feat(npu): Enhance Dockerfile.npu with verified vllm commit and run script

[miracle0517]

• Add compatibility handling to Dockerfile.npu for broader environment support.
• Add a training script that uses Ray job submit to launch distributed training jobs.

This improves the build robustness on NPU platforms and provides a unified entry point for job scheduling via Ray.

Verification with qwen3-4b image on A3 machine：

[gemini-code-assist]

Code Review

This pull request introduces a new script run-qwen3-4B-npu-ray.sh to run Qwen3-4B training on NPU using Ray, and updates Dockerfile.npu to refine the installation of vLLM and its Ascend backend. Feedback on the changes includes a critical fix for a missing file operand in a sed command in the Dockerfile, and a security concern regarding globally disabling SSL verification. For the new script, recommendations include using pkill for cleaner process termination, avoiding hardcoded script directories to maintain portability, and adding the --wait flag to the Ray job submission to stream logs synchronously.

[gemini-code-assist]

The sed command is missing the target file operand /etc/apt/sources.list. This will cause the Docker build to fail with a sed: no input files error.

RUN sed -i '/ports\.ubuntu\.com/ {h;s|ports\.ubuntu\.com|mirrors.tuna.tsinghua.edu.cn|g;G}' /etc/apt/sources.list && \

[gemini-code-assist]

The awk '{print $4}' command extracts the process name (or type) instead of the PID from the npu-smi info output, which will cause the subsequent kill -9 command to fail since it expects a numeric PID. Additionally, npu-smi info output format can vary. A more robust and standard way to clean up the Ray worker processes is to use pkill -9 -f rayWorker.

Suggested change

	npu-smi info 2>/dev/null | grep rayWorker | awk '{print $4}' | xargs -r kill -9 2>/dev/null || true
	pkill -9 -f rayWorker 2>/dev/null || true

[gemini-code-assist]

Disabling SSL verification globally (git config --global http.sslVerify false) is a security risk as it exposes the build process to Man-in-the-Middle (MITM) attacks when cloning repositories. Since all the cloned repositories are from public hosts like GitHub, SSL verification should remain enabled. If a specific proxy or environment requires it, consider configuring the specific CA certificate instead.

[gemini-code-assist]

Hardcoding SCRIPT_DIR to /root/vime/scripts/ reduces the portability of the script. It is better to use the dynamic directory resolution logic (which is currently commented out) so that the script works regardless of where the repository is cloned.

Suggested change

	# SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
	SCRIPT_DIR="/root/vime/scripts/"
	SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"

[gemini-code-assist]

By default, ray job submit submits the job asynchronously and returns immediately. To ensure the script blocks until the training job completes and streams the logs to stdout, you should pass the --wait flag.

Suggested change

	ray job submit --address="http://127.0.0.1:${RAY_DASHBOARD_PORT}" \
	--runtime-env-json="${RUNTIME_ENV_JSON}" \
	--working-dir="/root/vime" \
	-- python3 train.py \
	ray job submit --address="http://127.0.0.1:${RAY_DASHBOARD_PORT}" \
	--runtime-env-json="${RUNTIME_ENV_JSON}" \
	--working-dir="/root/vime" \
	--wait \
	-- python3 train.py \

[read-the-docs-community]

Documentation build overview

📚 vime | 🛠️ Build #33281415 | 📁 Comparing 315794f against latest (e62d44f)

  🔍 Preview build  

28 files changed · + 1 added · ± 26 modified · - 1 deleted

+ Added

• get_started/NPU.html

± Modified

• genindex.html
• index.html
• search.html
• advanced/arch-support-beyond-megatron.html
• advanced/fault-tolerance.html
• advanced/megatron-config.html
• advanced/pd-disaggregation.html
• advanced/reproducibility.html
• advanced/speculative-decoding.html
• advanced/vllm-config.html
• and 16 more...

- Deleted

• _examples_synced/tau-bench/README.html

[CalvinXKY]

Conflicts need fixing; DCO error — commit with sign-off.

[CalvinXKY]

To limit the number of SHs, we need to modify the existing script based on the GPU example and use Ray for submission.
TODO left: keep only one SH per model type.

[CalvinXKY]

List verification tests performed in the description.

[Meihan-chen]

LGTM