All pull requests

Pull requests

• tune llama 70B model for mi300x; reduce mi355x llama 70B model train steps

  #1306·amd-fuyuajin opened Oct 25, 2026Collaborator·

  ·

• [Primer] Inkling best run: feat/inkling/add inkling bda8ce55

  #1305·clairesonglee opened Oct 24, 2026Collaborator·

  ·
  1

• chore(config): update MI300X Llama-3.1-70B FP8 batch size

  #1303·clairesonglee opened Oct 23, 2026Collaborator·4 of 5 tasks·

  ·

• fix(runner): stop MoE hangs on hosts with amdgpu DRM older than 3.64

  ci:gpu
  #1301·amd-fuyuajin opened Oct 23, 2026Collaborator·

  ·

• fix(projection): price DeepSeek-V4 attention as sparse MLA and use CDNA4 peaks

  #1300·araina-amd opened Oct 23, 2026Collaborator·

  ·

• fix(megatron_bridge): use gpt_step for text-only models in posttrain

  #1298·WangLingxun opened Oct 23, 2026Collaborator·

  ·
  1

• fix(megatron_bridge): prefer rope_parameters over default_theta when resolving rope_theta

  #1297·WangLingxun opened Oct 23, 2026Collaborator·

  ·
  1

• feat(examples): hipBLASLt GEMM tuning for GPT-OSS 20B and 120B on MI455X

  #1296·HuangWei-95 opened Oct 23, 2026Collaborator·

  ·

• feat(examples): MI455X DeepSeek-V3 and Llama-3.1-8B launchers use tuned hipBLASLt kernels

  ci:gpu
  #1295·lhzhang333 opened Oct 23, 2026Collaborator·4 tasks done·

  ·

• update the mi325x yaml config files for v26.9 image

  ci:gpu
  #1294·amd-fuyuajin opened Oct 23, 2026Collaborator·

  ·

• feat(tools/perf): report why a run failed, from the rank that failed; guard TRAIN_STEPS; keep downloads out of the checkout

  ci:gpu
  #1293·amd-fuyuajin opened Oct 23, 2026Collaborator·

  ·

• fix(megatron): store the Turbo FP8 weight cache in the op's padded layout

  #1292·vidushi8 opened Oct 22, 2026Collaborator·

  ·
  3

• fix(megatron-bridge): load SQuAD by its namespaced id in the Zebra SFT configs

  #1291·clairesonglee opened Oct 22, 2026Collaborator·

  ·

• fix(hybrid): read the FLA switches without requiring Megatron-LM args

  #1290·clairesonglee opened Oct 22, 2026Collaborator·

  ·

• maxtext: bump to release/v26.9, silence TE return_max_logit warnings, default autotune level 3

  ci:gpu
  #1289·amd-fuyuajin opened Oct 22, 2026Collaborator·6 tasks done·

  ·
  1

• chore(deps): bump certifi from 2026.7.22 to 2026.10.19 in /docs/sphinx

  dependenciespython
  #1282·dependabot[bot] opened Oct 21, 2026Bot·

  ·

• add zayz1 model into the benchmarking test list

  #1278·amd-fuyuajin opened Oct 20, 2026Collaborator·

  ·

• feat(zaya1): size MI355X 8B batches to GPU memory

  ci:gpu
  #1276·clairesonglee opened Oct 20, 2026Collaborator·3 tasks done·

  ·

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

  #1263·clairesonglee opened Oct 16, 2026Collaborator·

  ·
  1

• fix(hybrid): point the Zebra converter wrappers at a KDA checkpoint

  ci:gpu
  #1262·vanshbhatia-amd opened Oct 16, 2026Member·

  ·
  5

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

  #1256·vidushi8 opened Oct 14, 2026Collaborator·

  ·
  20

• Notes/qwen3 30b a3b fp8 turbo campaign

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

  ·
  20

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

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

  ·
  26

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

  #1205·clairesonglee opened Sep 30, 2026Collaborator·

  ·
  4

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

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

  ·
  4