【昇騰】NPU ID：物理ID、邏輯ID、芯片映射關系

起因：
https://www.hiascend.com/document/detail/zh/Atlas%20200I%20A2/23.0.0/re/npu/npusmi_013.html
npu-smi info -l查詢所有NPU設備：

[naie@notebook-npu-bd130045-55bbffd786-lr6t8 DCNN]$ npu-smi info -lTotal Count                    : 1NPU ID                         : 6Chip Count                     : 1

運行腳本：

import torch_npu
from torch_npu.contrib import transfer_to_npu
import torchimport torch
import torch.nn as nnclass SingleConv(nn.Module):def __init__(self, in_ch, out_ch, kernel_size, stride, padding):super(SingleConv, self).__init__()self.single_conv = nn.Sequential(nn.Conv2d(in_ch, out_ch, kernel_size=kernel_size, padding=padding, stride=stride, bias=False),nn.BatchNorm2d(out_ch),nn.ReLU(inplace=True))def forward(self, x):return self.single_conv(x)class DenseFeaureAggregation(nn.Module):def __init__(self, in_ch, out_ch, base_ch):super(DenseFeaureAggregation, self).__init__()self.conv1 = nn.Sequential(nn.BatchNorm2d(num_features=1 * in_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch, base_ch, dilation=2, kernel_size=3, padding=2, stride=1, bias=True),)self.conv2 = nn.Sequential(nn.BatchNorm2d(num_features=in_ch + base_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch + base_ch, base_ch, dilation=3, kernel_size=3, padding=3, stride=1, bias=True),)self.conv3 = nn.Sequential(nn.BatchNorm2d(num_features=in_ch + 2 * base_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch + 2 * base_ch, base_ch, dilation=5, kernel_size=3, padding=5, stride=1, bias=True),)self.conv4 = nn.Sequential(nn.BatchNorm2d(num_features=in_ch + 3 * base_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch + 3 * base_ch, base_ch, dilation=7, kernel_size=3, padding=7, stride=1, bias=True),)self.conv5 = nn.Sequential(nn.BatchNorm2d(num_features=in_ch + 4 * base_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch + 4 * base_ch, base_ch, dilation=9, kernel_size=3, padding=9, stride=1, bias=True),)self.conv_out = nn.Sequential(nn.BatchNorm2d(num_features=in_ch + 5 * base_ch, eps=1e-5, affine=True),nn.ReLU(inplace=True),nn.Conv2d(in_ch + 5 * base_ch, out_ch, dilation=1, kernel_size=1, padding=0, stride=1, bias=True),)def forward(self, x):out_ = self.conv1(x)concat_ = torch.cat((out_, x), dim=1)out_ = self.conv2(concat_)concat_ = torch.cat((concat_, out_), dim=1)out_ = self.conv3(concat_)concat_ = torch.cat((concat_, out_), dim=1)out_ = self.conv4(concat_)concat_ = torch.cat((concat_, out_), dim=1)out_ = self.conv5(concat_)concat_ = torch.cat((concat_, out_), dim=1)out_ = self.conv_out(concat_)return out_class Encoder(nn.Module):def __init__(self, in_ch, list_ch):super(Encoder, self).__init__()self.encoder_1 = nn.Sequential(SingleConv(in_ch, list_ch[1], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[1], list_ch[1], kernel_size=3, stride=1, padding=1))self.encoder_2 = nn.Sequential(nn.MaxPool2d(kernel_size=2, stride=2, padding=0),SingleConv(list_ch[1], list_ch[2], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[2], list_ch[2], kernel_size=3, stride=1, padding=1))self.encoder_3 = nn.Sequential(nn.MaxPool2d(kernel_size=2, stride=2, padding=0),SingleConv(list_ch[2], list_ch[3], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[3], list_ch[3], kernel_size=3, stride=1, padding=1))self.encoder_4 = nn.Sequential(nn.MaxPool2d(kernel_size=2, stride=2, padding=0),SingleConv(list_ch[3], list_ch[4], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[4], list_ch[4], kernel_size=3, stride=1, padding=1))self.DFA = DenseFeaureAggregation(list_ch[4], list_ch[4], list_ch[4])def forward(self, x):out_encoder_1 = self.encoder_1(x)out_encoder_2 = self.encoder_2(out_encoder_1)out_encoder_3 = self.encoder_3(out_encoder_2)out_encoder_4 = self.encoder_4(out_encoder_3)out_encoder_4 = self.DFA(out_encoder_4)return [out_encoder_1, out_encoder_2, out_encoder_3, out_encoder_4]class Decoder(nn.Module):def __init__(self, out_ch, list_ch):super(Decoder, self).__init__()self.upconv_3_1 = nn.ConvTranspose2d(list_ch[4], list_ch[3], kernel_size=2, stride=2, bias=True)self.decoder_conv_3_1 = nn.Sequential(SingleConv(2 * list_ch[3], list_ch[3], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[3], list_ch[3], kernel_size=3, stride=1, padding=1))self.upconv_2_1 = nn.ConvTranspose2d(list_ch[3], list_ch[2], kernel_size=2, stride=2, bias=True)self.decoder_conv_2_1 = nn.Sequential(SingleConv(2 * list_ch[2], list_ch[2], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[2], list_ch[2], kernel_size=3, stride=1, padding=1))self.upconv_1_1 = nn.ConvTranspose2d(list_ch[2], list_ch[1], kernel_size=2, stride=2, bias=True)self.decoder_conv_1_1 = nn.Sequential(SingleConv(2 * list_ch[1], list_ch[1], kernel_size=3, stride=1, padding=1),SingleConv(list_ch[1], list_ch[1], kernel_size=3, stride=1, padding=1))self.conv_out = nn.Sequential(nn.Conv2d(list_ch[1], out_ch, kernel_size=1, padding=0, bias=True))def forward(self, out_encoder):out_encoder_1, out_encoder_2, out_encoder_3, out_encoder_4 = out_encoderout_decoder_3_1 = self.decoder_conv_3_1(torch.cat((self.upconv_3_1(out_encoder_4), out_encoder_3), dim=1))out_decoder_2_1 = self.decoder_conv_2_1(torch.cat((self.upconv_2_1(out_decoder_3_1), out_encoder_2), dim=1))out_decoder_1_1 = self.decoder_conv_1_1(torch.cat((self.upconv_1_1(out_decoder_2_1), out_encoder_1), dim=1))output = self.conv_out(out_decoder_1_1)return [output]class Model(nn.Module):def __init__(self, in_ch, out_ch, list_ch):super(Model, self).__init__()self.encoder = Encoder(in_ch, list_ch)self.decoder = Decoder(out_ch, list_ch)# initself.initialize()@staticmethoddef init_conv_deconv_BN(modules):for m in modules():if isinstance(m, nn.Conv2d):nn.init.kaiming_uniform_(m.weight, mode='fan_in', nonlinearity='relu')if m.bias is not None:nn.init.constant_(m.bias, 0.)elif isinstance(m, nn.ConvTranspose2d):nn.init.kaiming_uniform_(m.weight, mode='fan_in', nonlinearity='relu')if m.bias is not None:nn.init.constant_(m.bias, 0.)elif isinstance(m, nn.BatchNorm2d):nn.init.constant_(m.weight, 1.)nn.init.constant_(m.bias, 0.)def initialize(self):print('# random init encoder weight using nn.init.kaiming_uniform !')self.init_conv_deconv_BN(self.decoder.modules)print('# random init decoder weight using nn.init.kaiming_uniform !')self.init_conv_deconv_BN(self.encoder.modules)def forward(self, x):out_encoder = self.encoder(x)out_decoder = self.decoder(out_encoder)  # is a listreturn out_decoderimport re
import subprocessdef get_npu_id():try:# 執行命令并捕獲輸出output = subprocess.check_output(['npu-smi', 'info', '-l'], text=True)# 使用正則表達式查找NPU IDmatch = re.search(r'NPU ID\s+:\s+(\d+)', output)if match:return match.group(1)  # 返回匹配的第一個組，即NPU IDelse:return "NPU ID not found"except subprocess.CalledProcessError as e:return f"An error occurred: {e}"network = Model(in_ch=4, out_ch=1,list_ch=[-1, 32, 64, 128, 256])npu_id = get_npu_id()
# list_GPU_ids = [npu_id]
device = torch.device('cuda:' + str(npu_id))
network.to(device)
print("device:",npu_id)

報錯：

Traceback (most recent call last):File "/home/work/user-job-dir/app/notebook/RTDosePrediction-main/RTDosePrediction/Src/DCNN/test_device_id.py", line 211, in <module>network.to(device)File "/home/naie/.local/lib/python3.9/site-packages/torch_npu/contrib/transfer_to_npu.py", line 56, in decoratedreturn fn(*args, **kwargs)File "/home/naie/.local/lib/python3.9/site-packages/torch_npu/utils/module.py", line 68, in toreturn self._apply(convert)File "/home/naie/.local/lib/python3.9/site-packages/torch/nn/modules/module.py", line 810, in _applymodule._apply(fn)File "/home/naie/.local/lib/python3.9/site-packages/torch/nn/modules/module.py", line 810, in _applymodule._apply(fn)File "/home/naie/.local/lib/python3.9/site-packages/torch/nn/modules/module.py", line 810, in _applymodule._apply(fn)[Previous line repeated 2 more times]File "/home/naie/.local/lib/python3.9/site-packages/torch/nn/modules/module.py", line 833, in _applyparam_applied = fn(param)File "/home/naie/.local/lib/python3.9/site-packages/torch_npu/utils/module.py", line 66, in convertreturn t.to(device, dtype if t.is_floating_point() or t.is_complex() else None, non_blocking)File "/home/naie/.local/lib/python3.9/site-packages/torch_npu/contrib/transfer_to_npu.py", line 56, in decoratedreturn fn(*args, **kwargs)
RuntimeError: exchangeDevice:torch_npu/csrc/aten/common/CopyKernel.cpp:37 NPU error, error code is 107001
[ERROR] 2024-12-13-10:47:03 (PID:38196, Device:0, RankID:-1) ERR00100 PTA call acl api failed
[Error]: Invalid device ID.Check whether the device ID is valid.
EE1001: 2024-12-13-10:47:03.815.272 The argument is invalid.Reason: Set device failed, invalid device, set device=6, valid device range is [0, 1)Solution: 1.Check the input parameter range of the function. 2.Check the function invocation relationship.TraceBack (most recent call last):rtSetDevice execute failed, reason=[device id error][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]open device 6 failed, runtime result = 107001.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]

猜想也許是進行了從物理ID到邏輯ID的映射。
查了一下華為的官方文檔：
https://www.hiascend.com/document/detail/zh/Atlas%20200I%20A2/23.0.0/re/npu/npusmi_013.html
還真的存在這么一個映射。
在這里插入圖片描述

遂用這個命令查看了當前環境下的芯片映射關系：

[naie@notebook-npu-bd130045-55bbffd786-lr6t8 DCNN]$ npu-smi info -mNPU ID                         Chip ID                        Chip Logic ID                  Chip Name                     6                              0                              0                              Ascend 910B36                              1                              -                              Mcu

確實物理ID為6的NPU被映射成了0。這是因為當前環境下（notebook）中只存在一個NPU。

但是還有一個問題：什么時候使用物理ID什么時候使用邏輯ID呢？