torch_npu.npu_quant_matmul

功能描述

完成量化的矩阵乘计算,最小支持输入维度为2维,最大支持输入维度为6维。

接口原型

npu_quant_matmul(Tensor x1, Tensor x2, Tensor scale, *, Tensor? offset=None, Tensor? pertoken_scale=None, Tensor? bias=None, ScalarType? output_dtype=None) -> Tensor

参数说明

输出说明

一个Tensor类型的输出,代表量化matmul的计算结果。如果output_dtype为torch.float16,输出的数据类型为FLOAT16;如果output_dtype为torch.bfloat16,输出的数据类型为BFLOAT16;如果output_dtype为torch.int8或者None,输出的数据类型为INT8;如果output_dtype非以上数据类型,返回错误码。

约束说明

支持的PyTorch版本

支持的型号

调用示例

1.单算子调用:
# Atlas 推理系列产品:在单算子模式下不支持使能高带宽的x2数据排布,如果想追求极致性能,请使用图模式
import torch
import torch_npu
import logging
import os

cpu_x1 = torch.randint(-5, 5, (1, 256, 768), dtype=torch.int8)
cpu_x2 = torch.randint(-5, 5, (31, 768, 16), dtype=torch.int8)
scale = torch.randn(16, dtype=torch.float32)
offset = torch.randn(16, dtype=torch.float32)
bias = torch.randint(-5, 5, (31, 1, 16), dtype=torch.int32)
# Method 1:You can directly call npu_quant_matmul
npu_out = torch_npu.npu_quant_matmul(cpu_x1.npu(), cpu_x2.npu(), scale.npu(), offset=offset.npu(), bias=bias.npu())

# Method 2: You can first call npu_trans_quant_param to convert scale and offset from float32 to int64 when output dtype is torch.int8 or torch.float16
scale_1 = torch_npu.npu_trans_quant_param(scale.npu(), offset.npu())
npu_out = torch_npu.npu_quant_matmul(cpu_x1.npu(), cpu_x2.npu(), scale_1,  bias=bias.npu())


2.图模式:
2.1 通用
import torch
import torch_npu
import torchair as tng
from torchair.ge_concrete_graph import ge_apis as ge
from torchair.configs.compiler_config import CompilerConfig
import logging
from torchair.core.utils import logger
logger.setLevel(logging.DEBUG)
import os
import numpy as np
os.environ["ENABLE_ACLNN"] = "true"
config = CompilerConfig()
npu_backend = tng.get_npu_backend(compiler_config=config)

class MyModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
    def forward(self, x1, x2, scale, offset, bias):
        scale_1 = torch_npu.npu_trans_quant_param(scale, offset)
        return torch_npu.npu_quant_matmul(x1, x2, scale_1, offset=offset, bias=bias)
cpu_model = MyModel()
model = cpu_model.npu()
cpu_x1 = torch.randint(-1, 1, (15, 1, 512), dtype=torch.int8)
cpu_x2 = torch.randint(-1, 1, (15, 512, 128), dtype=torch.int8)
scale = torch.randn(1, dtype=torch.float32)
offset = torch.randn(1, dtype=torch.float32)
bias = torch.randint(-1,1, (15, 1, 128), dtype=torch.int32)
model = torch.compile(cpu_model, backend=npu_backend, dynamic=True)
npu_out = model(cpu_x1.npu(), cpu_x2.npu(), scale.npu(), offset.npu(), bias.npu())

2.2 Atlas 推理系列产品Atlas A2 训练系列产品不可用)高性能调用方式
import torch
import torch_npu
import torchair as tng
from torchair.ge_concrete_graph import ge_apis as ge
from torchair.configs.compiler_config import CompilerConfig
import logging
from torchair.core.utils import logger
logger.setLevel(logging.DEBUG)
import os
import numpy as np
os.environ["ENABLE_ACLNN"] = "true"
config = CompilerConfig()
npu_backend = tng.get_npu_backend(compiler_config=config)

class MyModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
    def forward(self, x1, x2, scale, offset, bias):
        return torch_npu.npu_quant_matmul(x1, x2.transpose(2,1), scale, offset=offset, bias=bias)
cpu_model = MyModel()
model = cpu_model.npu()
cpu_x1 = torch.randint(-1, 1, (15,1, 512), dtype=torch.int8).npu()
cpu_x2 = torch.randint(-1, 1, (15,512, 128), dtype=torch.int8).npu()
# Process x2 into a high-bandwidth format(29) offline to improve performance, please ensure that the input is continuous with (batch,n,k) layout
cpu_x2_t_29 = torch_npu.npu_format_cast(cpu_x2.transpose(2,1).contiguous(), 29)
scale = torch.randn(1, dtype=torch.float32).npu()
offset = torch.randn(1, dtype=torch.float32).npu()
bias = torch.randint(-1,1, (128,), dtype=torch.int32).npu()
# Process scale from float32 to int64 offline to improve performance
scale_1 = torch_npu.npu_trans_quant_param(scale, offset)
model = torch.compile(cpu_model, backend=npu_backend, dynamic=False)
npu_out = model(cpu_x1, cpu_x2_t_29, scale_1, offset, bias)