All pull requests

Pull requests

• feat(data): intra-microbatch reordering for MegatronMIMO scalable reads

  community-request
  #6306·going-song opened Oct 3, 2026Contributor·

  ·
  6

• fix(model): reject get_model_config() for families without a config mapping

  area:modelbugcommunity-requestneeds-review
  #6305·jayzuccarelli opened Oct 3, 2026·

  ·
  1

  1

• fix(training): count attention output gates in hybrid FLOPs

  area:trainingbugneeds-review
  #6301·cuichenx opened Oct 2, 2026Contributor·3 of 5 tasks·

  ·
  1

• feat(recipe): add Qwen3.5 text MXFP8 long-context SFT

  area:recipeblockedfeature
  #6300·cuichenx opened Oct 2, 2026Contributor·

  ·
  2

• [build] chore: bump MCore main to feba4e1e

  area:buildcifull-test-suiteneeds-more-testsneeds-review
  #6299·cuichenx opened Oct 2, 2026Contributor·

  ·
  1

• fix(ci): prevent functional test filename shell injection

  area:buildci
  #6296·balasaajay opened Oct 2, 2026Contributor·4 tasks done·

  ·
  2

• docs(model-card): record distributed inference candidates

  area:modelbug
  #6294·cuichenx opened Oct 2, 2026Contributor·

  ·
  1

• docs(models): correct Super and GPT-OSS verification provenance

  area:modelbug
  #6293·cuichenx opened Oct 2, 2026Contributor·

  ·
  1

• feat: support pipeline-stage-local HF export

  area:ckptcommunity-requestfeatureneeds-more-testsneeds-review
  #6291·Bsgg1 opened Oct 2, 2026·

  ·
  1

  7

• build: install Mixture of Kittens in Docker on amd64 and arm64

  area:buildci
  #6290·balasaajay opened Oct 2, 2026Contributor·4 tasks done·

  ·
  2

• feat: add DiffusionGemma support

  area:modelcommunity-requestfeatureneeds-more-testsneeds-review
  #6284·BenDGeist opened Oct 1, 2026·3 of 5 tasks·

  ·
  3

• no non det algo

  area:recipebug
  #6282·malay-nagda opened Oct 1, 2026Contributor·5 tasks·

  ·
  2

• feat(recipe): align DSv4 Pro GB300 benchmark and add proxy

  area:recipefeaturehigh-complexityneeds-more-tests
  #6279·malay-nagda opened Oct 1, 2026Contributor·2 of 4 tasks·

  ·
  1

• feat(model): add Shensi model support

  area:modelcommunity-requestfeatureneeds-review
  #6278·zongzhilou opened Oct 1, 2026·3 of 5 tasks·

  ·
  3

• feat(training): dynamic context parallel on global-batch packing with a Qwen3.5-35B-A3B recipe

  area:trainingfeature
  #6275·ilml opened Oct 1, 2026Contributor·5 of 6 tasks·

  ·
  1

• fix(peft): make shared-outer grouped-expert LoRA checkpointable

  area:peftbugcommunity-requestneeds-review
  #6274·JackRao123 opened Oct 1, 2026·

  ·
  1

• fix(training): reuse Megatron-Core determinism setup under --deterministic

  area:trainingbreaking-changebugneeds-review
  #6271·ZhiyuLi-Nvidia opened Sep 30, 2026Contributor·2 of 5 tasks·

  ·
  1

• feat(diffusion): support response-relative asymmetric attention

  area:diffusionfeature
  #6270·sajadn opened Sep 30, 2026Contributor·

  ·
  1

• docs: fix compare script path in README

  area:misccommunity-requestdocsdocs-onlyneeds-review
  #6269·pratikgx opened Sep 30, 2026·2 of 4 tasks·

  ·
  1

• reuse train state from MLM

  area:trainingbreaking-changefeature
  #6263·dimapihtar opened Sep 30, 2026Contributor·5 tasks·

  ·
  1

• build: update PyTorch base image to 26.09

  area:buildbuildcineeds-more-testsneeds-review
  #6258·balasaajay opened Sep 30, 2026Contributor·

  ·
  4

• fix(model): store Muse Glimmer final norm uncentered for exact bf16 round trip

  area:modelbreaking-changebug
  #6256·kamran-nvidia opened Sep 30, 2026Contributor·4 of 8 tasks·

  ·
  4

• fix(training): align native DeepSeek V4 packed CP layouts

  area:trainingbugneeds-review
  #6249·Connor-XY opened Sep 29, 2026·

  ·
  7

• fix(recipe): resync native DeepSeek V4 pipeline layout after overrides

  area:recipebugneeds-review
  #6248·Connor-XY opened Sep 29, 2026·

  ·
  5

• fix(training): preserve restored RNG during data iterator setup

  area:trainingbugneeds-review
  #6247·Connor-XY opened Sep 29, 2026·

  ·
  5