All pull requests

Pull requests

• docs(examples): Qwen3.5-35B-A3B profile and result for the TimeLens recipe

  #4746·Rockdu opened Nov 9, 2026Collaborator·8/8·5 tasks done·

  ·

• refactor(megatron): pass train_one_step input and output as dataclasses

  #4745·guapisolo opened Nov 9, 2026Collaborator·

  ·

• build(docker): retry transient wheel downloads

  run-ci-image
  #4743·HJSang opened Nov 9, 2026Collaborator·

  ·

• docs(examples): geo3k token-space recipes for Qwen3.5-35B-A3B, Qwen3.8-27B, and Inkling

  #4742·Rockdu opened Nov 9, 2026Collaborator·7/8·4 of 5 tasks·

  ·

• feat(checkpoint): decay old checkpoints, keeping the newest N and every Mth rollout

  #4741·andreas-inherent opened Nov 9, 2026·

  ·
  1

• Honor FlashInfer tie policies in DeepSeek-V4.1 candidate selection

  #4740·zianglih opened Nov 9, 2026Contributor·

  ·
  1

• [feat] Add --eval-source-path to build the eval datasets from a function

  #4739·ben-av opened Nov 9, 2026·

  ·
  1

• [megatron] perf: optimize CP all-to-all relayout

  #4737·YolandaLyj opened Nov 9, 2026Contributor·

  ·

• Carve rollout engines out of training for each eval (--eval-pool-num-engines)

  #4736·yueming-yuan opened Nov 9, 2026Collaborator·

  ·

• feat(mm): run Inkling audio and image RL on the token-space processor

  run-cirun-ci-fsdprun-ci-lorarun-ci-model-scriptsrun-ci-precisionrun-ci-qwen35run-ci-replayrun-ci-sglangrun-ci-weight-update
  #4732·Rockdu opened Nov 8, 2026Collaborator·6/8·

  ·

• fix(weight-update): avoid multi-root all-gather in ExpertGather

  #4731·LiRunGuo opened Nov 7, 2026·

  ·
  1

• [TPU] Score each CP rank's share of the logit rows instead of gathering [T, V] logits

  run-ci-torchtitan
  #4728·HJSang opened Nov 6, 2026Collaborator·

  ·
  4

• [TPU] Context parallelism as Ulysses around the flash attention

  run-ci-torchtitan
  #4727·HJSang opened Nov 6, 2026Collaborator·

  ·
  4

• [Perf] Batch NVFP4 expert quantization during weight sync

  #4725·zianglih opened Nov 6, 2026Contributor·

  ·

• TPU trainer: compile the first step while the engine starts (--compile-warmup)

  run-ci-rolloutrun-ci-torchtitan
  #4724·HJSang opened Nov 6, 2026Collaborator·

  ·
  40

• [DeepSeek-V4.1-Flash] Sync FP8 weights in serving's layouts with FP32 block scales

  #4723·adrenaline21 opened Nov 6, 2026Collaborator·

  ·

• Add per-op pins of the miles-ext-kernels submodule

  #4722·yyihuang opened Nov 6, 2026·

  ·
  1

• Re-derive DeepSeek-V4 session turns from the sampled tokens with the vendored parser

  #4719·HJSang opened Nov 6, 2026Collaborator·

  ·
  6

• test(session): show the assistant_text diff when the soft gate fails

  #4717·HJSang opened Nov 6, 2026Collaborator·

  ·
  10

• [LoRA] torchtitan multi-slot LoRA and AdapterSlots compute members

  run-ci-lorarun-ci-torchtitan
  #4716·yueming-yuan opened Nov 6, 2026Collaborator·

  ·

• [WIP] feat(algo): add REINFORCE Zero (--advantage-estimator reinforce_zero)

  #4715·catherinelee274 opened Nov 6, 2026·1 task·

  ·
  1

  1

• fix(megatron): bump the weight version once per rollout, not per optimizer step

  #4714·guapisolo opened Nov 6, 2026Collaborator·

  ·

• [TPU] Launchers for Qwen3-30B-A3B and GLM-4.7-Flash 4-layer, and tools/prune_hf_layers.py

  #4713·Zhichenzzz opened Nov 6, 2026Collaborator·3 tasks·

  ·

• feat(tau2-bench): classify cut-off and failed episodes in the example adapter

  #4711·guapisolo opened Nov 6, 2026Collaborator·

  ·
  8

• Tinker gateway on native LoRA: raw launcher mode and padded-vocab HF load

  run-ci-lorarun-unit-test
  #4710·yueming-yuan opened Nov 6, 2026Collaborator·4/4·

  ·
  13