All pull requests

Pull requests

• perf(megatron): join MoE router and expert input gradients

  #1372·HuangWei-95 opened Nov 3, 2026Collaborator·3 tasks·

  ·

• perf(megatron): align persistent DEOSC state views with logical quantization tiles

  #1371·HuangWei-95 opened Nov 3, 2026Collaborator·3 tasks·

  ·

• fix(megatron): finish the core_v0.19.2 port (PP schedules, V schedule, FSDP, config args)

  ci:gpu
  #1370·WangLingxun opened Nov 3, 2026Collaborator·

  ·
  6

• feat(inkling_mm_model): add thinkingmachines/Inkling

  #1369·peymanr opened Nov 3, 2026Contributor·

  ·
  5

• Zyphra/moe expert placement

  #1367·RadhaGulhane13 opened Nov 3, 2026·

  ·

• Port MoE token dedup (ROCm/Megatron-LM PR #165) to the Megatron backend

  #1366·RadhaGulhane13 opened Nov 3, 2026·

  ·

• feat(minimax-m3): MiniMax-M3 proxy pretraining on MI455X (gfx1250)

  ci:gpu
  #1364·HuangWei-95 opened Nov 2, 2026Collaborator·6 of 7 tasks·

  ·
  5

• feat(kimi_k3): Kimi K3.1 Pro preset, 4-layer proxy and MI455X BF16 recipe

  #1363·boyang-amd opened Nov 2, 2026Collaborator·3 of 4 tasks·

  ·

• fix(megatron): reduce deferred output-layer wgrads with their grad bucket

  ci:gpu
  #1362·WangLingxun opened Nov 2, 2026Collaborator·

  ·
  1

  9

• fix(megatron): mix the resume iteration into the Flux per-DP-rank seed

  #1361·WangLingxun opened Nov 2, 2026Collaborator·

  ·

• chore(deps): bump actions/setup-python from 6.3.0 to 7.1.0 in the github-actions group

  dependenciesgithub-actions
  #1356·dependabot[bot] opened Nov 1, 2026Bot·

  ·

• perf(megatron): tune MI325X GPT-OSS-20B and ZAYA1-8B pretrain configs

  ci:gpu
  #1354·amd-fuyuajin opened Oct 31, 2026Collaborator·

  ·
  5

• feat(olmo3): add allenai/Olmo-3-1025-7B

  #1353·peymanr opened Oct 31, 2026Contributor·

  ·
  1

• feat(smollm3): add HuggingFaceTB/SmolLM3-3B-Base

  #1352·peymanr opened Oct 31, 2026Contributor·

  ·
  2

• fix(sft): align the unpacked SFT loss mask with the shifted labels

  #1351·peymanr opened Oct 30, 2026Contributor·

  ·
  12

• perf(automodel): compile-friendly FP8 linears and deferred train-loop host syncs

  ci:gpu
  #1350·luiza-amd opened Oct 30, 2026Contributor·

  ·
  20

  

• perf(configs): tune MI355X FLUX.1-dev (JAX) and Mixtral 8x7B / DeepSeek-V2-Lite BF16 (Megatron); add FP8 MoE models to the perf catalog

  ci:gpu
  #1348·amd-fuyuajin opened Oct 30, 2026Collaborator·

  ·
  3

• feat(granite): add ibm-granite/granite-4.1-3b

  #1347·peymanr opened Oct 30, 2026Contributor·

  ·
  3

• Make training and the projection benchmark both take the fused FP8 grouped-MLP path plus remove calibration constants

  #1346·araina-amd opened Oct 30, 2026Collaborator·

  ·
  13

• feat(diffusion): Add AMDB Mochi-1 benchmark support

  ci:gpu
  #1345·kailashg26 opened Oct 29, 2026Contributor·1 task done·

  ·
  38

• feat(diffusion): Add AMDB Wan2.1 image-to-video benchmark support

  ci:gpu
  #1344·kailashg26 opened Oct 29, 2026Contributor·1 task done·

  ·
  33

• feat(diffusion): Add AMDB HunyuanVideo benchmark support

  ci:gpu
  #1343·kailashg26 opened Oct 29, 2026Contributor·1 task done·

  ·
  32

• feat(diffusion): Add AMDB SDXL benchmark support

  ci:gpu
  #1342·kailashg26 opened Oct 29, 2026Contributor·1 task done·

  ·
  46

• feat(diffusion): Add AMDB FLUX.1-dev benchmark support

  ci:gpu
  #1341·kailashg26 opened Oct 29, 2026Contributor·1 task done·

  ·
  36

• [draft] feat(megatron): add MiniMax-M3.1-Flash and fix M3's FlyDSL MSA import on rocm/primus:v26.8

  #1339·HuangWei-95 opened Oct 29, 2026Collaborator·3 of 5 tasks·

  ·