Skip to content

[doc] feat: split GPU/NPU install docs - #515

Merged
SamitHuang merged 2 commits into
verl-project:mainfrom
chi030303:doc/split-gpu-npu-install
Sep 7, 2026
Merged

[doc] feat: split GPU/NPU install docs#515
SamitHuang merged 2 commits into
verl-project:mainfrom
chi030303:doc/split-gpu-npu-install

Conversation

@chi030303

@chi030303 chi030303 commented Sep 2, 2026

Copy link
Copy Markdown
Contributor

What does this PR do?

Fixes #226.

Split docs/start/install.md into hardware-specific pages, following this repo's flowgrpo_quickstart.md / flowgrpo_quickstart_npu.md naming:

  • docs/start/install.md — NVIDIA GPU (existing path kept so links still work)
  • docs/start/install_npu.md — Ascend NPU

Each page is self-contained with a one-line cross-link at the top. The Getting Started toctree lists both. The NPU FlowGRPO quickstart now points at install_npu.md.

On the GPU page, the extras table / FA3 section document kernels==0.16.0 to match the current gpu extra in pyproject.toml (the table on main still says 0.14.1; see #440). This PR does not change Docker/CI pins; #497 is already unifying those to 0.16.0. Do not close #440 from this PRdocker/Dockerfile.cuda's KERNELS_VERSION default and its "defaults match pyproject.toml extras" header comment are still inconsistent with 0.16.0, and that part of #440 is unresolved.

Checklist Before Starting

Test

Docs-only. No new CI, no training.

  • cd docs && make htmlstart/install.html and start/install_npu.html render; cross-links open; GPU extras show kernels==0.16.0
  • tests/special_sanity/check_docs_time_info.py passed
  • SKIP=autogen-trainer-cfg pre-commit run --files docs/start/install.md docs/start/install_npu.md docs/index.md docs/start/flowgrpo_quickstart_npu.md — applicable hooks passed. The skipped hook regenerates trainer configs and needs omegaconf / a full install; this docs-only change does not affect it.

API and Usage Example

No API change. GPU users keep using docs/start/install.md; NPU users use docs/start/install_npu.md.

Design & Code Changes

  • Split the mixed GPU/NPU install page; did not invent new install steps
  • docs/index.md: add start/install_npu.md under Getting Started
  • docs/start/flowgrpo_quickstart_npu.md: install link → install_npu.md

Checklist Before Submitting

  • Read the Contribute Guide
  • pre-commit on the changed docs files (SKIP=autogen-trainer-cfg; see Test)
  • Add / Update the documentation
  • No new unit/e2e CI: documentation-only split of existing pages

AI-assisted contribution disclosure

This PR was prepared with AI assistance (Cursor agent). I (the submitter) reviewed every changed line, ran the test commands listed above myself in this environment, and can defend/discuss the change.

Split docs/start/install.md into an NVIDIA GPU page (kept at the same
path so existing links still work) and a new docs/start/install_npu.md
for Ascend NPU, matching the flowgrpo_quickstart.md /
flowgrpo_quickstart_npu.md split already used in this repo. Each page
is self-contained with a one-line cross-link to the other, and the
NPU-specific Docker section moves to the new page.

On the GPU page, correct the kernels version documented for the gpu
extra from 0.14.1 to 0.16.0 to match pyproject.toml. Docker/CI 0.14.1
overrides (docker/Dockerfile.cuda default, gpu-smoke-prepare action)
are intentionally left as-is since they are documented overrides, not
the project extra default.

Fixes verl-project#226

Relates to verl-project#440: fixes the docs-reported kernels version, but does not
close verl-project#440 since docker/Dockerfile.cuda's KERNELS_VERSION default and
its "defaults match pyproject.toml extras" header comment are still
inconsistent with 0.16.0.

Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: chi030303 <106855944+chi030303@users.noreply.github.com>

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟢 Approval recommended

The doc split is consistent with existing intra-doc linking patterns and the GPU kernels==0.16.0 documentation matches pyproject.toml’s gpu extra.

Pull request overview

This PR splits the mixed GPU/NPU installation documentation into two hardware-specific pages (keeping the existing docs/start/install.md path for GPU link stability) and updates the GPU install page’s documented kernels version to match the gpu extra in pyproject.toml.

Changes:

  • Split install docs into NVIDIA GPU (docs/start/install.md) and Ascend NPU (docs/start/install_npu.md) with cross-links between them.
  • Update GPU install docs to state kernels==0.16.0 for the gpu extra (matching pyproject.toml).
  • Update docs navigation and NPU quickstart to point to the new NPU install page.
File summaries
File Description
docs/start/install.md GPU-specific install page; adds cross-link to NPU page and corrects kernels version to 0.16.0.
docs/start/install_npu.md New NPU-specific install page containing the previously mixed-in NPU instructions and Docker guidance.
docs/start/flowgrpo_quickstart_npu.md Updates the quickstart to link to the new NPU install page.
docs/index.md Adds start/install_npu.md to the Getting Started toctree.
Review details
  • Files reviewed: 4/4 changed files
  • Comments generated: 0
  • Review effort level: Lite

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Keep the hardware-specific pages, and take main's Python 3.11, vLLM 0.28 extras, and FA fail-fast docs.

Signed-off-by: chi030303 <106855944+chi030303@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

@SamitHuang SamitHuang left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@SamitHuang
SamitHuang merged commit 7545a00 into verl-project:main Sep 7, 2026
1 check passed
@chi030303
chi030303 deleted the doc/split-gpu-npu-install branch September 7, 2026 11:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug] CUDA install guide reports the wrong kernels version for the gpu extra [Docs] Split installation documentation into GPU and NPU

3 participants