All pull requests

Pull requests

• feat(recipe): add DeepSeek V4 Pro BF16 GB300 performance recipe

  area:recipefeatureneeds-review
  #6746·cuichenx opened Oct 31, 2026Contributor·5 tasks done·

  ·
  1

• chore(deps): AUT-2692 bump Megatron-Core main to ba495158

  area:buildfull-test-suite
  #6745·svcnemo-autobot opened Oct 31, 2026Contributor·

  ·
  3

• fix(ckpt): preserve Omni task boundaries during weight export

  area:ckptbugneeds-review
  #6744·mxinO opened Oct 31, 2026Contributor·4 tasks done·

  ·
  2

• refactor(model): build Step-3.7-Flash from a builder-backed model config

  area:modelfeatureneeds-more-testsneeds-review
  #6743·kamran-nvidia opened Oct 30, 2026Contributor·

  ·
  4

• refactor(model): build EXAONE 4.5 from a builder-backed model config

  area:modelfeatureneeds-review
  #6742·kamran-nvidia opened Oct 30, 2026Contributor·3 of 5 tasks·

  ·
  1

• refactor(model): build MiniMax-M3 from builder-backed model configs

  area:modelfeatureneeds-review
  #6739·kamran-nvidia opened Oct 30, 2026Contributor·3 of 5 tasks·

  ·
  4

• Migrate Qwen 3.5 to hybrid model

  #6733·xuantengh opened Oct 30, 2026Contributor·5 tasks·

  ·
  5

• feat(quant): stream KV-cache scales in ModelOpt refits

  area:quantblockedfeature
  #6726·mxinO opened Oct 29, 2026Contributor·5 of 8 tasks·

  ·
  7

• feat(model): enable full-iteration CUDA graphs for Qwen3-VL text and Qwen3.5-VL images

  area:perffeatureneeds-review
  #6725·tomlifu opened Oct 29, 2026Contributor·2 of 5 tasks·

  ·
  6

• fix(ckpt): copy the source tokenizer and processor files on HF export

  area:ckptbugneeds-review
  #6724·kevalmorabia97 opened Oct 29, 2026Contributor·3 of 5 tasks·

  ·
  16

• perf(recipe): run the VRNVL72 DeepSeek V3 MXFP8 recipes with MLPerf execution settings

  area:recipefeatureneeds-review
  #6723·malay-nagda opened Oct 29, 2026Contributor·4 tasks done·

  ·
  1

• perf(recipe): enable NCCL EP for Qwen3 on B200 and B300

  26.10.01area:perfarea:recipefeatureneeds-reviewr0.7.0
  #6722·rhmukundan opened Oct 29, 2026Contributor·

  ·
  1

• perf(recipes): tune Nemotron 3 Super B200 and B300 NCCL-EP configs

  26.10.01area:perffeaturer0.7.0
  #6718·rhmukundan opened Oct 29, 2026Contributor·

  ·
  1

• Remove the data_parallel_sharding_strategy in MFSDPv2's config valida…

  area:trainingbug
  #6716·XiaodaNV opened Oct 29, 2026·5 tasks·

  ·
  1

• [Blocked on MLM #8234] fix(config): adapt Bridge serialization to generic Core helpers

  area:miscblockedbug
  #6715·cuichenx opened Oct 29, 2026Contributor·3 of 5 tasks·

  ·
  1

• fix(model): make full recompute work for Gemma 4 E2B and E4B

  area:modelbug
  #6713·kamran-nvidia opened Oct 28, 2026Contributor·

  ·
  2

• fix(model): match Gemma 4 dense VL image masks to Hugging Face

  area:modelbug
  #6712·kamran-nvidia opened Oct 28, 2026Contributor·3 of 5 tasks·

  ·
  3

• cp: Restore MLPerf GPTOSS 20B configs after 26.08 perf recipe refactor (6473) into r0.7.0

  cherry-pickRun CICD
  #6708·svcnvidia-nemo-ci opened Oct 28, 2026Contributor·

  ·
  2

• test: isolate fake shell commands from BASH_ENV

  area:cicicommunity-requestneeds-review
  #6705·bvolpato opened Oct 28, 2026·1 of 5 tasks·

  ·
  1

• perf(data): bound index workers by input count

  area:datacommunity-requestfeature
  #6704·bvolpato opened Oct 28, 2026·2 of 5 tasks·

  ·
  1

• test: batch Qwen VL recipe smoke tests

  area:recipecicommunity-request
  #6703·bvolpato opened Oct 28, 2026·1 of 5 tasks·

  ·
  1

• test: prepare functional release data on demand

  area:misccicommunity-requestneeds-review
  #6702·bvolpato opened Oct 28, 2026·1 of 5 tasks·

  ·
  1

• ci: mark release docs latest during code freeze

  area:misccineeds-review
  #6700·balasaajay opened Oct 27, 2026Contributor·

  ·
  22

• feat(quant): support LSQ scale learning for ModelOpt-quantized models

  community-request
  #6699·rohansjoshi opened Oct 27, 2026·3 of 5 tasks·

  ·
  5

• fix(ckpt): keep per-group LR overrides when restoring scheduler bounds on resume

  area:ckptbugcommunity-requestneeds-review
  #6698·rohansjoshi opened Oct 27, 2026·2 of 5 tasks·

  ·
  4