.quant_scale', 'model.layers.26.self_attn.o_proj.quant_scale', 'model.layers.3.self_attn.o_proj.quant_scale', 'model.layers.10.mlp.down_proj.quant_scale', 'model.layers.23.self_attn.o_proj.quant_scale', 'model.layers.8.mlp.down_proj.quant_scale', 'model.layers.4.mlp.down_proj.quant_scale', 'model.layers.5.mlp.down_proj.quant_scale', 'model.layers.24.self_attn.o_proj.quant_scale', 'model.layers.27.mlp.down_proj.quant_scale', 'model.layers.18.mlp.down_proj.quant_scale', 'model.layers.29.self_attn.o_proj.quant_scale', 'model.layers.28.self_attn.o_proj.quant_scale', 'model.layers.30.self_attn.o_proj.quant_scale', 'model.layers.15.mlp.down_proj.quant_scale', 'model.layers.20.self_attn.o_proj.quant_scale', 'model.layers.21.mlp.down_proj.quant_scale', 'model.layers.9.self_attn.o_proj.quant_scale', 'model.layers.22.mlp.down_proj.quant_scale', 'model.layers.10.self_attn.o_proj.quant_scale', 'model.layers.28.mlp.down_proj.quant_scale', 'model.layers.23.mlp.down_proj.quant_scale', 'model.layers.25.mlp.down_proj.quant_scale', 'model.layers.14.mlp.down_proj.quant_scale', 'model.layers.9.mlp.down_proj.quant_scale', 'model.layers.7.self_attn.o_proj.quant_scale', 'model.layers.27.self_attn.o_proj.quant_scale', 'model.layers.16.mlp.down_proj.quant_scale', 'model.layers.1.self_attn.o_proj.quant_scale', 'model.layers.14.self_attn.o_proj.quant_scale', 'model.layers.31.self_attn.o_proj.quant_scale', 'model.layers.16.self_attn.o_proj.quant_scale', 'model.layers.11.mlp.down_proj.quant_scale', 'model.layers.20.mlp.down_proj.quant_scale', 'model.layers.2.self_attn.o_proj.quant_scale', 'model.layers.24.mlp.down_proj.quant_scale', 'model.layers.18.self_attn.o_proj.quant_scale', 'model.layers.8.self_attn.o_proj.quant_scale', 'model.layers.26.mlp.down_proj.quant_scale', 'model.layers.17.self_attn.o_proj.quant_scale', 'model.layers.17.mlp.down_proj.quant_scale', 'model.layers.2.mlp.down_proj.quant_scale', 'model.layers.22.self_attn.o_proj.quant_scale', 'model.layers.6.mlp.down_proj.quant_scale', 'model.layers.0.mlp.down_proj.quant_scale', 'model.layers.13.self_attn.o_proj.quant_scale', 'model.layers.4.self_attn.o_proj.quant_scale', 'model.layers.11.self_attn.o_proj.quant_scale', 'model.layers.0.self_attn.o_proj.quant_scale', 'model.layers.1.mlp.down_proj.quant_scale', 'model.layers.21.self_attn.o_proj.quant_scale', 'model.layers.7.mlp.down_proj.quant_scale', 'model.layers.12.self_attn.o_proj.quant_scale', 'model.layers.3.mlp.down_proj.quant_scale', 'model.layers.19.mlp.down_proj.quant_scale', 'model.layers.25.self_attn.o_proj.quant_scale', 'model.layers.31.mlp.down_proj.quant_scale', 'model.layers.13.mlp.down_proj.quant_scale', 'model.layers.15.self_attn.o_proj.quant_scale']

• This IS expected if you are initializing Int8LlamaForCausalLM from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
• This IS NOT expected if you are initializing Int8LlamaForCausalLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
  ["something to say I'keyvalton met im kyaswe, I'keyvalton met New L"]
  使用4.36.2版本,感觉这个输出不正常呀，
  还有就是，如果不更换transformers版本，我该如何修改llama.py文件呢？

[AlexMa0]

问1+1=？回答是这样：
['1+1=?"- c- c- m.Ъ, I'keyval\n wop.Ъ,']

[AniZpZ]

1. 检查calibration数据和过程是否正常
2. 使用per-token而非per-tensor量化

[AlexMa0]

：）hi 大佬
1、使用的数据集是：mit-han-lab/pile-val-backup/val.jsonl.zst
请问校准过程的正不正常的标准是啥呀？
2、我在量化时修改了quantconfig:{
"qkv": "per-token",
"out": "per-token",
"fc1": "per-token",
"fc2": "per-token"
}但是推理的时候依旧是有问题：['1+1=?\n Unterscheidung zwischen 1+1 und 1+2\n\n1+1=2\n']