All pull requests

Pull requests

• [WIP] feat(megatron): opt-in pinned-Triton fusions for DeepSeek-V3 on MI455X

  #1259·lhzhang333 opened Oct 15, 2026Collaborator·

  ·

• feat(inkling): add Inkling (Small and Large) text backbone to the Megatron backend

  #1256·vidushi8 opened Oct 14, 2026Collaborator·

  ·
  5

• docs(specforge): publish SpecForge guides on the ROCm docs site

  ci:gpu
  #1255·yudigege86 opened Oct 14, 2026Collaborator·3 tasks·

  ·
  32

• fix(glm5_next): accept text-only HF checkpoint names in the GLM-5.3 loader

  #1254·yuankaichen-amd opened Oct 14, 2026Contributor·

  ·

• Notes/qwen3 30b a3b fp8 turbo campaign

  #1246·Xiaoming-AMD opened Oct 10, 2026Collaborator·

  ·
  20

• feat(torchtitan): add Triton fused AdamW with a device-sized grid, on by default

  ci:gpu
  #1233·WangLingxun opened Oct 8, 2026Collaborator·

  ·

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

  ci:gpu
  #1216·vidushi8 opened Oct 2, 2026Collaborator·

  ·
  18

• [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

• [maxtext] gemma4

  #1192·yeandy opened Sep 25, 2026Contributor·

  ·
  5

• feat(megatron): add Triton fused AdamW on the shared kernels, on by default

  ci:gpu
  #1190·WangLingxun opened Sep 24, 2026Collaborator·

  ·

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

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

  ·
  15

• 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·

  ·
  107

• 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

• feat(mxfp6): MXFP6 (E2M3) Flux linears, a fused MLP, and the MLPerf recipe

  #1064·jasainio opened Aug 31, 2026Contributor·5 of 7 tasks·

  ·

• fix(flux): make MLPerf validation measure what it claims to

  #1055·jasainio opened Aug 28, 2026Contributor·5 of 6 tasks·

  ·
  16

• [megatron] Implement gemma4

  #1053·yeandy opened Aug 27, 2026Contributor·

  ·
  3

• feat(megatron): add Qwen3-1.7B model preset and BF16 pretrain configs

  #1042·clairesonglee opened Aug 26, 2026Collaborator·

  ·

• [WIP]feat(turbo): route attention through Primus-Turbo's flex_attention when enabled

  #1033·botaohu001 opened Aug 25, 2026Contributor·

  ·
  1