📺 Info Siaran: verl project/verl/pull/

ETOgaosion · 2025-07-18T06:40:14Z

What does this PR do?

[docker] feat: upgrade to torch 2.7, sglang 0.4.8

Stage 2: vllm 0.9.1
Stage 3: mcore 0.13.0

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 upgrades torch and sglang to newer versions, along with related dependency updates in setup.py and documentation. The changes are mostly straightforward. However, I've identified a redundant import in verl/workers/rollout/sglang_rollout/sglang_rollout.py which could impact future maintainability. Please see the detailed comment.

wuxibin89 · 2025-07-18T08:11:41Z

      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
    container:
-      image: verlai/verl:app-verl0.4-sglang0.4.6.post5-vllm0.8.5-mcore0.12.1
+      image: verlai/verl:app-verl0.5-sglang0.4.8-mcore0.12.2-te2.2


Is vllm removed in new image?

Yes, we split vllm and sglang image, since the different dependency(torch 2.7.0 and 2.7.1), is there requirements of collocation?

No, should we update the vllm image as well?

Yes, I split into 2 PRs as they are functional independent.

zhaochenyang20 · 2025-07-19T11:01:30Z

We are dyingly looking for SGLang 0.4.9+

ETOgaosion · 2025-07-21T02:29:12Z

@zhaochenyang20 transformers==4.53.2 requirement is not possible to support for now. I can use another PR to show the CI result. If the old transformers is compatible, we can consider upgrade.

ETOgaosion · 2025-07-22T08:58:18Z

There are possibilities of SGLang failure:

(WorkerDict pid=74770) [2025-07-22 08:48:28 TP1] Scheduler hit an exception: Traceback (most recent call last):
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 2645, in run_scheduler_process
(WorkerDict pid=74770)     scheduler.event_loop_overlap()
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
(WorkerDict pid=74770)     return func(*args, **kwargs)
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 756, in event_loop_overlap
(WorkerDict pid=74770)     result = self.run_batch(batch)
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 1679, in run_batch
(WorkerDict pid=74770)     model_worker_batch = batch.get_model_worker_batch()
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/schedule_batch.py", line 1694, in get_model_worker_batch
(WorkerDict pid=74770)     else self.seq_lens.cpu()
(WorkerDict pid=74770) RuntimeError: CUDA error: an illegal memory access was encountered
(WorkerDict pid=74770) CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.

zhaochenyang20 · 2025-07-22T14:41:44Z

There are possibilities of SGLang failure:

(WorkerDict pid=74770) [2025-07-22 08:48:28 TP1] Scheduler hit an exception: Traceback (most recent call last):
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 2645, in run_scheduler_process
(WorkerDict pid=74770)     scheduler.event_loop_overlap()
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
(WorkerDict pid=74770)     return func(*args, **kwargs)
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 756, in event_loop_overlap
(WorkerDict pid=74770)     result = self.run_batch(batch)
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/scheduler.py", line 1679, in run_batch
(WorkerDict pid=74770)     model_worker_batch = batch.get_model_worker_batch()
(WorkerDict pid=74770)   File "/usr/local/lib/python3.10/dist-packages/sglang/srt/managers/schedule_batch.py", line 1694, in get_model_worker_batch
(WorkerDict pid=74770)     else self.seq_lens.cpu()
(WorkerDict pid=74770) RuntimeError: CUDA error: an illegal memory access was encountered
(WorkerDict pid=74770) CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.

A usual fix to SGLang illegal mem access is to upgrade sglang version. 😂

zhaochenyang20 · 2025-07-23T22:27:38Z

@ETOgaosion how is the process for now?

ETOgaosion · 2025-07-24T02:00:28Z

@zhaochenyang20 I think no problem for merge now, but CI is unstable.

### What does this PR do? [docker] feat: upgrade to torch 2.7, sglang 0.4.8 Stage 2: vllm 0.9.1 Stage 3: mcore 0.13.0 ### 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: hebiao064 <hebiaobuaa@gmail.com>

* origin/mindspeed: (39 commits) [perf] feat: add optional role selection in discrete mode for NPU Profiler (verl-project#2750) [rollout] feat: remove chat scheduler (verl-project#2725) [trainer] refactor: Make sure to keep the type checking (verl-project#2634) [doc] style: change resize handle from gradient to plain color (verl-project#2746) [CI] feat: add `mypy` to pre-commit (verl-project#2614) [megatron] feat: a bunch of optimzation on vram, sequence packing (verl-project#2678) [docker] feat: upgrade to torch 2.7, sglang 0.4.8 (verl-project#2617) [doc] feat: add resizable sidebar and improve layout (verl-project#2577) [ci] fix: release ascend test time, fix one step off-policy CI (verl-project#2731) [recipe] chore: add retool training script (verl-project#2732) [ci] fix: checkpoint_convertor ci miss a hf model download (verl-project#2730) [doc] feat: Add agent-lightning in the list of "awesome works using verl (verl-project#2726) [tool] fix: geo3k create return str instead of tuple (verl-project#2714) [megatron] fix: resolve backward propagation error in megatron_actor due to shared logits tensor in-place modification (verl-project#2484) [misc] chore: bump main branch version to v0.5.0.dev (verl-project#2718) [sglang] fix: Adding strict naming sanity for sglang (verl-project#2719) [ray] feat: RayWorkerGroup support set worker env (verl-project#2685) [ci] test: add CriticWorker unit test, make some util CPU friendly (verl-project#2717) [cfg] refactor: add ActorConfig, EngineConfig, and ActorWorker unit test, refactor validation code (verl-project#2621) [misc] chore: bump version to v0.5.0 (verl-project#2716) ...

### What does this PR do? [docker] feat: upgrade to torch 2.7, sglang 0.4.8 Stage 2: vllm 0.9.1 Stage 3: mcore 0.13.0 ### 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: hebiao064 <hebiaobuaa@gmail.com>

ETOgaosion requested review from PeterSH6, SwordFaith, chenhaiq, eric-haibin-lin, tongyx361, vermouth1992 and zhaochenyang20 as code owners July 18, 2025 06:40

gemini-code-assist Bot reviewed Jul 18, 2025

View reviewed changes

Comment thread verl/workers/rollout/sglang_rollout/sglang_rollout.py Resolved

wuxibin89 reviewed Jul 18, 2025

View reviewed changes

ETOgaosion force-pushed the upgrade_sglang_0.4.8 branch from 009b7ae to 92c1dc9 Compare July 21, 2025 02:27

ETOgaosion force-pushed the upgrade_sglang_0.4.8 branch from 174bea0 to c3ac3d2 Compare July 22, 2025 08:35

ETOgaosion changed the title ~~[docker] feat: upgrade to torch 2.7, sglang 0.4.8~~ [docker] feat: upgrade to torch 2.7, sglang 0.4.8, mcore 0.13 Jul 23, 2025

ETOgaosion force-pushed the upgrade_sglang_0.4.8 branch from 25c2759 to 5f54adc Compare July 23, 2025 11:05

ETOgaosion changed the title ~~[docker] feat: upgrade to torch 2.7, sglang 0.4.8, mcore 0.13~~ [docker] feat: upgrade to torch 2.7, sglang 0.4.8 Jul 23, 2025

ETOgaosion force-pushed the upgrade_sglang_0.4.8 branch 2 times, most recently from 341e43f to 662d352 Compare July 24, 2025 07:58

hebiao064 and others added 8 commits July 24, 2025 08:02

Upgrade SGLang 0.4.8

903b876

temp store

639f779

try update images

db99e83

type check no need to install sglang

e44ad1a

fix not robust import

c6537a1

type check need to use sglang image

f076c90

fix transformers

74fefdd

fix convertor

b2ca26b

ETOgaosion added 7 commits July 24, 2025 08:02

pre-commit not work

5497253

fix qwen2_5_vl fused kernel, disable SHM and P2P to solve SGlang

3630b02

fix qwen2_5_vl fused kernel, disable SHM and P2P to solve SGlang

7878c63

fix one step off policy

5465653

fix Qwen2_5_VL new API

7836a59

sglang still go to find vllm

48ae25a

a mess up between sgl and vllm

662d352

zhaochenyang20 approved these changes Jul 24, 2025

View reviewed changes

eric-haibin-lin merged commit 4879d61 into verl-project:main Jul 24, 2025

Navigation Menu

Repository navigation

[docker] feat: upgrade to torch 2.7, sglang 0.4.8 - #2617

[docker] feat: upgrade to torch 2.7, sglang 0.4.8#2617
eric-haibin-lin merged 15 commits into
verl-project:mainfrom
ETOgaosion:upgrade_sglang_0.4.8

ETOgaosion commented Jul 18, 2025 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

wuxibin89 Jul 18, 2025

Uh oh!

ETOgaosion Jul 18, 2025

Uh oh!

wuxibin89 Jul 18, 2025

Uh oh!

ETOgaosion Jul 18, 2025

Uh oh!

zhaochenyang20 commented Jul 19, 2025

Uh oh!

ETOgaosion commented Jul 21, 2025 •

edited

Loading

Uh oh!

ETOgaosion commented Jul 22, 2025

Uh oh!

zhaochenyang20 commented Jul 22, 2025

Uh oh!

zhaochenyang20 commented Jul 23, 2025

Uh oh!

ETOgaosion commented Jul 24, 2025

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

ETOgaosion commented Jul 18, 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!

wuxibin89 Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

ETOgaosion Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

wuxibin89 Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

ETOgaosion Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

zhaochenyang20 commented Jul 19, 2025

Uh oh!

ETOgaosion commented Jul 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ETOgaosion commented Jul 22, 2025

Uh oh!

zhaochenyang20 commented Jul 22, 2025

Uh oh!

zhaochenyang20 commented Jul 23, 2025

Uh oh!

ETOgaosion commented Jul 24, 2025

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

ETOgaosion commented Jul 18, 2025 •

edited

Loading

ETOgaosion commented Jul 21, 2025 •

edited

Loading