All pull requests

Pull requests

• perf(pipeline): pool receive buffers for interleaved 1F1B

  #4308·HuiyingLi opened Oct 12, 2026Contributor·3 tasks done·

  ·
  1

• fix(vlm): use parallel CE for Muse Glimmer TP2 OOM

  r0.7.0
  #4307·akoumpa opened Oct 12, 2026Contributor·2 tasks done·

  ·
  5

• fix(qwen3.5-moe): isolate packed text metadata from vision

  r0.7.0
  #4306·akoumpa opened Oct 12, 2026Contributor·3 tasks done·

  ·
  2

• fix(mimo): pin HybridEP domains to GB200 hosts

  docs-onlyr0.7.0
  #4305·akoumpa opened Oct 12, 2026Contributor·4 tasks done·

  ·
  2

• fix(vlm): pad variable-resolution images in Ministral recipes

  docs-onlyr0.7.0
  #4304·akoumpa opened Oct 12, 2026Contributor·3 tasks done·

  ·
  3

• fix(ci): route MoK benchmarks to four-GPU GB200 nodes

  docs-onlyr0.7.0
  #4302·akoumpa opened Oct 12, 2026Contributor·4 tasks done·

  ·
  3

• ci: reuse release dependency caches in cherry-pick PRs

  #4299·akoumpa opened Oct 12, 2026Contributor·3 of 4 tasks·

  ·
  2

• fix(nemotron-parse): preserve processor normalization flags

  #4298·akoumpa opened Oct 12, 2026Contributor·3 tasks done·

  ·
  1

  2

• feat(dspark): support sequence packing for DeepSeek V4 and V4.1

  community-request
  #4295·khazic opened Oct 12, 2026Contributor·

  ·
  1

• fix(dspark): train DeepSeek-V4.1 on its released chat template

  community-request
  #4294·khazic opened Oct 12, 2026Contributor·

  ·
  1

• fix(recipes): use return_hidden_states for the fused-loss hidden states

  #4288·yisongbetter opened Oct 12, 2026Contributor·4 tasks done·

  ·
  3

• perf(moe): fused clamped-SwiGLU activations for shared and routed experts

  #4286·yisongbetter opened Oct 12, 2026Contributor·4 tasks done·

  ·
  4

• perf(fsdp): cached FusedAdam lists and a cross-block FSDP unit for fp32 holders

  #4285·yisongbetter opened Oct 12, 2026Contributor·4 tasks done·

  ·
  3

• perf(deepseek-v41): compile island for the cache quantize/dequantize core

  #4284·yisongbetter opened Oct 12, 2026Contributor·4 tasks done·

  ·
  5

• fix(checkpoint): optionally use Gloo for load metadata

  #4279·HuiyingLi opened Oct 11, 2026Contributor·4 tasks done·

  ·
  8

• chore: CI proxy for #4178

  #4276·HuiyingLi opened Oct 11, 2026Contributor·

  ·
  3

• chore: CI proxy for #4182

  #4275·HuiyingLi opened Oct 11, 2026Contributor·

  ·
  3

• fix(peft): preserve MoE LoRA base parameter storage

  #4260·HuiyingLi opened Oct 10, 2026Contributor·3 tasks done·

  ·
  1

• feat(packaging): ship tools and examples under nemo_automodel

  #4257·akoumpa opened Oct 9, 2026Contributor·3 tasks done·

  ·
  18

• docs(fern): add new model performance by GPU type

  docs-onlyr0.7.0
  #4256·akoumpa opened Oct 9, 2026Contributor·3 tasks done·

  ·
  1

• fix(checkpoint): apply Qwen MTP FP8 scales (AM-864)

  r0.7.0
  #4250·HuiyingLi opened Oct 9, 2026Contributor·3 of 4 tasks·

  ·
  4

• perf(recipe): target GB200 for Super 3.5 VL 256K packed SFT

  docs-only
  #4243·HuiyingLi opened Oct 9, 2026Contributor·3 tasks done·

  ·
  4

• chore: CI proxy for #4151

  #4236·HuiyingLi opened Oct 9, 2026Contributor·

  ·
  4

• fix(vlm): use generation masks for GLM training labels

  #4233·HuiyingLi opened Oct 9, 2026Contributor·3 tasks done·

  ·
  8

• refactor(moe): move DeepEP/HybridEP dispatch replay onto the shared channel

  #4232·Agoniii opened Oct 9, 2026Collaborator·3 tasks done·

  ·
  4