[megatron, sglang, ckpt, recipe] feat: add MiMo-7B MTP training on Ascend

[RordChang]

What does this PR do?

Adds MiMo-7B-RL MTP training on Ascend NPU with Megatron/MindSpeed for training, SGLang for rollout, and direct HCCL weight synchronization between the two sides.

• Adds the registered sglang_hccl checkpoint engine for direct Megatron-to-SGLang weight updates.
• Pairs each Megatron rank with a different colocated SGLang replica and creates persistent two-rank HCCL process groups.
• Sends complete Hugging Face tensors in bounded buckets through SGLang's distributed weight-update API.
• Restores actor parameter offload after non-naive checkpoint-engine synchronization.
• Extends the canonical MiMo-7B MTP Megatron script with an Ascend mode while retaining the existing NVIDIA defaults.
• Updates the existing MTP and checkpoint-engine documentation.

Related task: verl-project/verl-ascend-recipe#20

Checklist Before Starting

• Search for similar PRs. Query: https://github.com/verl-project/verl/pulls?q=is%3Apr+is%3Aopen+MiMo+MTP+SGLang+Ascend
• Format the PR title as [{modules}] {type}: {description}.

Test

• pre-commit run --all-files --show-diff-on-failure --color=always: all hooks passed.
• CPU validation: 12 passed, 1 warning.
• bash -n examples/mtp_trainer/run_mimo_7b_mtp_megatron.sh: passed.

Experiment Results

Platform: 4 x Ascend 910B.

Metric	Result
Completed training steps	100/100
Reward mean, first 10 steps	-0.8080
Reward mean, last 10 steps	0.0891
Reward change	+0.8971
Reward linear slope	+0.01224/step
MTP loss mean, first 10 steps	0.5979
MTP loss mean, last 10 steps	0.3559
MTP loss change	-0.2420
End-to-end throughput	533.3 token/s
Steady-state throughput	801.9 token/s

At step 100, each SGLang replica received 451 tensors in 12 buckets (15.67 GB) in 1.32-1.37 seconds; checkpoint tracking reached step 100.

Full redacted 100-step training log:
training_100step_sanitized.log

API and Usage Example

DEVICE=npu \
NPUS_PER_NODE=4 \
MODEL_PATH=/path/to/MiMo-7B-RL \
DATA_ROOT=/path/to/math \
bash examples/mtp_trainer/run_mimo_7b_mtp_megatron.sh

DATA_ROOT should contain train.parquet and test.parquet. Additional Hydra overrides can be appended to the command.

The recipe trains the MTP auxiliary head (enable_train=True) while SGLang performs rollout; rollout-side MTP acceleration remains disabled for this training configuration.

Design & Code Changes

• SGLangHCCLCheckpointEngine prepares per-pair HCCL endpoints and streams complete named tensors without splitting a model parameter.
• SGLangHCCLCheckpointEngineManager builds a cyclic derangement between trainer ranks and rollout replicas, initializes the SGLang update groups, and coordinates memory/generation state.
• ActorRolloutRefWorker.update_weights returns offloaded actors to CPU after checkpoint-engine synchronization when parameter offload is enabled.
• The canonical MiMo launcher selects sglang_hccl only for DEVICE=npu; existing CUDA behavior remains unchanged.

Checklist Before Submitting

• 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.
• CPU validation results are reported above.
• Once the PR is ready for CI, send a message in the ci-request channel or the Feishu group.
• The recipe submodule is not changed, so no submodule reference update is required.

AI assistance disclosure: AI tools assisted with implementation, validation, and documentation preparation. The author reviewed the complete diff and takes responsibility for the submission.

[Copilot]

Pull request overview

Adds an Ascend NPU synchronous MiMo-7B MTP training path that performs direct Megatron→SGLang weight updates via paired HCCL process groups, plus CPU-side unit tests and documentation/recipe updates to support the new backend.

Changes:

• Introduces the sglang_hccl checkpoint engine + a dedicated manager to build persistent paired (2-rank) HCCL groups and stream full HF-keyed tensors in bounded buckets.
• Updates the actor worker weight-update path to re-offload parameters to CPU after non-naive checkpoint sync when parameter offload is enabled.
• Extends the canonical MiMo-7B MTP Megatron launcher + docs to support DEVICE=npu, and adds CPU tests for bucketing/topology/HTTP error handling and global-step propagation.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
verl/workers/engine_workers.py	Re-offloads actor parameters after checkpoint-engine sync when param offload is enabled.
verl/checkpoint_engine/sglang_hccl_checkpoint_engine.py	Adds direct paired HCCL Megatron→SGLang weight sync engine + custom manager.
verl/checkpoint_engine/README.md	Documents the new sglang_hccl backend in the supported backends table.
verl/checkpoint_engine/init.py	Exposes the new engine and manager from the checkpoint engine package.
tests/checkpoint_engine/test_sglang_hccl_on_cpu.py	Adds CPU unit tests for bucketing, derangement mapping, HTTP handling, and group setup wiring.
tests/checkpoint_engine/test_global_steps_on_cpu.py	Adds CPU coverage for param re-offload after checkpoint-engine send.
examples/mtp_trainer/run_mimo_7b_mtp_megatron.sh	Extends the canonical MiMo-7B MTP launcher with Ascend mode and HCCL sync configuration.
examples/mtp_trainer/README.md	Documents how to run the Ascend synchronous training mode via DEVICE=npu.
docs/advance/mtp.md	Updates MTP dependency guidance for SGLang and Megatron-Bridge references.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[wucong25]

hccl checkpoint engine的设计是和推理后端无关的，这里单独写了一个sglang的hccl checkpoint engine，建议复用原来的hccl checkpoint engine

[wucong25]

统一先提到verl-ascend-recipe：verl-ascend-recipe哈，此pr暂时关闭