All pull requests

Pull requests

• [OOB Release] add JAX v26.8 dockerfile

  #1226·hlira-amd opened Oct 6, 2026Collaborator·

  ·
  1

• feat(projection): DLRM/HSTU explicit backward GEMMs and arch-scaled attention epilogue

  #1224·araina-amd opened Oct 6, 2026Collaborator·2 of 3 tasks·

  ·

• perf(mi355x): tune pretrain configurations

  #1218·vidushi8 opened Oct 2, 2026Collaborator·

  ·
  8

• feat(glm5_next): add GLM-5.3-Flash (KDA + DSA hybrid with mHC) support

  #1217·yuankaichen-amd opened Oct 2, 2026Contributor·

  ·

• fix(torchtitan): bound DeepSeek compile at routed experts

  #1216·vidushi8 opened Oct 2, 2026Collaborator·

  ·
  11

• chore(deps): bump tornado from 6.5.8 to 6.5.9 in /docs/sphinx

  dependenciespython
  #1214·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• chore(deps): bump urllib3 from 2.7.0 to 2.8.0 in /docs/sphinx

  dependenciespython
  #1213·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• chore(deps): bump gitpython from 3.1.59 to 3.1.62 in /docs/sphinx

  dependenciespython
  #1212·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• chore(deps): bump pyjwt from 2.13.0 to 2.15.0 in /docs/sphinx

  dependenciespython
  #1211·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• chore(deps): bump actions/setup-python from 6.3.0 to 7.0.0 in the github-actions group

  dependenciesgithub-actions
  #1210·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• [Hybrid Models] Add a dense-to-hybrid checkpoint upcycler for non-MoE models.

  #1205·clairesonglee opened Sep 30, 2026Collaborator·

  ·

• fix(diffusion): keep Flux sensitive layers out of FP8 and MXFP4

  #1204·jianhan-amd opened Sep 30, 2026·

  ·
  4

• fix(diffusion): average the masked flow-matching loss over elements

  #1200·jianhan-amd opened Sep 29, 2026·

  ·
  4

• [Megatron-LM] Add ZAYA1-8B Megatron pretraining implementation

  ci:gpu
  #1199·clairesonglee opened Sep 28, 2026Collaborator·3 tasks·

  ·
  21

• [maxtext] gemma4

  #1192·yeandy opened Sep 25, 2026Contributor·

  ·
  5

• feat(megatron): add opt-in Triton fused AdamW with a device-sized grid

  #1190·WangLingxun opened Sep 24, 2026Collaborator·

  ·

• fix(megatron): QK-Norm on MLA decoupled-RoPE to stop hybrid divergence

  #1176·vanshbhatia-amd opened Sep 18, 2026Member·

  ·
  13

• fix(megatron): stop the tp_group deprecation warning from breaking compiled graphs

  #1169·jianhan-amd opened Sep 17, 2026·

  ·
  21

• Move the MaxDiffusion backend to transformers 5

  #1159·olehtika opened Sep 15, 2026Contributor·

  ·
  9

• fix(te): recover gfx950 hipBLASLt wgrad failures via workspace retry and torch fallback

  #1141·kailashg26 opened Sep 11, 2026Contributor·6 tasks done·

  ·
  3

• Feat/diffusion/wan21 megatron training

  #1132·jianhan-amd opened Sep 10, 2026·

  ·
  89

• refactor(examples): group per-backend material under its backend dire…

  #1130·Xiaoming-AMD opened Sep 10, 2026Collaborator·

  ·
  1

• test(mxfp6): cover BF16 sensitive-layer routing at the real 57-block shape

  #1115·gphuang opened Sep 8, 2026Contributor·2 of 4 tasks·

  ·
  2

• perf(mxfp6): fuse Flux QK norm with RoPE and the LN-modulate backward

  #1107·jasainio opened Sep 6, 2026Contributor·

  ·

• fix: exclude FP8 quantized weights from MXFP4 de-osc

  #1081·hann-wang opened Sep 3, 2026·

  ·
  8