All pull requests

Pull requests

• fix(training): use model config for kernel warmup

  complexity: mediumnemotron
  #7653·JF-D opened Sep 24, 2026Contributor·

  ·
  1

  

• Fix 32-bit input offsets in batch-invariant weighted SwiGLU

  community-request
  #7651·arcusbuilds opened Sep 24, 2026·

  ·
  1

  1

• Preserve unsharded cyclic samples across DP changes

  community-request
  #7649·Silas-Zeng opened Sep 24, 2026·

  ·
  1

  1

• Reject unsafe sharded cyclic dataloader resumes

  community-request
  #7648·Silas-Zeng opened Sep 24, 2026·

  ·
  1

  1

• dist_ckpt: make torch_dist save planning linear in the expert-weight chunk count

  #7644·asolergi-nv opened Sep 24, 2026Contributor·

  ·
  1

• Fix DSv4 fused RoPE autograd aliasing

  community-request
  #7643·Silas-Zeng opened Sep 24, 2026·

  ·
  1

  2

• [DSv4.1] Add cuDNN saved-state gate to Engram training

  #7642·YangXu1990uiuc opened Sep 24, 2026·

  ·
  2

• Add VPP support to typed pipeline boundaries

  community-request
  #7640·Silas-Zeng opened Sep 24, 2026·

  ·
  1

• Fix multimodal dataloader restoration during full-dataset resume

  #7639·tylerpoon opened Sep 24, 2026Contributor·

  ·
  1

• Test MFSDP v2 attention activation offload

  #7638·wujingyue opened Sep 24, 2026Contributor·

  ·
  1

• Clean up MFSDP parameter-group annotations

  complexity: low
  #7636·wujingyue opened Sep 24, 2026Contributor·

  ·
  1

• perf(fsdp): combine multi-axis all-gathers into one collective

  #7634·wujingyue opened Sep 24, 2026Contributor·

  ·
  1

• fix(fsdp): support independent hybrid ZeRO-2 and expert sharding

  #7633·wujingyue opened Sep 23, 2026Contributor·

  ·
  1

• ci: Add source-to-bucket overrides for selective unit tests

  #7632·balasaajay opened Sep 23, 2026Contributor·

  ·
  1

• [Dev] Stage async checkpoint tensors through pageable memory on ROCm

  community-request
  #7631·HJSang opened Sep 23, 2026·

  ·
  2

• [Dev] Import override from typing_extensions in training model builders

  community-request
  #7630·HJSang opened Sep 23, 2026·

  ·
  3

• [Dev] Preserve HDO step across DP-reshardable resume

  community-request
  #7629·HJSang opened Sep 23, 2026·

  ·
  4

• Bump nvidia-modelopt minimum to 0.47

  #7628·chadvoegele opened Sep 23, 2026·

  ·
  1

• Fix GDN model builder for MTP in GPT model

  Approvedcomplexity: low
  #7626·zhongbozhu opened Sep 23, 2026Contributor·

  ·
  2

• [DEV] Add opt-in aligned HCA backward

  #7625·layalir opened Sep 23, 2026Contributor·

  ·
  1

• ci: persist native dependency wheel caches across container builds

  #7624·balasaajay opened Sep 23, 2026Contributor·

  ·
  8

• fix(mtp): compute acceptance before fused CE overwrites logits

  community-request
  #7623·dundysm opened Sep 23, 2026·

  ·
  1

  3

• [Dev] lite: Add MiniMax-M3 support (MSA primitive, MagiAttention MSA backend, HybridEP dispatcher)

  #7621·conver334 opened Sep 23, 2026Contributor·

  ·
  1

• fix(dist_checkpointing): pass only dtype to torch.empty in load_tensors_metadata

  #7620·michal2409 opened Sep 23, 2026Contributor·

  ·
  1

• Speed up single-round EP MoE return with overlapped output initialization

  community-request
  #7619·0z5a opened Sep 23, 2026·

  ·
  1