[mcore] verl+megatron development tracking

[ISEEKYAN]

veRL Megatron-core Development Tracking

This page focuses on development of verl+mcore. The milestone target is to enable training deepseek-v3 on veRL as #708 and the further target is to continuously enhance the verl training experience of the mcore backend.

Progress and TODO

Recent

• [x] update mcore version to 0.11 #392
• [x] use mcore GPTModel api instead of huggingface workaround with sequence packing #706
• [x] support context parallel #970
• [x] support loading mcore dist_checkpointing #1030
• [x] support Megatron 0.11.0 and vLLM 0.8.2 #851
• [x] support qwen2moe training #1139
• [x] support Moonlight-16B-A3B training (WIP) #1284
• [ ] support Qwen2.5-VL training #1286
• [x] support EP(expert parallel) #1467

Further

• [ ] FP8 training
• [ ] training efficiency related optimization
• [ ] support sglang inference engine
• [ ] support trtllm inference engine

[eric-haibin-lin]

Could you merge the todo list from this as well? https://github.com/volcengine/verl/issues/825

[ISEEKYAN]

Could you merge the todo list from this as well? #825

ok

[rbao2018]

mark👀

[yushuiwx]

mark👀

[Aurelius84]

mark👀

[Wolfwjs]

How does the mcore backend support make_vocab_size_divisible_by, and where to pad the vocabulary to meet the splitting requirements?

[supermancmk]

Could you support qwen to use Megatron for SFT training?

[eric-haibin-lin]

Referencing a related issue: https://github.com/volcengine/verl/issues/708

[bhks]

The SGlang backend feature support sglang inference engine is already there in verl right ?