All pull requests

Pull requests

• fix(datasets): honor answer_only_loss_mask=False for generation templates

  community-request
  #4756·sohailshk opened Nov 13, 2026Contributor·2 of 3 tasks·

  ·
  1

• feat(inkling2): add native text and image training support

  #4755·snowmanwwg opened Nov 13, 2026Contributor·2 of 6 tasks·

  ·
  10

• fix(deepseek-v4): pool packed compressor windows per document at CP size 1

  community-request
  #4754·linmuchuiyang opened Nov 13, 2026Contributor·3 tasks done·

  ·
  1

  1

• feat(inkling2): add Inkling-2-Small training support

  community-request
  #4751·ischencheng opened Nov 13, 2026Contributor·9 tasks done·

  ·
  1

• fix(vlm): don't pad clips shorter than min_frames in the token estimate

  community-request
  #4750·dundysm opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

  3

• fix(flops): count every shared expert in deepseekv3_flops

  community-request
  #4748·dundysm opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

  3

• fix(kd): match teacher vocab shards to student DTensor ranges

  community-request
  #4746·ZenAlexa opened Nov 12, 2026Contributor·

  ·
  3

• fix(config): make dotted in checks agree with get()

  community-request
  #4745·dundysm opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

  3

• fix(ernie4_5): detect the HF model prefix from expert keys

  community-request
  #4741·dundysm opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

  3

• fix(checkpoint): materialize TE FusedAdam state for native optimizer loads

  #4739·HuiyingLi opened Nov 12, 2026Contributor·3 tasks done·

  ·
  2

• ci: proxy CI run for PR #4722

  #4738·HuiyingLi opened Nov 12, 2026Contributor·

  ·
  2

• refactor(peft): type LoRA expert adapter state and initialization contracts

  community-request
  #4737·taran-dev4u opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

• fix(shared): map complex64 dtype strings to torch.complex64

  community-request
  #4736·sohailshk opened Nov 12, 2026Contributor·3 tasks done·

  ·
  2

• fix(datasets): clamp limit_dataset_samples in column-mapped instruction data

  community-request
  #4735·sohailshk opened Nov 12, 2026Contributor·3 tasks done·

  ·
  3

• feat(te): NVFP4 training through Transformer Engine

  #4734·HuiyingLi opened Nov 12, 2026Contributor·2 of 3 tasks·

  ·
  1

• perf(checkpoint): read consolidation shards through open handles

  community-request
  #4733·dogeONE-bit opened Nov 12, 2026·3 tasks done·

  ·
  7

• fix(config): handle excess layer types in NeMoAutoConfig

  #4732·janbernloehr opened Nov 11, 2026Contributor·3 tasks done·

  ·
  2

• feat(speculative): warm-start DSpark drafts from existing weights

  community-request
  #4725·ischencheng opened Nov 11, 2026Contributor·3 tasks done·

  ·
  21

• fix(peft): let a leading "*." wildcard match top-level modules

  community-request
  #4723·sohailshk opened Nov 11, 2026Contributor·2 of 3 tasks·

  ·
  3

• fix(datasets): hash named chat templates in the DSpark cache manifest

  community-request
  #4722·sohailshk opened Nov 11, 2026Contributor·2 of 3 tasks·

  ·
  8

• feat(decision): add Deciffusion decision post-training recipe

  community-request
  #4721·zlapp opened Nov 11, 2026Contributor·3 tasks done·

  ·
  2

• fix(transformers): load pinned snapshots that miss only unread listed files

  #4720·jQizhang opened Nov 11, 2026Contributor·3 tasks done·

  ·
  1

• ci: proxy CI run for PR #4685

  #4719·HuiyingLi opened Nov 11, 2026Contributor·

  ·
  2

• chore: CI proxy for #4663

  #4718·akoumpa opened Nov 11, 2026Contributor·

  ·
  3

• fix(validation): support empty ranks and partial pipeline batches

  #4717·akoumpa opened Nov 11, 2026Contributor·3 tasks done·

  ·
  3