Skip to content

Pull requests: ISEEKYAN/Megatron-LM

New pull request New
36 Open 93 Closed
36 Open 93 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

mlite: fix cross_entropy_fusion never taking effect (all four models)
#133 opened Jul 27, 2026 by ISEEKYAN Owner Loading…
Bound vocab-parallel cross entropy memory by token chunks
#132 opened Jul 25, 2026 by ISEEKYAN Owner Loading…
[lite] feat: plug-and-play LoRA primitive (apply_lora_to_chunks)
#131 opened Jul 25, 2026 by ISEEKYAN Owner Loading…
[lite] feat: Muon optimizer for MLite — DistOpt (ZeRO-1) lowering + FSDP2 implementation
#130 opened Jul 25, 2026 by ISEEKYAN Owner Loading…
Add M-FSDP CPU optimizer offload (offload_fraction support)
#129 opened Jul 24, 2026 by conver334 • Draft
6 tasks done
[qat] feat: weight-only QAT primitive (int8/int4/fp8_e4m3/mxfp4) — review boundary
#127 opened Jul 23, 2026 by ISEEKYAN Owner Loading…
Fix M-FSDP "active buffers" error on device migration / scratch release
#126 opened Jul 22, 2026 by conver334 Loading…
Fix M-FSDP DCP checkpoint parameter scope
#125 opened Jul 22, 2026 by conver334 Loading…
1
[experimental] Add Agent Compose package and skills skeleton
#124 opened Jul 21, 2026 by ISEEKYAN Owner • Draft
verl/ds4: env-overridable MoE backend & offload fraction — Blackwell (SM100) validation of the resync-format KL matrix
#123 opened Jul 20, 2026 by Meirtz Loading…
Typed model composition kernel (compose.assemble) + explicit specialization points (#118)
#120 opened Jul 19, 2026 by ISEEKYAN Owner Loading…
Cross-cutting contract conformance + fail-loud memory-feature guards (#114)
#119 opened Jul 19, 2026 by ISEEKYAN Owner Loading…
verl example: move training hyperparameters out of launchers into config
#103 opened Jul 13, 2026 by ISEEKYAN Owner Loading…
Add Muon comparison study: verl + Megatron vs. Megatron-Lite
#99 opened Jul 13, 2026 by ISEEKYAN Owner Loading…
Document CUDA Graph design for THD-only MLite training
#98 opened Jul 10, 2026 by ISEEKYAN Owner Loading…
Add Hopper blockwise FP8 precision profiles
#96 opened Jul 10, 2026 by ISEEKYAN Owner Loading…
Document Muon post-training configuration
#95 opened Jul 10, 2026 by ISEEKYAN Owner Loading…
Analyze incremental RL weight synchronization
#93 opened Jul 9, 2026 by ISEEKYAN Owner Loading…
Document MLite FP8 training architecture
#92 opened Jul 9, 2026 by ISEEKYAN Owner Loading…
MLite Muon optimizer step 1: native config + parameter-routing contract (WIP)
#90 opened Jul 9, 2026 by ISEEKYAN Owner Loading…
MLite M-FSDP: fully standalone optimizer (follow-up to #26; no vendoring / mcore / fsdp2 deps)
#89 opened Jul 9, 2026 by ISEEKYAN Owner Loading…
Document Muon optimizer integration design
#88 opened Jul 9, 2026 by ISEEKYAN Owner Loading…
[lite] experiments: memory-capacity harness (capacity law + module ablation, first results)
#83 opened Jul 6, 2026 by Meirtz Loading…
[lite] hardening: fsdp2 rank0-only export deadlock + packed loss-mask misalignment guard
#82 opened Jul 6, 2026 by Meirtz Loading…
[lite] R3 router replay: record rollout routing, replay in training (multi-microbatch, recompute-safe, vLLM route ingest)
#79 opened Jul 5, 2026 by Meirtz Loading…
ProTip! Find all pull requests that aren't related to any open issues with -linked:issue.