[sglang] feat: update weights in batch with FSDP by zhaochenyang20 · Pull Request #2559 · verl-project/verl

zhaochenyang20 · 2025-07-15T22:56:44Z

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:

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.
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

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

This pull request introduces an optimization for updating weights in FSDP by batching tensor updates. There are two critical issues that must be addressed: the tensor bucketing logic is not synchronized across ranks when using FSDP v1 with sharded state dicts, and the use of zip(..., strict=False) can hide bugs by silently truncating data if batches from different ranks have inconsistent lengths.

zhaochenyang20 · 2025-07-15T23:14:08Z

Move #2499 into this

hebiao064

LGTM with one minor question

zhaochenyang20 · 2025-07-15T23:20:42Z

thanks @hebiao064 @Yangruipis

zhaochenyang20 · 2025-07-15T23:38:44Z

zhaochenyang20/Awesome-ML-SYS-Tutorial#169

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

@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: verl-project#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 <zhaochenyang20@gmail.com>

zhaochen20 and others added 9 commits July 12, 2025 13:59

batch update fsdp weights

f993441

update bash

230cc4b

use infer tp rather than tp [unaligned in megatron]

024d181

use preprocess_tensor_for_update_weights

86030b9

set default value to 16

28e9de7

add raise chain

950b4ed

Merge branch 'main' into main

ace1a9f

Merge branch 'main' of github.com:zhaochenyang20/verl

a7092c8

get_named_tensor_buckets

1378f9f

zhaochenyang20 requested review from PeterSH6, eric-haibin-lin, tongyx361 and vermouth1992 as code owners July 15, 2025 22:56

gemini-code-assist Bot reviewed Jul 15, 2025

View reviewed changes

Comment thread verl/workers/sharding_manager/fsdp_sglang.py

Comment thread verl/workers/sharding_manager/fsdp_sglang.py Outdated

test locally on FSDP

eed6bc5

zhaochenyang20 changed the title ~~Fsdp buffer~~ [sglang] feat: update weights in batch with FSDP Jul 15, 2025

hebiao064 reviewed Jul 15, 2025

View reviewed changes

Comment thread verl/workers/sharding_manager/fsdp_sglang.py Outdated

hebiao064 approved these changes Jul 15, 2025

View reviewed changes

hebiao064 reviewed Jul 15, 2025

View reviewed changes

Comment thread verl/trainer/config/rollout/rollout.yaml Outdated

fix lint

efb7c9c

hebiao064 approved these changes Jul 15, 2025

View reviewed changes

hebiao064 mentioned this pull request Jul 15, 2025

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

Closed

zhaochenyang20 merged commit f0d4c76 into verl-project:main Jul 15, 2025
28 of 29 checks passed

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

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

[sglang] feat: update weights in batch with FSDP#2559
zhaochenyang20 merged 11 commits intoverl-project:mainfrom
zhaochenyang20:fsdp_buffer

zhaochenyang20 commented Jul 15, 2025

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

Uh oh!

hebiao064 left a comment

Uh oh!

Uh oh!

zhaochenyang20 commented Jul 15, 2025 •

edited

Loading

Uh oh!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

zhaochenyang20 commented Jul 15, 2025

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!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

Uh oh!

hebiao064 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

zhaochenyang20 commented Jul 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhaochenyang20 commented Jul 15, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

zhaochenyang20 commented Jul 15, 2025 •

edited

Loading