[megatron] feat: Using MTP in RL Training and Inference by ArronHZG · Pull Request #4936 · verl-project/verl

ArronHZG · 2026-01-15T09:15:45Z

Guide to Using MTP in RL Training and Inference

Author: https://github.com/meituan-search

Last Updated: 2026/01/14

1. Scope of Support

Currently, RL training can be performed on mimo-7B-RL, Qwen-next, and Deepseek series models based on the MTP architecture. The support rules for training and inference engines are as follows:

Training Engine: Only supports the mbridge + megatron combination; other training engines are not compatible at this time;
Inference Engine: Compatible with all engines, but the model must be in the corresponding engine's compatibility list;
Dependency Versions:
- mbridge: Use the specified branch: https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp (will be merged into the main branch in the future);
- megatron: Use the latest dev version (commit: 23e092f41ec8bc659020e401ddac9576c1cfed7e), which supports MTP + CP training methods.

2. MTP Training Configuration (Core Parameters)

The MTP training process can be flexibly controlled through the following configurations. All configurations are based on the actor_rollout_ref.model.mtp prefix:

Configuration Scenario	Core Parameters	Description
Load MTP Parameters Only	`enable=True`	VRAM usage will increase, but the exported parameters include the MTP module and can be directly used for online deployment
Full-Parameter MTP Training	`enable=True` `enable_train=True` `mtp_loss_scaling_factor=0.1`	MTP Loss will apply to all model parameters
MTP Parameter-Only Training	`enable=True` `enable_train=True` `detach_encoder=True`	Freeze the Encoder layer, update only MTP module parameters, MTP Loss applies only to MTP parameters
MTP Accelerated Rollout	1. vLLM configuration: `enable=True` `enable_rollout=True` `method="mtp"` `num_speculative_tokens=1` 2. SGLang configuration: `enable=True` `enable_rollout=True` `speculative_algorithm="EAGLE"` `speculative_num_steps=2` `speculative_eagle_topk=2` `speculative_num_draft_tokens=4`	Achieve inference acceleration during the Rollout phase based on MTP

3. Experimental Results

The experiment was conducted as follows:

model = mimo-7B-math
max_response_length = 8k

Experiment chart:

Scenarios with No Significant Effect

The following configurations will not have a noticeable impact on training results:

The base model does not carry MTP parameters;
The base model carries MTP parameters, but the MTP module is not trained;
The base model carries MTP parameters and trains MTP, with mtp_loss_scaling_factor=0;
The base model carries MTP parameters, trains MTP and detaches the encoder, with mtp_loss_scaling_factor=0.1.

Scenarios with Significant Effect

Only the following configuration will have a noticeable impact on training results:

The base model carries MTP parameters, MTP Loss applies to all model parameters, and mtp_loss_scaling_factor=0.1.

Recommended Training Method

It is recommended to adopt the detach_encoder=True approach for MTP training.

4. Performance Notes for MTP in Rollout Inference

The effectiveness of MTP-accelerated Rollout is significantly affected by model size and inference hardware. Key reference information is as follows:

Hardware Tensor Core Performance

Hardware Model	FP16 Performance (TFLOPS)
H20	148
H800	1,671
H200	1,979

Measured Performance and Recommendations

Taking the mimo-7B model deployed separately on H20 hardware using SGLang as an example: After enabling MTP speculative decoding, the Rollout throughput decreases by approximately 50%.

Current priority recommendation: Do not enable MTP acceleration during the inference phase for now;
Future planning: Further optimization of the speculative logic in the Rollout phase will be conducted to improve throughput performance.

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, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
- 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.

ISEEKYAN · 2026-01-19T08:27:46Z

Great job

some further TODOs:

Compatible with megatron-bridge
Compatible with model engine
Use as less patches as possible by contributing some code back to megatron-core.
More experiments on industry size models, such as deepseekv3, qwen3next.

…#4936) # Guide to Using MTP in RL Training and Inference **Author**: `https://github.com/meituan-search` **Last Updated**: 2026/01/14 # 1. Scope of Support Currently, RL training can be performed on mimo-7B-RL, Qwen-next, and Deepseek series models based on the MTP architecture. The support rules for training and inference engines are as follows: - **Training Engine**: Only supports the `mbridge + megatron` combination; other training engines are not compatible at this time; - **Inference Engine**: Compatible with all engines, but the model must be in the corresponding engine's compatibility list; - **Dependency Versions**: - mbridge: Use the specified branch: [https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp](https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp) (will be merged into the main branch in the future); - megatron: Use the latest dev version (commit: [23e092f41ec8bc659020e401ddac9576c1cfed7e](https://github.com/NVIDIA/Megatron-LM/tree/23e092f41ec8bc659020e401ddac9576c1cfed7e)), which supports MTP + CP training methods. # 2. MTP Training Configuration (Core Parameters) The MTP training process can be flexibly controlled through the following configurations. All configurations are based on the `actor_rollout_ref.model.mtp` prefix: | Configuration Scenario | Core Parameters | Description | |------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------| | Load MTP Parameters Only | `enable=True` | VRAM usage will increase, but the exported parameters include the MTP module and can be directly used for online deployment | | Full-Parameter MTP Training | `enable=True` `enable_train=True` `mtp_loss_scaling_factor=0.1` | MTP Loss will apply to all model parameters | | MTP Parameter-Only Training | `enable=True` `enable_train=True` `detach_encoder=True` | Freeze the Encoder layer, update only MTP module parameters, MTP Loss applies only to MTP parameters | | MTP Accelerated Rollout | 1. vLLM configuration: `enable=True` `enable_rollout=True` `method="mtp"` `num_speculative_tokens=1` 2. SGLang configuration: `enable=True` `enable_rollout=True` `speculative_algorithm="EAGLE"` `speculative_num_steps=2` `speculative_eagle_topk=2` `speculative_num_draft_tokens=4` | Achieve inference acceleration during the Rollout phase based on MTP | # 3. Experimental Results The experiment was conducted as follows: * model = mimo-7B-math * max_response_length = 8k Experiment chart: ![fully_async_policy_revenue]( https://github.com/ArronHZG/verl-community/blob/main/docs/mimo-7b-mtp.png?raw=true) **Scenarios with No Significant Effect** The following configurations will not have a noticeable impact on training results: 1. The base model does not carry MTP parameters; 2. The base model carries MTP parameters, but the MTP module is not trained; 3. The base model carries MTP parameters and trains MTP, with `mtp_loss_scaling_factor=0`; 4. The base model carries MTP parameters, trains MTP and detaches the encoder, with `mtp_loss_scaling_factor=0.1`. **Scenarios with Significant Effect** Only the following configuration will have a noticeable impact on training results: - The base model carries MTP parameters, MTP Loss applies to all model parameters, and `mtp_loss_scaling_factor=0.1`. **Recommended Training Method** It is recommended to adopt the `detach_encoder=True` approach for MTP training. # 4. Performance Notes for MTP in Rollout Inference The effectiveness of MTP-accelerated Rollout is significantly affected by **model size** and **inference hardware**. Key reference information is as follows: **Hardware Tensor Core Performance** | Hardware Model | FP16 Performance (TFLOPS) | |----------------|---------------------------| | H20 | 148 | | H800 | 1,671 | | H200 | 1,979 | **Measured Performance and Recommendations** Taking the mimo-7B model deployed separately on H20 hardware using SGLang as an example: After enabling MTP speculative decoding, the Rollout throughput decreases by approximately 50%. - Current priority recommendation: Do not enable MTP acceleration during the inference phase for now; - Future planning: Further optimization of the speculative logic in the Rollout phase will be conducted to improve throughput performance. > 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 - [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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [x] 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`.

### 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. The SFT training with MTP is supported, using the same MTP training configuration as RL training. An example configuration for running SFT can be found in `examples/sft/gsm8k/run_mimo_megatron_mtp.sh` **SFT result** The experiment was conducted using following data: - model = mimo-7B-math - dataset = gsm8k The result: [wandb link](https://wandb.ai/hou-zg-meituan/mimo-7b-sft-mtp?nw=nwuserhouzg) The presence of mtp layer has limited effect on main loss. However, when MTP layer is detached, the mtp_loss converges to a higher value. ### 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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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. ### Design & Code Changes Support SFT training with MTP, with the core change based on #4936 ### 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: Using none-standard mbridge/mcore - [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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [X] 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`. --------- Co-authored-by: ArronHZG <hou.zg@foxmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

### 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. The SFT training with MTP is supported, using the same MTP training configuration as RL training. An example configuration for running SFT can be found in `examples/sft/gsm8k/run_mimo_megatron_mtp.sh` **SFT result** The experiment was conducted using following data: - model = mimo-7B-math - dataset = gsm8k The result: [wandb link](https://wandb.ai/hou-zg-meituan/mimo-7b-sft-mtp?nw=nwuserhouzg) The presence of mtp layer has limited effect on main loss. However, when MTP layer is detached, the mtp_loss converges to a higher value. ### 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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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. ### Design & Code Changes Support SFT training with MTP, with the core change based on verl-project#4936 ### 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: Using none-standard mbridge/mcore - [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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [X] 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`. --------- Co-authored-by: ArronHZG <hou.zg@foxmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

…#4936) # Guide to Using MTP in RL Training and Inference **Author**: `https://github.com/meituan-search` **Last Updated**: 2026/01/14 # 1. Scope of Support Currently, RL training can be performed on mimo-7B-RL, Qwen-next, and Deepseek series models based on the MTP architecture. The support rules for training and inference engines are as follows: - **Training Engine**: Only supports the `mbridge + megatron` combination; other training engines are not compatible at this time; - **Inference Engine**: Compatible with all engines, but the model must be in the corresponding engine's compatibility list; - **Dependency Versions**: - mbridge: Use the specified branch: [https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp](https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp) (will be merged into the main branch in the future); - megatron: Use the latest dev version (commit: [23e092f41ec8bc659020e401ddac9576c1cfed7e](https://github.com/NVIDIA/Megatron-LM/tree/23e092f41ec8bc659020e401ddac9576c1cfed7e)), which supports MTP + CP training methods. # 2. MTP Training Configuration (Core Parameters) The MTP training process can be flexibly controlled through the following configurations. All configurations are based on the `actor_rollout_ref.model.mtp` prefix: | Configuration Scenario | Core Parameters | Description | |------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------| | Load MTP Parameters Only | `enable=True` | VRAM usage will increase, but the exported parameters include the MTP module and can be directly used for online deployment | | Full-Parameter MTP Training | `enable=True` `enable_train=True` `mtp_loss_scaling_factor=0.1` | MTP Loss will apply to all model parameters | | MTP Parameter-Only Training | `enable=True` `enable_train=True` `detach_encoder=True` | Freeze the Encoder layer, update only MTP module parameters, MTP Loss applies only to MTP parameters | | MTP Accelerated Rollout | 1. vLLM configuration: `enable=True` `enable_rollout=True` `method="mtp"` `num_speculative_tokens=1` 2. SGLang configuration: `enable=True` `enable_rollout=True` `speculative_algorithm="EAGLE"` `speculative_num_steps=2` `speculative_eagle_topk=2` `speculative_num_draft_tokens=4` | Achieve inference acceleration during the Rollout phase based on MTP | # 3. Experimental Results The experiment was conducted as follows: * model = mimo-7B-math * max_response_length = 8k Experiment chart: ![fully_async_policy_revenue]( https://github.com/ArronHZG/verl-community/blob/main/docs/mimo-7b-mtp.png?raw=true) **Scenarios with No Significant Effect** The following configurations will not have a noticeable impact on training results: 1. The base model does not carry MTP parameters; 2. The base model carries MTP parameters, but the MTP module is not trained; 3. The base model carries MTP parameters and trains MTP, with `mtp_loss_scaling_factor=0`; 4. The base model carries MTP parameters, trains MTP and detaches the encoder, with `mtp_loss_scaling_factor=0.1`. **Scenarios with Significant Effect** Only the following configuration will have a noticeable impact on training results: - The base model carries MTP parameters, MTP Loss applies to all model parameters, and `mtp_loss_scaling_factor=0.1`. **Recommended Training Method** It is recommended to adopt the `detach_encoder=True` approach for MTP training. # 4. Performance Notes for MTP in Rollout Inference The effectiveness of MTP-accelerated Rollout is significantly affected by **model size** and **inference hardware**. Key reference information is as follows: **Hardware Tensor Core Performance** | Hardware Model | FP16 Performance (TFLOPS) | |----------------|---------------------------| | H20 | 148 | | H800 | 1,671 | | H200 | 1,979 | **Measured Performance and Recommendations** Taking the mimo-7B model deployed separately on H20 hardware using SGLang as an example: After enabling MTP speculative decoding, the Rollout throughput decreases by approximately 50%. - Current priority recommendation: Do not enable MTP acceleration during the inference phase for now; - Future planning: Further optimization of the speculative logic in the Rollout phase will be conducted to improve throughput performance. > 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 - [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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [x] 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`.

### 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. The SFT training with MTP is supported, using the same MTP training configuration as RL training. An example configuration for running SFT can be found in `examples/sft/gsm8k/run_mimo_megatron_mtp.sh` **SFT result** The experiment was conducted using following data: - model = mimo-7B-math - dataset = gsm8k The result: [wandb link](https://wandb.ai/hou-zg-meituan/mimo-7b-sft-mtp?nw=nwuserhouzg) The presence of mtp layer has limited effect on main loss. However, when MTP layer is detached, the mtp_loss converges to a higher value. ### 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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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. ### Design & Code Changes Support SFT training with MTP, with the core change based on verl-project#4936 ### 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: Using none-standard mbridge/mcore - [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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [X] 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`. --------- Co-authored-by: ArronHZG <hou.zg@foxmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

…#4936) # Guide to Using MTP in RL Training and Inference **Author**: `https://github.com/meituan-search` **Last Updated**: 2026/01/14 # 1. Scope of Support Currently, RL training can be performed on mimo-7B-RL, Qwen-next, and Deepseek series models based on the MTP architecture. The support rules for training and inference engines are as follows: - **Training Engine**: Only supports the `mbridge + megatron` combination; other training engines are not compatible at this time; - **Inference Engine**: Compatible with all engines, but the model must be in the corresponding engine's compatibility list; - **Dependency Versions**: - mbridge: Use the specified branch: [https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp](https://github.com/ArronHZG/mbridge/tree/feature/verl_mtp) (will be merged into the main branch in the future); - megatron: Use the latest dev version (commit: [23e092f41ec8bc659020e401ddac9576c1cfed7e](https://github.com/NVIDIA/Megatron-LM/tree/23e092f41ec8bc659020e401ddac9576c1cfed7e)), which supports MTP + CP training methods. # 2. MTP Training Configuration (Core Parameters) The MTP training process can be flexibly controlled through the following configurations. All configurations are based on the `actor_rollout_ref.model.mtp` prefix: | Configuration Scenario | Core Parameters | Description | |------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------| | Load MTP Parameters Only | `enable=True` | VRAM usage will increase, but the exported parameters include the MTP module and can be directly used for online deployment | | Full-Parameter MTP Training | `enable=True` `enable_train=True` `mtp_loss_scaling_factor=0.1` | MTP Loss will apply to all model parameters | | MTP Parameter-Only Training | `enable=True` `enable_train=True` `detach_encoder=True` | Freeze the Encoder layer, update only MTP module parameters, MTP Loss applies only to MTP parameters | | MTP Accelerated Rollout | 1. vLLM configuration: `enable=True` `enable_rollout=True` `method="mtp"` `num_speculative_tokens=1` 2. SGLang configuration: `enable=True` `enable_rollout=True` `speculative_algorithm="EAGLE"` `speculative_num_steps=2` `speculative_eagle_topk=2` `speculative_num_draft_tokens=4` | Achieve inference acceleration during the Rollout phase based on MTP | # 3. Experimental Results The experiment was conducted as follows: * model = mimo-7B-math * max_response_length = 8k Experiment chart: ![fully_async_policy_revenue]( https://github.com/ArronHZG/verl-community/blob/main/docs/mimo-7b-mtp.png?raw=true) **Scenarios with No Significant Effect** The following configurations will not have a noticeable impact on training results: 1. The base model does not carry MTP parameters; 2. The base model carries MTP parameters, but the MTP module is not trained; 3. The base model carries MTP parameters and trains MTP, with `mtp_loss_scaling_factor=0`; 4. The base model carries MTP parameters, trains MTP and detaches the encoder, with `mtp_loss_scaling_factor=0.1`. **Scenarios with Significant Effect** Only the following configuration will have a noticeable impact on training results: - The base model carries MTP parameters, MTP Loss applies to all model parameters, and `mtp_loss_scaling_factor=0.1`. **Recommended Training Method** It is recommended to adopt the `detach_encoder=True` approach for MTP training. # 4. Performance Notes for MTP in Rollout Inference The effectiveness of MTP-accelerated Rollout is significantly affected by **model size** and **inference hardware**. Key reference information is as follows: **Hardware Tensor Core Performance** | Hardware Model | FP16 Performance (TFLOPS) | |----------------|---------------------------| | H20 | 148 | | H800 | 1,671 | | H200 | 1,979 | **Measured Performance and Recommendations** Taking the mimo-7B model deployed separately on H20 hardware using SGLang as an example: After enabling MTP speculative decoding, the Rollout throughput decreases by approximately 50%. - Current priority recommendation: Do not enable MTP acceleration during the inference phase for now; - Future planning: Further optimization of the speculative logic in the Rollout phase will be conducted to improve throughput performance. > 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 - [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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [x] 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`.

### 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. The SFT training with MTP is supported, using the same MTP training configuration as RL training. An example configuration for running SFT can be found in `examples/sft/gsm8k/run_mimo_megatron_mtp.sh` **SFT result** The experiment was conducted using following data: - model = mimo-7B-math - dataset = gsm8k The result: [wandb link](https://wandb.ai/hou-zg-meituan/mimo-7b-sft-mtp?nw=nwuserhouzg) The presence of mtp layer has limited effect on main loss. However, when MTP layer is detached, the mtp_loss converges to a higher value. ### 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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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. ### Design & Code Changes Support SFT training with MTP, with the core change based on verl-project#4936 ### 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: Using none-standard mbridge/mcore - [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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [X] 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`. --------- Co-authored-by: ArronHZG <hou.zg@foxmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

### 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. The SFT training with MTP is supported, using the same MTP training configuration as RL training. An example configuration for running SFT can be found in `examples/sft/gsm8k/run_mimo_megatron_mtp.sh` **SFT result** The experiment was conducted using following data: - model = mimo-7B-math - dataset = gsm8k The result: [wandb link](https://wandb.ai/hou-zg-meituan/mimo-7b-sft-mtp?nw=nwuserhouzg) The presence of mtp layer has limited effect on main loss. However, when MTP layer is detached, the mtp_loss converges to a higher value. ### 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`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward` - 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. ### Design & Code Changes Support SFT training with MTP, with the core change based on verl-project/verl#4936 ### 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: Using none-standard mbridge/mcore - [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). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [X] 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`. --------- Co-authored-by: ArronHZG <hou.zg@foxmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

ArronHZG added 25 commits December 16, 2025 15:25

init mtp config

ae09ece

enable_mtp

a1825c2

add MIMO model

e0c71b7

add mimo model

420ea22

mtp_patch

1409af9

mimo_7b shell

c0c4f45

mtp worker

a3ee092

mtp rollout

793c3ba

mtp pp2

1cfe236

mtp pp2 worker

25a50c0

mtp rollout and train work

95e06c5

mtp loss

6524ce0

mtp loss

9df9904

detach mtp

a52aa28

detach mtp

3043c73

enable_memory_saver = True

86d73db

for sglang 0.5.6

58b2316

merge main

5e0aa9f

mtp train

d7f7b0b

mtp train

02500d0

mtp detach_encoder

edc447d

rm some log

6c75dbd

merge main

9a059a7

change ppo yaml

137c196

readme

7840b93

ArronHZG requested review from ISEEKYAN, SwordFaith, chenhaiq, vermouth1992 and zhaochenyang20 as code owners January 15, 2026 09:15

ArronHZG requested review from PeterSH6, eric-haibin-lin, tongyx361 and wuxibin89 as code owners January 15, 2026 09:15

ArronHZG added 6 commits January 16, 2026 11:46

refactor

faedcb5

fix ci

9b6ea4a

rm log

59e91e0

Merge branch 'main' into mtp

ff1348f

Merge branch 'main' into mtp

5d3a0c8

from megatron.core.utils import unwrap_model

729b818

wuxibin89 approved these changes Jan 19, 2026

View reviewed changes

wuxibin89 mentioned this pull request Jan 19, 2026

[roadmap] verl 26Q1 roadmap #4880

Closed

16 of 28 tasks

wuxibin89 merged commit df340d7 into verl-project:main Jan 19, 2026
68 of 71 checks passed

arvyanh mentioned this pull request Jan 21, 2026

[megatron] feat: Support MTP training in SFT #4981

Merged

8 tasks done

Navigation Menu

Repository navigation

[megatron] feat: Using MTP in RL Training and Inference - #4936

[megatron] feat: Using MTP in RL Training and Inference#4936
wuxibin89 merged 31 commits into
verl-project:mainfrom
meituan-search:mtp

ArronHZG commented Jan 15, 2026

Uh oh!

Uh oh!

ISEEKYAN commented Jan 19, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

ArronHZG commented Jan 15, 2026

Guide to Using MTP in RL Training and Inference

1. Scope of Support

2. MTP Training Configuration (Core Parameters)

3. Experimental Results

4. Performance Notes for MTP in Rollout Inference

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

Uh oh!

ISEEKYAN commented Jan 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ISEEKYAN commented Jan 19, 2026 •

edited

Loading