Files
miles/docker/npu_patch

Miles NPU Patch Installation Guide

This guide provides instructions for installing Miles with NPU support, including all required dependencies and patches.

Component Version Mapping

Component Version/Commit Source
Miles 551d15914c GitHub
SGLang sglang-miles GitHub
SGL Kernel NPU 2026.05.01 GitHub
Megatron-Bridge bridge GitHub
Megatron-LM 3714d81d418c9f1bca4594fc35f9e8289f652862 GitHub
MindSpeed fc63de5c48426dd019c3b3f39e65f5bdf56e4086 GitCode
HDK 25.3.RC1 Ascend
CANN 8.5.0 Ascend

Preparing the Running Environment

Python Version

Only python==3.11 is supported currently.

conda create -n miles_release python=3.11
conda activate miles_release

Working Directory Setup

mkdir <WORKSPACE> && cd <WORKSPACE>

CANN Environment

Prior to start work with miles on Ascend you need to install CANN Toolkit, Kernels operator package and NNAL version 8.5.0, check the installation guide

source <CANN_PATH>/ascend-toolkit/set_env.sh
source <CANN_PATH>/nnal/atb/set_env.sh

PyTorch and PyTorch NPU

pip install torch-npu==2.8.0

Installing Dependencies

SGLang

cd <WORKSPACE>
git clone https://github.com/sgl-project/sglang.git && cd sglang
git checkout sglang-miles
mv python/pyproject.toml python/pyproject.toml.backup
mv python/pyproject_other.toml python/pyproject.toml
pip install -e "python[srt_npu]"
pip install torch-npu==2.8.0

SGL Kernel NPU and Torch Memory Saver

Download Source code(zip) from the release link, then install:

bash bulid.sh
bash build.sh -a memory-saver
pip install output/sgl_kernel_npu*.whl
pip install output/torch_memory_saver*.whl

Megatron-Bridge

pip install git+https://github.com/ISEEKYAN/mbridge.git@89eb10887887bc74853f89a4de258c0702932a1c --no-deps

cd <WORKSPACE>
git clone https://github.com/radixark/Megatron-Bridge.git -b bridge
pip install nvidia-modelopt[torch]>=0.37.0 --no-build-isolation

Megatron-LM

cd <WORKSPACE>
git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
  cd Megatron-LM/ && git checkout 3714d81d418c9f1bca4594fc35f9e8289f652862 && \
  pip install -e .

MindSpeed

cd <WORKSPACE>
git clone https://gitcode.com/Ascend/MindSpeed.git && \
  cd MindSpeed/ && git checkout fc63de5c48426dd019c3b3f39e65f5bdf56e4086 && \
  pip install -e .

Miles

cd <WORKSPACE>
git clone https://github.com/radixark/miles.git && cd miles
cp -r docker/npu_patch ../npu_patch
pip install -e .

Applying Patches

cd <WORKSPACE>/miles
git apply ../npu_patch/miles.patch

cd <WORKSPACE>/sglang
git apply ../npu_patch/sglang.patch

cd <WORKSPACE>/Megatron-LM
git apply ../npu_patch/megatron_common.patch
git apply ../npu_patch/megatron.patch

cd <WORKSPACE>/Megatron-Bridge
git apply ../npu_patch/megatron-bridge.patch

cd <WORKSPACE>/MindSpeed
git apply ../npu_patch/mindspeed.patch

Additional Dependencies

cd <WORKSPACE>/miles
pip install triton-ascend
pip install torch-npu==2.8.0
pip install torchvision==0.23.0
pip install numpy==1.26.0