Issues

• GPT-OSS post-training enables YaRN and a learned position table together

  community-request
  NVIDIA/Megatron-LM#7900 · Filter by author Aravind-11 opened on Oct 6, 2026

• [BUG] EP=1 disaggregated inference can crash on dummy recurrent-state allocation

  NVIDIA/Megatron-LM#7898 · Filter by author jthomson04 opened on Oct 5, 2026

• [RL] Make the advantage estimator and loss aggregation configurable (RLOO, Dr. GRPO, GSPO)

  community-request
  NVIDIA/Megatron-LM#7893 · Filter by author Aravind-11 opened on Oct 5, 2026

• [Design proposal]: shared residual-connection integration for mHC and Wide Residual

  enhancement
  Enhancement

  NVIDIA/Megatron-LM#7888 · Filter by author kwangjunahn opened on Oct 5, 2026
  1

  

• Megatron-LM Monthly Activity Report — October 2026 (through October 4, 2026)

  NVIDIA/Megatron-LM#7882 · Filter by author lhb8125 opened on Oct 5, 2026

• [BUG] Low-precision temperature scaling changes TorchSampling output at temperature=0.1 and 0

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7857 · Filter by author Waterpine opened on Oct 4, 2026

• [BUG] BF16 Mamba/GDP multi-token decode and rollback states differ from serial decode

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7856 · Filter by author Waterpine opened on Oct 4, 2026
  

• [BUG] Fused softmax path ignores window_size for sliding-window attention

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7847 · Filter by author DhineshPonnarasan opened on Oct 3, 2026

• [BUG] GDP prefix caching produces different tokens with CUDA graphs enabled

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7845 · Filter by author Waterpine opened on Oct 3, 2026
  

• fix(inference): handle SSM-only handoff when all KV blocks are cached

  NVIDIA/Megatron-LM#7837 · Filter by author jthomson04 opened on Oct 2, 2026

• [BUG] One-token disaggregated handoff stalls subsequent requests with EP enabled

  bug
  NVIDIA/Megatron-LM#7835 · Filter by author jthomson04 opened on Oct 2, 2026

• [Tracking] Bitwise train/generation parity for RL: batch-invariant inference + the training forward

  community-requestenhancementwaiting-on-maintainers
  NVIDIA/Megatron-LM#7831 · Filter by author YigongQin opened on Oct 2, 2026

• [BUG] Block CUDA graphs can replay stale pipeline receive buffers in hybrid inference

  NVIDIA/Megatron-LM#7829 · Filter by author jthomson04 opened on Oct 2, 2026

• Support request-local sampling seeds in dynamic inference

  NVIDIA/Megatron-LM#7827 · Filter by author wdykas opened on Oct 2, 2026
  1

• MFSDP v2: frozen weights resharded before backward when inputs require no gradients

  bugMFSDPv2
  NVIDIA/Megatron-LM#7823 · Filter by author wujingyue opened on Oct 2, 2026

• [dev] Megatron Lite DSA: memory-bounded, exact indexer top-k with an optional LiteTopK backend for long-context GLM-5 prefill

  community-requestenhancementwaiting-on-maintainers
  NVIDIA/Megatron-LM#7818 · Filter by author Heisenberg-Yin opened on Oct 2, 2026

• Enable FP32 indexer projection outputs with cuDNN DSA

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7816 · Filter by author zianglih opened on Oct 2, 2026

• MFSDP v2: static prefetch order does not represent repeated module calls

  bug
  NVIDIA/Megatron-LM#7814 · Filter by author wujingyue opened on Oct 2, 2026

• Support configurable native cuDNN DSA top-k tie breaking

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7812 · Filter by author zianglih opened on Oct 2, 2026

• [BUG] A100 LTS GPT and MoE exact-reference mismatches across allocations

  bug
  NVIDIA/Megatron-LM#7801 · Filter by author balasaajay opened on Oct 1, 2026

• MFSDP v2: revisit gradient counting as the post-backward trigger

  enhancement
  NVIDIA/Megatron-LM#7791 · Filter by author wujingyue opened on Oct 1, 2026

• [QUESTION] Will this repo accept contributions for other vendor devices

  community-requestquestion
  NVIDIA/Megatron-LM#7784 · Filter by author polisettyvarma opened on Oct 1, 2026
  4

  

• MFSDP v2: cross-unit checkpoint replay reads sharded child weights

  bug
  Bug

  NVIDIA/Megatron-LM#7777 · Filter by author wujingyue opened on Oct 1, 2026

• [BUG] Expert RNG seed depends on pp_rank + ep_rank only, so experts in different pipeline stages get identical initial weights

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#7776 · Filter by author SeverinVisionary opened on Oct 1, 2026

• Megatron-LM Monthly Activity Report — September 2026

  NVIDIA/Megatron-LM#7773 · Filter by author lhb8125 opened on Oct 1, 2026