All pull requests

Pull requests

• chore(deps): bump Megatron minimum to 0.18

  #247·hjh0119 opened Oct 30, 2026Collaborator·

  ·

• fix(glm5_next): align Core mHC configuration and config copying

  #246·WangKangAandy opened Oct 29, 2026·

  ·

• fix(glm5_next): probe KDA fields on the hybrid layer config

  #245·WangKangAandy opened Oct 29, 2026·

  ·
  2

• feat(gemma4): optional ffpa-attn kernel for head_dim 512 full-attention layers

  #243·1tex opened Oct 28, 2026Contributor·

  ·

• fix(gdn): toggle sequence parallelism on the projection parameters too

  #242·li-lizhe opened Oct 25, 2026Contributor·

  ·
  1

  2

• fix: convert grouped-expert fc1 bias together with its FP8 scale

  #227·shiaho777 opened Oct 9, 2026Contributor·

  ·
  3

• fix: reject checkpoint filenames that leave the model directory

  #226·shiaho777 opened Oct 9, 2026Contributor·

  ·

• fix: parse layer patterns instead of evaluating them

  #225·shiaho777 opened Oct 9, 2026Contributor·

  ·

• fix: run MTP layers in the FP4 context when FP4 is enabled

  #224·shiaho777 opened Oct 9, 2026Contributor·

  ·

• fix: zero left-padded tokens before the GDN recurrence

  #223·shiaho777 opened Oct 9, 2026Contributor·

  ·

• fix: bound QSA keys by the key length under context parallelism

  #222·shiaho777 opened Oct 9, 2026Contributor·

  ·
  3

• fix: place collectives and new modules on the process accelerator

  #218·shiaho777 opened Oct 9, 2026Contributor·

  ·
  1

• fix: do not install the QSA sparse kernel off CUDA

  #217·shiaho777 opened Oct 9, 2026Contributor·

  ·
  3

• fix: pass DSA hidden states through selective attention recompute

  #216·shiaho777 opened Oct 9, 2026Contributor·

  ·

• Bound PLE backward workspace by chunking token reductions

  #213·yuchenwang3 opened Oct 3, 2026Contributor·

  ·
  1

  

• Preserve low-precision rounding in gated residual mixing

  #212·yuchenwang3 opened Oct 3, 2026Contributor·

  ·
  2

• feat(model): add generic chunked linear cross entropy

  #177·maomaocun opened Aug 31, 2026·

  ·

• Support local and unfused transformer paths and LoRA

  #159·hazelduan opened Aug 6, 2026Contributor·

  ·

• [bugfix] support MTP-only pipeline stages

  #155·taking-lying-flat opened Jul 31, 2026Contributor·

  ·
  5

• gpt_bridge support fsdp

  #154·Jintao-Huang opened Jul 21, 2026Collaborator·

  ·
  2

• support gemma4 use_hybrid_sliding

  #129·Jintao-Huang opened Jun 18, 2026Collaborator·

  ·
  6

• optimize cp gdn

  #128·Jintao-Huang opened Jun 17, 2026Collaborator·

  ·
  2

• [model] support minimax_m3

  #121·Jintao-Huang opened Jun 16, 2026Collaborator·

  ·
  2

• Fix assertion error when packed_seq_params is passed without proper qkv_format

  #109·tomzw11 opened Jun 4, 2026·

  ·
  3

• feat(kimi_k25): implement multimodal get_inputs_embeds for Megatron training

  #106·zhihanliu-collab opened Jun 3, 2026·

  ·
  4