Skip to content

[cfg] feat: support NPU profiling - #507

Draft
SelfDLer wants to merge 1 commit into
verl-project:mainfrom
SelfDLer:profiler-config
Draft

[cfg] feat: support NPU profiling#507
SelfDLer wants to merge 1 commit into
verl-project:mainfrom
SelfDLer:profiler-config

Conversation

@SelfDLer

@SelfDLer SelfDLer commented Sep 1, 2026

Copy link
Copy Markdown

What does this PR do?

Add NPU profiler configuration for Ascend training.

The shared profiler config now exposes NPUToolConfig, allowing diffusion and Omni trainers to select npu profiling and configure collection contents, level, analysis mode, database layout, and response-token profiling bounds. The generated trainer reference configs are updated to include the same fields.

Checklist Before Starting

Test

  • scripts/generate_trainer_config.sh - passed (All good).
  • git diff --check upstream/main...HEAD - passed.
  • PR_TITLE='[cfg] feat: support NPU profiling' python tests/special_sanity/check_pr_title.py - passed.
  • pre-commit run --all-files --show-diff-on-failure --color=always - not fully runnable in this Windows checkout: upstream symlinks are materialized as link-target text and two repository hooks require POSIX sh. No failures came from this PR's YAML diff.

API and Usage Example

global_profiler:
  tool: npu

Role-level options are available under actor_rollout_ref.<role>.profiler.tool_config.npu.

Design & Code Changes

  • Add npu to the supported profiler tool choices.
  • Add the NPU tool configuration fields to the shared role profiler config.
  • Regenerate diffusion, VeOmni diffusion, Omni FSDP, and Omni Megatron reference configs.

Checklist Before Submitting

  • Read the contribution guide.
  • Human submitter has reviewed every changed line and can defend the change end-to-end.
  • Full pre-commit completed in a POSIX checkout with symlink support.
  • Config generation check passed.
  • The config comments and generated reference YAML document the new fields.

AI assistance (OpenAI Codex) was used to prepare this Draft PR and run checks.

Expose the upstream NPU profiler configuration so Ascend runs can select NPU profiling without local config overrides.

AI assistance (OpenAI Codex) was used to reorganize this change.

Co-authored-by: OpenAI Codex <noreply@openai.com>
Signed-off-by: leovzhang <coffeeofzl@gmail.com>
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.

1 participant