Skip to content

Pull requests: ISEEKYAN/Megatron-LM

New pull request New
57 Open 121 Closed
57 Open 121 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

refactor(lite): progressively reduce DS4 vLLM alignment path
#187 opened Aug 21, 2026 by ISEEKYAN Owner • Draft
[Lite] Remove full-size FP32 temporaries from the FSDP2 optimizer step
#186 opened Aug 21, 2026 by ISEEKYAN Owner Loading…
1
test(lite): gate DS4 vLLM forward and backward parity
#185 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
fix(lite): complete DS4 RL training integration
#184 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
feat(lite): add deterministic DeepEP and FP8 primitives
#183 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
lite: replace use_deepep with moe_token_dispatcher_type and add a HybridEP backend
#182 opened Aug 19, 2026 by ISEEKYAN Owner Loading…
feat(lite): add DS4 vLLM-aligned training model
#181 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
fix(verl): consume lazy weight exports on actor thread
#180 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
1
feat(lite): add DS4 vLLM-aligned training path
#179 opened Aug 19, 2026 by ISEEKYAN Owner • Draft
lite: mark test_hf_weights_gpu with its real GPU requirements
#178 opened Aug 19, 2026 by ISEEKYAN Owner Loading…
Restore FSDP2 master alias offload fix on PR157
#171 opened Aug 13, 2026 by conver334 Loading…
feat(mlite): add training and rollout offload for standalone M-FSDP
#169 opened Aug 12, 2026 by ISEEKYAN Owner • Draft
feat(mlite): add standalone Megatron FSDP optimizer
#166 opened Aug 10, 2026 by ISEEKYAN Owner • Draft
feat(lite): add stacked multi-LoRA MoE support
#161 opened Aug 8, 2026 by ISEEKYAN Owner • Draft
test(lite): add Megatron-Bridge LoRA parity coverage
#160 opened Aug 8, 2026 by ISEEKYAN Owner • Draft
Share FP32 FSDP2 optimizer master storage
#157 opened Aug 6, 2026 by ISEEKYAN Owner Loading…
Review-only: M-FSDP checkpoint shard refresh (already on the #89 branch)
#156 opened Aug 6, 2026 by ISEEKYAN Owner Loading…
Allow dist-opt gradient reduction in parameter dtype
#152 opened Aug 5, 2026 by ISEEKYAN Owner Loading…
Fix MXFP4 actor export for K3 resync
#151 opened Aug 5, 2026 by ISEEKYAN Owner Loading…
Fix full M-FSDP optimizer offload with FP32 gradients
#148 opened Aug 4, 2026 by ISEEKYAN Owner • Draft
Align Lite isort and Black formatting
#147 opened Aug 4, 2026 by ISEEKYAN Owner Loading…
[lite] feat: declarative fused weight layout primitive
#142 opened Aug 1, 2026 by ISEEKYAN Owner Loading…
Harden canonical and physical HF checkpoint coverage
#138 opened Jul 30, 2026 by ISEEKYAN Owner Loading…
Propagate torchrun thread default to Ray workers
#137 opened Jul 28, 2026 by ISEEKYAN Owner Loading…
Add token-wise ChunkedEP overlap for MLite MoE models
#134 opened Jul 27, 2026 by ISEEKYAN Owner Loading…
3
ProTip! Adding no:label will show everything without a label.