assert m % 4 == 0 and n % 4 == 0 and k % 4 == 0

从 bminf/quantization/init.py 和 traceback 中显示的 cpm_kernels/kernels/gemm.py 调用开始，然后复现 issue 中的 RobertaForQuestionAnswering 示例。检查到达 assertion 的 GEMM 维度，并确定有效维度和无效维度的预期行为；当报告的示例具有明确的结果和回归覆盖时，即表示完成。

import torch
from bminf.wrapper import wrapper
from bminf.scheduler import TransformerBlockList
from transformers import RobertaTokenizer, RobertaForQuestionAnswering

tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = RobertaForQuestionAnswering.from_pretrained('roberta-base')

将模型移动到CUDA设备上

model.to('cuda')

使用wrapper函数对模型进行量化

quantized_model = wrapper(model, quantization=True)

确保quantized_model是一个包含nn.Module的列表

quantized_model_list = [quantized_model] # 将模型封装成列表

print(quantized_model)

创建TransformerBlockList实例

gpus = [(0, None)] # 假设有一个GPU，编号为0

torch.cuda.empty_cache()

创建TransformerBlockList实例

scheduled_model = TransformerBlockList(quantized_model_list, gpus)

文本数据

texts = ["Hello, my dog is cute", "I love natural language processing"]

定义一个函数来确保序列长度是4的倍数

def pad_to_multiple_of_4(input_ids, attention_mask):
seq_len = input_ids.shape[-1]
padding_length = (4 - seq_len % 4) % 4 # 确保padding_length是0到3之间的值
if padding_length > 0:
input_ids = torch.cat([input_ids, torch.zeros((input_ids.shape[0], padding_length), dtype=torch.long)], dim=1)
attention_mask = torch.cat([attention_mask, torch.zeros((attention_mask.shape[0], padding_length), dtype=torch.long)], dim=1)
return input_ids, attention_mask

使用分词器编码文本，并确保序列长度是4的倍数

inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt", padding_side="right", max_length=512)
input_ids = inputs['input_ids']
attention_mask = inputs.get('attention_mask')
input_ids, attention_mask = pad_to_multiple_of_4(input_ids, attention_mask)

验证序列长度

if input_ids.shape[-1] % 4 != 0:
raise ValueError("Input sequence length is not a multiple of 4 after padding.")

执行前向传播

with torch.no_grad():
output = scheduled_model(input_ids, attention_mask=attention_mask)

输出结果

print(output)

Traceback (most recent call last):
File "D:\LLM\模型压缩\BMinf_\demo.py", line 52, in
output = scheduled_model(input_ids, attention_mask=attention_mask)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1736, in wrapped_call_impl
return self.call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1747, in call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\bminf\scheduler_init.py", line 417, in forward
x = sched.forward(x, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\bminf\scheduler_init.py", line 348, in forward
return OpDeviceLayer.apply(self, x, len(kwargs), *cuda_lst)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\autograd\function.py", line 575, in apply
return super().apply(*args, **kwargs) # type: ignore[misc]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\bminf\scheduler_init.py", line 105, in forward
hidden_state = self._layers[i](hidden_state, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 1634, in forward
logits = self.qa_outputs(sequence_output)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1736, in _wrapped_call_impl
return self.call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\nn\modules\module.py", line 1747, in call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\bminf\quantization_init.py", line 81, in forward
out = OpLinear.apply(x, self.weight_quant, self.weight_scale)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\torch\autograd\function.py", line 575, in apply
return super().apply(*args, **kwargs) # type: ignore[misc]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\bminf\quantization_init.py", line 31, in forward
gemm_int8(
File "C:\Users\Zeng xiang xi.conda\envs\LLM\Lib\site-packages\cpm_kernels\kernels\gemm.py", line 139, in gemm_int8
assert m % 4 == 0 and n % 4 == 0 and k % 4 == 0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError