[trainer] feat: add mindspeedmm backend engine support on NPU.support Qwen3.5-27B、Qwen3.5-35B by OneMondy · Pull Request #6199 · verl-project/verl

OneMondy · 2026-04-29T02:58:46Z

… Qwen3.5-27B、Qwen3.5-35B

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.

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, veomni, sglang, vllm, vllm_omni, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
- 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. (If not accessible, please try the Feishu group (飞书群).)
If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

… Qwen3.5-27B、Qwen3.5-35B Co-authored-by: pengnuoheng <18720048515@163.com>

CLAassistant · 2026-04-29T02:58:54Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

gemini-code-assist

Code Review

This pull request introduces support for MindSpeed FSDP and refactors the MindSpeed Megatron backend, including new training scripts for Qwen3.5 models and a dedicated MindSpeed optimizer configuration. Key changes involve updating MindSpeedEngineConfig to support multiple strategies, enhancing VLM support through improved 3D position ID handling in TensorDict, and implementing MindSpeedFSDPEngineWithLMHead. Review feedback correctly identified several typos in strategy names (minspeed_megatron) within the shell scripts and missing imports for torch, re, and shutil in the MindSpeed engine implementation.

gemini-code-assist · 2026-04-29T03:02:03Z

-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_num_layers=1
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.overlap_grad_reduce=True
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.overlap_param_gather=True
+    actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron


Typo in strategy name: minspeed_megatron should be mindspeed_megatron. This will cause an assertion failure in MindSpeedEngineConfig.__post_init__. Additionally, use + as the prefix for overriding configuration values in Hydra.

Suggested change

actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron

+actor_rollout_ref.actor.mindspeed.strategy=mindspeed_megatron

References

Use + instead of ++ as the prefix for overriding configuration values in Hydra.

gemini-code-assist · 2026-04-29T03:02:03Z

-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_method=uniform
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_granularity=full
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_num_layers=1
+    actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron


Typo in strategy name: minspeed_megatron should be mindspeed_megatron. Additionally, use + as the prefix for overriding configuration values in Hydra.

Suggested change

actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron

+actor_rollout_ref.actor.mindspeed.strategy=mindspeed_megatron

References

Use + instead of ++ as the prefix for overriding configuration values in Hydra.

gemini-code-assist · 2026-04-29T03:02:03Z

-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_method=uniform
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_granularity=full
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_num_layers=1
+    actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron


Typo in strategy name: minspeed_megatron should be mindspeed_megatron. Additionally, use + as the prefix for overriding configuration values in Hydra.

Suggested change

actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron

+actor_rollout_ref.actor.mindspeed.strategy=mindspeed_megatron

References

Use + instead of ++ as the prefix for overriding configuration values in Hydra.

gemini-code-assist · 2026-04-29T03:02:03Z

-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.recompute_num_layers=1
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.overlap_grad_reduce=True
-    +actor_rollout_ref.actor.mindspeed.llm_kwargs.overlap_param_gather=True
+    actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron


Typo in strategy name: minspeed_megatron should be mindspeed_megatron. Additionally, use + as the prefix for overriding configuration values in Hydra.

Suggested change

actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron

+actor_rollout_ref.actor.mindspeed.strategy=mindspeed_megatron

References

Use + instead of ++ as the prefix for overriding configuration values in Hydra.

gemini-code-assist · 2026-04-29T03:02:03Z

+        if not self.mm_args.training.no_save_optim:
+            state["optimizer"] = self.optimizer
+        if not self.mm_args.training.no_save_rng:
+            state["extra_state"]["torch_rng_state"] = torch.get_rng_state()


The torch module is not imported in this scope. This will cause a NameError when attempting to save the RNG state.

Suggested change

state["extra_state"]["torch_rng_state"] = torch.get_rng_state()

import torch

state["extra_state"]["torch_rng_state"] = torch.get_rng_state()

gemini-code-assist · 2026-04-29T03:02:03Z

+            if "lr_scheduler" in state["extra_state"]:
+                self.lr_scheduler.load_state_dict(state["extra_state"]["lr_scheduler"])
+            if not self.mm_args.training.no_load_rng and "torch_rng_state" in state["extra_state"]:
+                torch.set_rng_state(state["extra_state"]["torch_rng_state"])


The torch module is not imported in this scope. This will cause a NameError when attempting to load the RNG state.

Suggested change

torch.set_rng_state(state["extra_state"]["torch_rng_state"])

import torch

torch.set_rng_state(state["extra_state"]["torch_rng_state"])

gemini-code-assist · 2026-04-29T03:02:03Z

+    def _cleanup_old_checkpoints(self, base_path, max_to_keep):
+        iter_pattern = re.compile(r"iter_(\d+)")


The re module is not imported in this file. This will cause a NameError when attempting to compile the iteration pattern.

Suggested change

def _cleanup_old_checkpoints(self, base_path, max_to_keep):

iter_pattern = re.compile(r"iter_(\d+)")

def _cleanup_old_checkpoints(self, base_path, max_to_keep):

import re

iter_pattern = re.compile(r"iter_(\d+)")

gemini-code-assist · 2026-04-29T03:02:03Z

+            if os.path.isdir(old_dir):
+                shutil.rmtree(old_dir)


The shutil module is not imported in this file. This will cause a NameError when attempting to remove old checkpoint directories.

Suggested change

if os.path.isdir(old_dir):

shutil.rmtree(old_dir)

if os.path.isdir(old_dir):

import shutil

shutil.rmtree(old_dir)

wuxibin89 · 2026-04-29T03:07:27Z

Too many changes are coupled into this PR, please move all changes except for mindspeedllm to separate PR.

[trainer] feat: add mindspeedmm backend engine support on NPU.support…

2131603

… Qwen3.5-27B、Qwen3.5-35B Co-authored-by: pengnuoheng <18720048515@163.com>

OneMondy requested review from PeterSH6, eric-haibin-lin, ji-huazhong, tardis-key, tongyx361, vermouth1992 and wucong25 as code owners April 29, 2026 02:58

gemini-code-assist Bot reviewed Apr 29, 2026

View reviewed changes

wuxibin89 added the Ascend label Apr 29, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[trainer] feat: add mindspeedmm backend engine support on NPU.support Qwen3.5-27B、Qwen3.5-35B#6199

[trainer] feat: add mindspeedmm backend engine support on NPU.support Qwen3.5-27B、Qwen3.5-35B#6199
OneMondy wants to merge 1 commit intoverl-project:mainfrom
OneMondy:main

OneMondy commented Apr 29, 2026

Uh oh!

CLAassistant commented Apr 29, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Uh oh!

wuxibin89 commented Apr 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	actor_rollout_ref.actor.mindspeed.strategy=minspeed_megatron
	+actor_rollout_ref.actor.mindspeed.strategy=mindspeed_megatron

	state["extra_state"]["torch_rng_state"] = torch.get_rng_state()
	import torch
	state["extra_state"]["torch_rng_state"] = torch.get_rng_state()

	torch.set_rng_state(state["extra_state"]["torch_rng_state"])
	import torch
	torch.set_rng_state(state["extra_state"]["torch_rng_state"])

		def _cleanup_old_checkpoints(self, base_path, max_to_keep):
		iter_pattern = re.compile(r"iter_(\d+)")

Conversation

OneMondy commented Apr 29, 2026

What does this PR do?

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

CLAassistant commented Apr 29, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

wuxibin89 commented Apr 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants