Repository Issue Activity (beta)

#6501 [feat] GTP+DCP

#1784 [ENHANCEMENT] MoE support in report_theoretical_memory

#6747 [VPP] Defer embedding initialization sync until all local model chunks are built

#6757 [ROADMAP][2026 Q3] Megatron Core MoE Roadmap

#6729 triton KV append kernel hard-asserts CUDA, blocking non-CUDA Triton backends

#6708 GRPO crashes with --transformer-impl local when RL training CUDA graphs are disabled

#6540 [feat] GTP+Symmetric Memory Registration buffer

#6700 [RL][Performance] Avoid full-vocabulary TP gather for selected-token logprobs

#1455 [QUESTION]Is there any plan to make custom_fsdp compatible with PP?

#6600 MFSDP v2: Support EP composability with grouped 3D expert weights

#6585 [BUG] pretrain_vlm.py silently ignores --rotary-base: LLaVAModel always uses the default 10000

#6087 [MFSDP V2] Post-wrap load_state_dict(assign=True) does not resume from the loaded checkpoint

#4815 [ROADMAP][2026 Q2] Megatron Core MoE Roadmap

#6656 [BUG] FP32 backward fails in VocabParallelCrossEntropy with custom-autograd view/in-place error

#6363 [feat] GTP+2D Mxfp8

#6660 [Megatron-FSDP] Gradients doubled every microbatch (2^N amplification) when a grad buffer's DP group has size 1 (EP=DP or DP=1) with optim_grads_params

#6491 Offload MoE expert weights to pinned host memory to enable training at smaller EP

#6720 [GTP][Muon] Optimize Muon step performance

#6392 GLM-5.2 training support tracking

#6714 MFSDP v2: overlap DP-inner and DP-outer communication

#6111 MoE router padding mask fails with expert bias during training

#5755 [BUG] TransformerConfig accepts num_attention_heads not divisible by num_query_groups, then crashes with a cryptic shape error inside attention

#3083 Profiling args are missing on the main brench

#6719 Pipeline the layer-sharded Muon all_to_all to hide exchange latency (and cut peak memory)

#5787 [BUG] `ParamAndGradBuffer.update_main_grads` repeats the global-shape view for cached main gradients (rowwise-TP silent-wrong / deadlock)