[sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x #2418

Yangruipis · 2025-07-08T11:56:00Z

What does this PR do?

optimize the performance of sglang+megatron weight update refer to the bucketing implementation of THUDM/slime.

model	bucket size MB	boost
Moonlight16B @ 8xH20	512MB	175s -> 18s
DeepseekV3 671B @ 512xH20	512MB	ONGOING

releated to issues #2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169

similar fixes for FSDP: #2499

We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines.

Checklist Before Starting

Search for similar PRs. Paste at least one query link here: ...
Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
- {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
- 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

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

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.

gemini-code-assist

Code Review

The pull request optimizes the performance of sglang+megatron weight resharding. It includes a backport of the itertools.batched function for older Python versions and modifies the weight update process to use batched updates. I suggested using collections.deque for a potentially more efficient backport implementation.

verl/workers/sharding_manager/megatron_sglang.py

Yangruipis · 2025-07-12T05:09:32Z

since we can't reproduce the performance boost after pull the latest code, we mark it as a WIP pr until we figure it out

zhaochenyang20 · 2025-07-14T05:56:09Z

Great job!

vermouth1992 · 2025-07-14T05:58:13Z

Could you please fix CI? Each file must accompany with a license head

Yangruipis · 2025-07-14T05:59:54Z

Could you please fix CI? Each file must accompany with a license head

fixed

verl/workers/sharding_manager/megatron_sglang.py

Yangruipis · 2025-07-14T09:50:38Z

I'm using a small bucket size to fix CI oom, and followed @hebiao064 ’s advice about the default bucket size. could you please trigger the ci again? thanks! @vermouth1992

zhaochenyang20

LGTM

zhaochenyang20 · 2025-07-15T20:31:33Z

Thanks to Biao. @hebiao064 for this pic. We will use a samller bucket next time.

hebiao064

LGTM, many thanks!

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: #2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

… by 10x (volcengine#2418) ### What does this PR do? optimize the performance of sglang+megatron weight update refer to the bucketing implementation of [`THUDM/slime`](https://github.com/THUDM/slime/blob/fb7605cc5fb09af0f9369d37f7192f12bddee577/slime/ray/ppo_actor.py#L452). |model| bucket size MB |boost | | ---- | ----- | ---- | | Moonlight16B @ 8xH20 | 512MB | 175s -> 18s | |DeepseekV3 671B @ 512xH20| 512MB | ONGOING | releated to issues volcengine#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: volcengine#2499 > We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] 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). --------- Co-authored-by: Stefan He <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine/verl#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

… by 10x (volcengine#2418) ### What does this PR do? optimize the performance of sglang+megatron weight update refer to the bucketing implementation of [`THUDM/slime`](https://github.com/THUDM/slime/blob/fb7605cc5fb09af0f9369d37f7192f12bddee577/slime/ray/ppo_actor.py#L452). |model| bucket size MB |boost | | ---- | ----- | ---- | | Moonlight16B @ 8xH20 | 512MB | 175s -> 18s | |DeepseekV3 671B @ 512xH20| 512MB | ONGOING | releated to issues volcengine#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: volcengine#2499 > We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] 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). --------- Co-authored-by: Stefan He <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

… by 10x (volcengine#2418) ### What does this PR do? optimize the performance of sglang+megatron weight update refer to the bucketing implementation of [`THUDM/slime`](https://github.com/THUDM/slime/blob/fb7605cc5fb09af0f9369d37f7192f12bddee577/slime/ray/ppo_actor.py#L452). |model| bucket size MB |boost | | ---- | ----- | ---- | | Moonlight16B @ 8xH20 | 512MB | 175s -> 18s | |DeepseekV3 671B @ 512xH20| 512MB | ONGOING | releated to issues volcengine#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: volcengine#2499 > We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] 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). --------- Co-authored-by: Stefan He <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

… by 10x (volcengine#2418) ### What does this PR do? optimize the performance of sglang+megatron weight update refer to the bucketing implementation of [`THUDM/slime`](https://github.com/THUDM/slime/blob/fb7605cc5fb09af0f9369d37f7192f12bddee577/slime/ray/ppo_actor.py#L452). |model| bucket size MB |boost | | ---- | ----- | ---- | | Moonlight16B @ 8xH20 | 512MB | 175s -> 18s | |DeepseekV3 671B @ 512xH20| 512MB | ONGOING | releated to issues volcengine#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: volcengine#2499 > We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] 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). --------- Co-authored-by: Stefan He <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine/verl#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

… by 10x (volcengine#2418) ### What does this PR do? optimize the performance of sglang+megatron weight update refer to the bucketing implementation of [`THUDM/slime`](https://github.com/THUDM/slime/blob/fb7605cc5fb09af0f9369d37f7192f12bddee577/slime/ray/ppo_actor.py#L452). |model| bucket size MB |boost | | ---- | ----- | ---- | | Moonlight16B @ 8xH20 | 512MB | 175s -> 18s | |DeepseekV3 671B @ 512xH20| 512MB | ONGOING | releated to issues volcengine#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: volcengine#2499 > We are from the Large Model Post-Training Team of 📕 Xiaohongshu's AI Platform Technology Department , dedicated to developing high-performance, easily-scalable distributed post-training engines. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] 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). --------- Co-authored-by: Stefan He <[email protected]>

@Yangruipis

### What does this PR do? > Add **concise** overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review. Thanks so much to @Yangruipis and @zhuzilin, we implemented the group-wise weights update for SGLang in FSDP. We are still testing the speed up in megtron and FSDP. For megatron: volcengine#2418 At sgl, we're currently exploring two approaches to optimize resharding: 1. **Grouped calls to `update weights from tensor`**: Previously, we called this endpoint for each tensor individually. We're now grouping tensors to reduce the CPU overhead of these calls. 2. **Single large data buffer update**: We're investigating whether we can form a single large data buffer to update a group of tensors all at once. This would reduce the number of times the IPC handler is opened and closed. For the first approach, we're implementing it separately in Megatron and FSDP. I'm starting by merging the FSDP implementation, and then I'll create a common interface for Megatron. We're still evaluating the second approach to see if it's feasible. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### 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). - [x] 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: ... - [x] 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). --------- Co-authored-by: zhaochenyang <[email protected]>

perf: megatron sglang optimize

7124f01

Yangruipis mentioned this pull request Jul 8, 2025

SGLang + Megatron PPO takes a long time to reshard when generate_sequences #2419

Open

gemini-code-assist bot reviewed Jul 8, 2025

View reviewed changes

verl/workers/sharding_manager/megatron_sglang.py Outdated Show resolved Hide resolved

ETOgaosion reviewed Jul 8, 2025

View reviewed changes

verl/workers/sharding_manager/megatron_sglang.py Outdated Show resolved Hide resolved

Yangruipis added 3 commits July 11, 2025 09:50

fix: conflict

08eb4e5

fix

a744429

Merge branch 'main' into fix/wuhuan/megatron_sglang_reshard_perf

909e6ea

Yangruipis marked this pull request as draft July 12, 2025 05:09

hebiao064 mentioned this pull request Jul 12, 2025

RL SGLang Update Weight in Tensor Slow for MOE zhaochenyang20/Awesome-ML-SYS-Tutorial#169

Closed

zhaochenyang20 mentioned this pull request Jul 12, 2025

[sglang] feat: update weights in batch with FSDP #2499

Closed

7 tasks

Yangruipis added 2 commits July 14, 2025 13:12

Merge branch 'main' into fix/wuhuan/megatron_sglang_reshard_perf

1e5a7bc

fix: update by buckets

2ff3f4e

Yangruipis marked this pull request as ready for review July 14, 2025 05:35

Yangruipis requested review from SwordFaith, chenhaiq and zhaochenyang20 as code owners July 14, 2025 05:35

fix: add license

97af9f2

hebiao064 reviewed Jul 14, 2025

View reviewed changes

verl/workers/sharding_manager/megatron_sglang.py Outdated Show resolved Hide resolved

Yangruipis changed the title ~~[sglang, megatron, perf] fix: megatron sglang reshard performance~~ [sglang, megatron, perf] fix: megatron sglang weight update performance Jul 14, 2025

fix: add rollout config and fix ci oom

efb6c58

Yangruipis requested review from PeterSH6, eric-haibin-lin, tongyx361 and vermouth1992 as code owners July 14, 2025 09:36

fix

6b61569

Yangruipis and others added 4 commits July 15, 2025 10:30

fix: conflict

48531a8

fix: codegen pre-commit

27bd1b9

fix: use megabtyes instead of bytes

c196066

Merge branch 'main' into fix/wuhuan/megatron_sglang_reshard_perf

69ea1aa

hebiao064 changed the title ~~[sglang, megatron, perf] fix: megatron sglang weight update performance~~ [sglang, megatron, perf] fix: Speed up megatron sglang weight update by 10x Jul 15, 2025

hebiao064 changed the title ~~[sglang, megatron, perf] fix: Speed up megatron sglang weight update by 10x~~ [sglang, megatron, perf] fix: Speed Up megatron sglang weight update by 10x Jul 15, 2025

hebiao064 changed the title ~~[sglang, megatron, perf] fix: Speed Up megatron sglang weight update by 10x~~ [sglang, megatron, perf] feat: Speed Up megatron sglang weight update by 10x Jul 15, 2025

hebiao064 changed the title ~~[sglang, megatron, perf] feat: Speed Up megatron sglang weight update by 10x~~ [sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x Jul 15, 2025

zhaochenyang20 approved these changes Jul 15, 2025

View reviewed changes

hebiao064 approved these changes Jul 15, 2025

View reviewed changes

zhaochenyang20 merged commit 1fe5daf into volcengine:main Jul 15, 2025
55 of 56 checks passed

zhaochenyang20 mentioned this pull request Jul 15, 2025

[sglang] feat: update weights in batch with FSDP #2559

Merged

7 tasks

This was referenced Jul 24, 2025

Extract update_weights from RL Engine to SGLang to keep simplicity and fix torch reduce sgl-project/sglang#8267

Merged

[fsdp,megatron,sglang] feat: Accelerate and Simplify Update weights logic and bump SGLang to 0.4.9.post6 #2720

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x #2418

[sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x #2418

Uh oh!

Yangruipis commented Jul 8, 2025 •

edited

Loading

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

Uh oh!

Uh oh!

Yangruipis commented Jul 12, 2025

Uh oh!

zhaochenyang20 commented Jul 14, 2025

Uh oh!

vermouth1992 commented Jul 14, 2025

Uh oh!

Yangruipis commented Jul 14, 2025

Uh oh!

Uh oh!

Yangruipis commented Jul 14, 2025

Uh oh!

zhaochenyang20 left a comment

Uh oh!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

hebiao064 left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

[sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x #2418

[sglang, megatron, perf] feat: speed up megatron sglang weight update by 10x #2418

Uh oh!

Conversation

Yangruipis commented Jul 8, 2025 • 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

API and Usage Example

Design & Code Changes

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!

Yangruipis commented Jul 12, 2025

Uh oh!

zhaochenyang20 commented Jul 14, 2025

Uh oh!

vermouth1992 commented Jul 14, 2025

Uh oh!

Yangruipis commented Jul 14, 2025

Uh oh!

Uh oh!

Yangruipis commented Jul 14, 2025

Uh oh!

zhaochenyang20 left a comment

Choose a reason for hiding this comment

Uh oh!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

hebiao064 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Yangruipis commented Jul 8, 2025 •

edited

Loading