torch_npu.contrib.module.LinearWeightQuant

功能描述

LinearWeightQuant是对torch_npu接口torch_npu.npu_weight_quant_batchmatmul的封装类,完成矩阵乘计算中的weight输入和输出的量化操作,支持pertensor,perchannel,pergroup多场景量化。

接口原型

torch_npu.contrib.module.LinearWeightQuant(in_features, out_features, bias=True, device=None, dtype=None, antiquant_offset=False, quant_scale=False, quant_offset=False, antiquant_group_size=0)

参数说明

输入说明

变量说明

输出说明

输出为Device侧Tensor类型,代表计算结果。当输入存在quant_scale时输出数据类型为INT8,当输入不存在quant_sclae时输出数据类型和输入x一致。

约束说明

支持的PyTorch版本

支持的型号

Atlas A2 训练系列产品

调用示例

单算子模式:
import torch
import torch_npu
import torchair as tng
from torch_npu.contrib.module import LinearWeightQuant

x = torch.randn((8192, 320),device='npu',dtype=torch.bfloat16)
weight = torch.randn((320, 256),device='npu',dtype=torch.int8)
antiquantscale = torch.randn((1, 256),device='npu',dtype=torch.bfloat16)
antiquantoffset = torch.randn((1, 256),device='npu',dtype=torch.bfloat16)
quantscale = torch.randn((1, 256),device='npu',dtype=torch.float)
quantoffset = torch.randn((1, 256),device='npu',dtype=torch.float)

model = LinearWeightQuant(in_features=320,
​                 out_features=256,
​                 bias=False,
​                 dtype=torch.bfloat16,
​                 antiquant_offset=True,
​                 quant_scale=True,
​                 quant_offset=True,
​                 antiquant_group_size=0,
​                 device=torch.device(f'npu:0')
​                 )
model.npu()
model.weight.data = weight
model.antiquant_scale.data = antiquantscale
model.antiquant_offset.data = antiquantoffset
model.quant_scale.data = quantscale
model.quant_offset.data = quantoffset
tng.experimental.inference.use_internal_format_weight(model)
out = model.(x)

图模式:
import torch
import torch_npu
import torchair as tng
from torch_npu.contrib.module import LinearWeightQuant
from torchair.configs.compiler_config import CompilerConfig

config = CompilerConfig()
config.debug.graph_dump.type = "pbtxt"
npu_backend = tng.get_npu_backend(compiler_config=config)

x = torch.randn((8192, 320),device='npu',dtype=torch.bfloat16)
weight = torch.randn((320, 256),device='npu',dtype=torch.int8)
antiquantscale = torch.randn((1, 256),device='npu',dtype=torch.bfloat16)
antiquantoffset = torch.randn((1, 256),device='npu',dtype=torch.bfloat16)
quantscale = torch.randn((1, 256),device='npu',dtype=torch.float)
quantoffset = torch.randn((1, 256),device='npu',dtype=torch.float)

model = LinearWeightQuant(in_features=320,
​                 out_features=256,
​                 bias=False,
​                 dtype=torch.bfloat16,
​                 antiquant_offset=True,
​                 quant_scale=True,
​                 quant_offset=True,
​                 antiquant_group_size=0,
​                 device=torch.device(f'npu:0')
​                 )
model.npu()
model.weight.data = weight
model.antiquant_scale.data = antiquantscale
model.antiquant_offset.data = antiquantoffset
model.quant_scale.data = quantscale
model.quant_offset.data = quantoffset
tng.experimental.inference.use_internal_format_weight(model)
model = torch.compile(model, backend=npu_backend, dynamic=False)
out = model.(x)