[sglang] feat: update weights in batch with FSDP by zhaochenyang20 · Pull Request #2499 · verl-project/verl

zhaochenyang20 · 2025-07-12T21:07:41Z

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 PR introduces batched weight updates for SGLang with FSDP to improve performance. The changes include a new batched utility function and significant refactoring of the update_weights method in FSDPSGLangShardingManager.

My review has identified two critical issues:

The refactored update_weights logic no longer handles DTensor instances correctly, which could lead to incorrect weight updates.
A call to a non-existent batch method on a DataProto object was introduced, which will cause a runtime error.

Both issues need to be addressed before merging.

zhaochenyang20 · 2025-07-12T21:36:59Z

/gemini review

gemini-code-assist

Code Review

This pull request introduces a batched approach for updating model weights in SGLang with FSDP, aiming to improve performance by reducing communication overhead. The core change involves batching tensor serialization and gathering operations. A new batched utility function is added for this purpose.

My review has identified two high-severity issues. One is a behavioral inconsistency in the new batched utility function's fallback implementation compared to the native itertools.batched. The other is a potential for high memory consumption on the gathering rank (TP rank 0), which could lead to Out-of-Memory errors with large batch sizes or a high degree of tensor parallelism.

Tavish9

The project's minimum python version has been upgraded to 3.10.

#2421

Should fix UP006, B905.

Tavish9 · 2025-07-13T06:33:19Z


 import pickle
-from typing import Any, Optional
+from typing import Any, Iterable, Iterator, List, Optional


Suggested change

from typing import Any, Iterable, Iterator, List, Optional

from typing import Any, Iterable, Iterator, Optional

Tavish9 · 2025-07-13T06:33:37Z

        return data
+
+
+def batched(iterable: Iterable, n: int) -> Iterator[List]:


Suggested change

def batched(iterable: Iterable, n: int) -> Iterator[List]:

def batched(iterable: Iterable, n: int) -> Iterator[list]:

Tavish9 · 2025-07-13T06:36:06Z

+                # This groups the serialized parts for each individual tensor across all TP ranks.
+                # Example: from [[(n0, t0_tp0), (n1, t1_tp0)], [(n0, t0_tp1), (n1, t1_tp1)]]
+                # to [ ( (n0, t0_tp0), (n0, t0_tp1) ), ( (n1, t1_tp0), (n1, t1_tp1) ) ]
+                logical_tensors = zip(*gathered_serialized_batches)


Suggested change

logical_tensors = zip(*gathered_serialized_batches)

logical_tensors = zip(*gathered_serialized_batches, strict=True)

hebiao064 · 2025-07-15T18:13:07Z

+        >>> list(batched([1, 2, 3, 4, 5], 2))
+        [[1, 2], [3, 4], [5]]
+    """
+    try:


use chunk based batch like Megatron PR did? #2418

I will. Let's first merge 2418

… by 10x (#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 #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. ```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 <hebiaobuaa@gmail.com>

… by 10x (verl-project#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 verl-project#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: verl-project#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 <hebiaobuaa@gmail.com>

… by 10x (#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 verl-project/verl#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: verl-project/verl#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 <hebiaobuaa@gmail.com>

… by 10x (verl-project#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 verl-project#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: verl-project#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 <hebiaobuaa@gmail.com>

… by 10x (#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 verl-project/verl#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: verl-project/verl#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 <hebiaobuaa@gmail.com>

… by 10x (verl-project#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 verl-project#2419 , sgl-project/sglang#6762 zhaochenyang20/Awesome-ML-SYS-Tutorial#169 similar fixes for FSDP: verl-project#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 <hebiaobuaa@gmail.com>

zhaochenyang20 requested review from SwordFaith and chenhaiq as code owners July 12, 2025 21:07

gemini-code-assist Bot reviewed Jul 12, 2025

View reviewed changes

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

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

zhaochenyang20 changed the title ~~[sglang] feat: update weights in batch~~ [sglang] feat: update weights in batch [WIP] Jul 12, 2025

gemini-code-assist Bot reviewed Jul 12, 2025

View reviewed changes

Comment thread verl/workers/rollout/sglang_rollout/utils.py Outdated

Comment thread verl/workers/sharding_manager/fsdp_sglang.py

zhaochenyang20 changed the title ~~[sglang] feat: update weights in batch [WIP]~~ [sglang] feat: update weights in batch Jul 12, 2025

Tavish9 suggested changes Jul 13, 2025

View reviewed changes

Yangruipis mentioned this pull request Jul 14, 2025

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

Merged

7 tasks

hebiao064 changed the title ~~[sglang] feat: update weights in batch~~ [sglang] feat: update weights in batch with FSDP Jul 15, 2025

hebiao064 reviewed Jul 15, 2025

View reviewed changes

zhaochenyang20 closed this Jul 15, 2025

zhaochenyang20 force-pushed the main branch from ace1a9f to 1fe5daf Compare July 15, 2025 22:25

zhaochenyang20 mentioned this pull request Jul 15, 2025

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

Merged

7 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

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

[sglang] feat: update weights in batch with FSDP#2499
zhaochenyang20 wants to merge 0 commit intoverl-project:mainfrom
zhaochenyang20:main

zhaochenyang20 commented Jul 12, 2025 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

zhaochenyang20 commented Jul 12, 2025

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Tavish9 left a comment •

edited

Loading

Uh oh!

Tavish9 Jul 13, 2025

Uh oh!

Tavish9 Jul 13, 2025

Uh oh!

Tavish9 Jul 13, 2025

Uh oh!

hebiao064 Jul 15, 2025

Uh oh!

zhaochenyang20 Jul 15, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	from typing import Any, Iterable, Iterator, List, Optional
	from typing import Any, Iterable, Iterator, Optional

		return data


		def batched(iterable: Iterable, n: int) -> Iterator[List]:

	logical_tensors = zip(*gathered_serialized_batches)
	logical_tensors = zip(*gathered_serialized_batches, strict=True)

Conversation

zhaochenyang20 commented Jul 12, 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!

zhaochenyang20 commented Jul 12, 2025

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Tavish9 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Tavish9 Jul 13, 2025

Choose a reason for hiding this comment

Uh oh!

Tavish9 Jul 13, 2025

Choose a reason for hiding this comment

Uh oh!

Tavish9 Jul 13, 2025

Choose a reason for hiding this comment

Uh oh!

hebiao064 Jul 15, 2025

Choose a reason for hiding this comment

Uh oh!

zhaochenyang20 Jul 15, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

zhaochenyang20 commented Jul 12, 2025 •

edited

Loading

Tavish9 left a comment •

edited

Loading