All pull requests

Pull requests

• test(deps): AUT-2423 validate Megatron Core dev compatibility

  #6219·svcnemo-autobot opened Sep 27, 2026Contributor·

  ·
  1

• chore(deps): AUT-2423 bump Megatron Core compatibility pins

  #6218·svcnemo-autobot opened Sep 27, 2026Contributor·

  ·
  1

• feat(nemotron): add GTP configuration for Nemotron 3 Ultra on GB200 and GB300

  area:recipefeature
  #6216·ghadiaravi13 opened Sep 26, 2026Contributor·

  ·
  1

• feat(gemma4): enable dense context parallel attention

  area:modelcommunity-requestfeatureneeds-more-tests
  #6213·cj-55m opened Sep 26, 2026·

  ·
  4

• fix(perf): fix VR200 Qwen3.5-VL and Lightning recipes; add --segment override

  area:recipebug
  #6212·malay-nagda opened Sep 26, 2026Contributor·

  ·
  2

• fix(training): correct diagnostic rank selection and repeated logs

  area:trainingbugmlm-syncneeds-review
  #6211·cuichenx opened Sep 25, 2026Contributor·

  ·
  1

• fix(training): correct loss averages and hybrid FLOPs reporting

  area:trainingbugmlm-syncneeds-review
  #6210·cuichenx opened Sep 25, 2026Contributor·

  ·
  1

• fix(ckpt): sync training resume state and compatibility fixes

  area:ckptbugmlm-syncneeds-review
  #6209·cuichenx opened Sep 25, 2026Contributor·

  ·
  1

• fix(training): sync optimizer buffer and model setup correctness guards

  area:trainingbugmlm-syncneeds-review
  #6208·cuichenx opened Sep 25, 2026Contributor·

  ·
  1

• feat(recipe): add Nemotron 3 Ultra 256-GPU VR200 NVFP4 perf recipe

  area:recipefeature
  #6205·malay-nagda opened Sep 25, 2026Contributor·

  ·
  1

• fix(model): register both MTP sub-layer spellings in the dense Qwen3.5 bridge

  area:modelbugcommunity-requestneeds-review
  #6203·palanieppan-m opened Sep 24, 2026·

  ·
  1

• feat(model): add MiMo-V2.6-Flash text checkpoint conversion

  area:modelcommunity-requestfeature
  #6201·kalectory opened Sep 23, 2026·

  ·
  1

• fix(training): preserve independent MFSDP v2 sharding policies

  area:trainingbug
  #6200·wujingyue opened Sep 23, 2026Contributor·

  ·
  1

• feat(model): add DeepSeek-V4.1 Flash support

  area:modelcommunity-requestfeature
  #6199·HollowMan6 opened Sep 23, 2026Contributor·

  ·
  1

• fix(model): propagate language attention backend to Qwen3-VL and Exaone 4.5 vision configs

  area:modelbugcommunity-requestneeds-review
  #6198·erictang000 opened Sep 23, 2026Contributor·

  ·
  1

• feat(ckpt): export native Megatron-LM checkpoints without needing a run_config.yaml

  area:ckptcommunity-requestfeature
  #6197·andresnowak opened Sep 23, 2026·

  ·
  1

• Non-uniform FSDP

  area:recipebreaking-changefeature
  #6196·gautham-kollu opened Sep 23, 2026Contributor·

  ·
  1

• build: bump nvidia-modelopt to 0.47.0

  area:buildcicommunity-requestneeds-review
  #6195·chadvoegele opened Sep 23, 2026·

  ·
  1

• [sync][training] fix: honor checkpoint integrity verification

  area:ckptbug
  #6192·yaoyu-33 opened Sep 23, 2026Contributor·

  ·
  2

• fix(docker): preserve validated base runtime on opt-in builds

  area:buildbugneeds-review
  #6190·malay-nagda opened Sep 23, 2026Contributor·

  ·
  5

• fix(data): propagate unpacked SFT padding masks to MoE routers

  area:databugneeds-review
  #6188·cuichenx opened Sep 23, 2026Contributor·

  ·

• feat(qwen): configure vision recomputation independently of decoder

  area:modelfeatureneeds-review
  #6186·cuichenx opened Sep 22, 2026Contributor·

  ·

• feat(conversion): extract Qwen3.5 MoE text checkpoints with MTP

  area:ckptfeatureneeds-review
  #6185·cuichenx opened Sep 22, 2026Contributor·

  ·

• fix(peft): pass pg_collection in PackedPerExpertLinear.sharded_state_dict

  area:peftbugcommunity-requestneeds-review
  #6184·erictang000 opened Sep 22, 2026Contributor·

  ·
  2

• fix(training): unblock Super-3.5 LLM verification

  area:trainingbugneeds-review
  #6177·huvunvidia opened Sep 22, 2026Contributor·

  ·
  3