xllm910test

This repository validates the patched Ascend MatmulReduceScatterV2 operator in quay.io/jd_xllm/xllm-ai:xllm-dev-a3-arm-cann9-202607027.

The test calls torch_npu.npu_mm_reduce_scatter_base directly. It does not clone or build xLLM, launch a model service, or require model weights.

Documents

• Create the container
• Run the MMRS validation

Quick start on the A3 server

git clone https://github.com/Cooofish/xllm910test.git
cd xllm910test

# Change these two host paths if needed.
export HOST_REPO_DIR="$PWD"
export HOST_ARTIFACT_DIR="$PWD/artifacts"
bash scripts/create_container.sh

docker exec -it xllm910test bash
cd /workspace/xllm910test
bash scripts/collect_environment.sh
bash scripts/validate_mmrs_patch.sh

The full matrix needs eight idle NPUs. A failed AICore case stops the run immediately. Restart the container and verify that the devices are clean before running another case.