[ckpt] feat: add kimi ckpt engine backend by kip-cxj · Pull Request #4954 · verl-project/verl

kip-cxj · 2026-01-16T09:55:24Z

What does this PR do?

Based on ckpt engine abstraction add checkpoint-engine abstraction, in this PR, we add kimi_ckpt_engine backend to support both GPU and huawei Ascend NPU.

Since establishing communication domains across trainer and rollout workers is required, this PR also depends on the newly added communication domain support in kimi_ckpt_engine.

TODO:

Add detailed performance testing results in checkpoint engine README.

Checklist Before Starting

Search for similar PRs. Paste at least one query link here: add Hccl ckpt engine backend
Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
- {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
- If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
- {type} is in feat, fix, refactor, chore, test
- If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
- Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

We have verified the functionality on both GPU and NPU. Performance benchmarks on a 32 NPU environment show promising results; however, due to a lack of available GPU resources, performance data for GPU is still pending.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

Read the Contribute Guide.
Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
Add / Update the documentation.
Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

gemini-code-assist

Code Review

This pull request introduces a new checkpoint engine backend, kimi_ckpt_engine, designed to support both GPU and Huawei Ascend NPU environments. The implementation is comprehensive, including the core engine logic, integration with the existing checkpointing framework, and a new test suite. My review focuses on ensuring correctness and maintainability. I've identified a critical thread-safety issue in the weight sending logic that could lead to data corruption. Additionally, I've suggested a minor but important rename in the test suite to improve clarity.

verl/checkpoint_engine/kimi_checkpoint_engine.py

tests/checkpoint_engine/test_kimi_checkpoint_engine.py

wuxibin89 · 2026-01-21T12:34:51Z

verl/checkpoint_engine/kimi_checkpoint_engine.py

+            with concurrent.futures.ThreadPoolExecutor(max_workers=32) as executor:
+                futures = [
+                    executor.submit(
+                        offload_cpu,


Why we need to offload model to CPU first? Does kimi ckpt engine support D2D?

The kimi_checkpoint_engine only supports reading weights from the CPU. It first register all weights and all gather the metadata across all ranks.

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

wuxibin89 · 2026-02-05T08:04:25Z

Please format code first.

wuxibin89 · 2026-02-05T08:38:10Z

verl/checkpoint_engine/README.md

 |nixl|NIXL|all_gather+ring p2p|Various transport backends (D2D, H2H, H2D, etc)<br>- UCX<br>- UCCL<br>- Mooncacke|Medium/High|High: dynamic adjust ring topology|Off-policy training<br>- Trainer/rollout disaggregated<br>- Elastic rollout<br>- Rollout fault tolerance<br>- Heterogeneous hardware rollout
+|kimi_ckpt_engine|MOONCAKE+NCCL/HCCL|p2p+broadcast|NVIDIA/Ascend|High|Low: rebuild communication group|Off-policy training<br>- Trainer/rollout disaggregated<br>- Save checkpoint each time
+
+PS: kimi_ckpt_engine first offloads all weights to the CPU. Then, using Mooncake transfer engine, these weights are transmitted via P2P to a specific worker in the rollout, followed by a broadcast to all other rollout workers. This mode requires the P2P feature of checkpoint_engine. Please ensure you have installed it via pip install 'checkpoint-engine[p2p]' and that your version is 0.4.0 or higher.


Could you provide more details about how kimi_ckpt_engine works? For example, give an image to show the communication topology between trainer and rollout workers.

I have updated the description and added a schematic. Could you please review it again to see if the level of detail is sufficient? Also, I’ve been scheduled for the H100 environment and will supplement the performance data over the next two days.

The H100 environment was preempted by a higher-priority task, so the performance data will be pending for now. Will update later.

kip-cxj · 2026-02-24T02:39:13Z

@wuxibin89 I have updated the description. Could you please take another look to see if it’s detailed enough to be merged?

### What does this PR do? Based on ckpt engine abstraction [add checkpoint-engine abstraction](verl-project#4775), in this PR, we add kimi_ckpt_engine backend to support both GPU and huawei Ascend NPU. Since establishing communication domains across trainer and rollout workers is required, this PR also depends on the [newly added communication domain support](MoonshotAI/checkpoint-engine#66) in kimi_ckpt_engine. TODO: - [x] Add detailed performance testing results in checkpoint engine README. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: [add Hccl ckpt engine backend](verl-project#4885) - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - If this PR involves multiple modules, separate them with `,` like `[megatron, fsdp, doc]` - `{type}` is in `feat`, `fix`, `refactor`, `chore`, `test` - If this PR breaks any API (CLI arguments, config, function signature, etc.), add `[BREAKING]` to the beginning of the title. - Example: `[BREAKING][fsdp, megatron] feat: dynamic batching` ### Test We have verified the functionality on both GPU and NPU. Performance benchmarks on a 32 NPU environment show promising results; however, due to a lack of available GPU resources, performance data for GPU is still pending. ### Checklist Before Submitting > [!IMPORTANT] > Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review. - [x] Read the [Contribute Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [x] Apply [pre-commit checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting): `pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always` - [x] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add unit or end-to-end test(s) to [the CI workflow](https://github.com/volcengine/verl/tree/main/.github/workflows) to cover all the code. If not feasible, explain why: ... - [ ] Once your PR is ready for CI, send a message in [the `ci-request` channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the `verl` Slack workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [ ] If your PR is related to the `recipe` submodule, please also update the reference to the submodule commit via `git submodule update --remote` or `cd recipe && git pull origin main`. --------- Co-authored-by: kip-cxj <cuixiaojin@huawei.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

gemini-code-assist bot reviewed Jan 16, 2026

View reviewed changes

verl/checkpoint_engine/kimi_checkpoint_engine.py Outdated Show resolved Hide resolved

tests/checkpoint_engine/test_kimi_checkpoint_engine.py Outdated Show resolved Hide resolved

tests/checkpoint_engine/test_kimi_checkpoint_engine.py Outdated Show resolved Hide resolved

wuxibin89 reviewed Jan 21, 2026

View reviewed changes

wuxibin89 mentioned this pull request Jan 21, 2026

[roadmap] verl Q1 roadmap #4880

Open

28 tasks

kip-cxj force-pushed the kimi_ckpt_engine branch from 34b6bab to 5f985f9 Compare January 26, 2026 09:20

kip-cxj added 8 commits February 2, 2026 19:54

add kimi checkpoint engine

a43a8b4

fix bug

8ddd675

fix pre-commit

4215623

fix test case name

b9c1941

add readme

f884f93

fix pre-commit

5a08de2

fix bug

1b49a38

adapt new architecture

ff78792

wuxibin89 mentioned this pull request Feb 2, 2026

[ckpt] feat: add mooncake backend #5176

Open

9 tasks

fix pre-commit

ab708eb

kip-cxj force-pushed the kimi_ckpt_engine branch from 5f985f9 to ab708eb Compare February 3, 2026 11:35

modify ThreadPoolExecutor

2699df4

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

wuxibin89 reviewed Feb 5, 2026

View reviewed changes

kip-cxj force-pushed the kimi_ckpt_engine branch 2 times, most recently from cd6d4c1 to 93a988f Compare February 9, 2026 11:26

modify readme

8dc26f1

kip-cxj force-pushed the kimi_ckpt_engine branch from 93a988f to 8dc26f1 Compare February 24, 2026 08:47

wuxibin89 approved these changes Feb 25, 2026

View reviewed changes

wuxibin89 merged commit 6f4942b into verl-project:main Feb 25, 2026
79 of 87 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ckpt] feat: add kimi ckpt engine backend#4954

[ckpt] feat: add kimi ckpt engine backend#4954
wuxibin89 merged 11 commits intoverl-project:mainfrom
kip-cxj:kimi_ckpt_engine

kip-cxj commented Jan 16, 2026 •

edited

Loading

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wuxibin89 Jan 21, 2026

Uh oh!

kip-cxj Jan 22, 2026

Uh oh!

wuxibin89 commented Feb 5, 2026

Uh oh!

wuxibin89 Feb 5, 2026

Uh oh!

kip-cxj Feb 10, 2026

Uh oh!

kip-cxj Feb 24, 2026

Uh oh!

kip-cxj commented Feb 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

kip-cxj commented Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Checklist Before Starting

Test

Checklist Before Submitting

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wuxibin89 Jan 21, 2026

Choose a reason for hiding this comment

Uh oh!

kip-cxj Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

wuxibin89 commented Feb 5, 2026

Uh oh!

wuxibin89 Feb 5, 2026

Choose a reason for hiding this comment

Uh oh!

kip-cxj Feb 10, 2026

Choose a reason for hiding this comment

Uh oh!

kip-cxj Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

kip-cxj commented Feb 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

kip-cxj commented Jan 16, 2026 •

edited

Loading