All pull requests

Pull requests

• exp - feat(serving): SGLang MoE DSpark draft loader for Kan's MoE drafter (#811/#812), normalizer arch name and equivalence gate (stacked on #931)

  #940·yushengsu-thu opened Oct 6, 2026·

  ·

• feat(mooncake): read server captures over multi-node NVLink

  #939·maocheng23 opened Oct 6, 2026Collaborator·

  ·

• feat(capture): NVLink publication in the v0.5.18 capture sink

  #938·maocheng23 opened Oct 6, 2026Collaborator·

  ·

• ci: fix live server-capture gate (runner lost NO_PROXY) + print server.log on failure

  #937·maocheng23 opened Oct 6, 2026Collaborator·

  ·

• perf(dflash): compiled create_block_mask / flex_attention wrappers [3/3]

  #936·maocheng23 opened Oct 6, 2026Collaborator·

  ·

• feat(dflash): share DFlash masks with inference, default masks for spec_generate [2/3]

  #935·maocheng23 opened Oct 6, 2026Collaborator·

  ·
  3

• fix(dflash): honor config.is_causal in DFlash training masks [1/3]

  #934·maocheng23 opened Oct 6, 2026Collaborator·

  ·

• fix(data): supervise historical Qwen thinking turns

  #933·cheater6666 opened Oct 6, 2026·4 tasks done·

  ·
  1

• (fsdp2 dev.) feat(moe): expert parallelism for MoE drafters on the FSDP2 backend (stacked on #915 + #931)

  #932·yushengsu-thu opened Oct 5, 2026·

  ·

• exp.[rebase] MoE FFN for DFlash-family drafters (#811/#812) on current main, to run the DeepSeek-V4-Flash MoE arm

  #931·yushengsu-thu opened Oct 5, 2026·

  ·
  1

• exp - feat(moe): qwen3 preset and the Qwen3.8-27B DFlash2 MoE ablation arm (stacked on #811/#812)

  #930·yushengsu-thu opened Oct 5, 2026·

  ·
  2

• (fsdp2 dev.) - follow-up : Add training.static_shape_buckets: a few static lengths for compiled blocks

  #929·yushengsu-thu opened Oct 4, 2026·

  ·
  1

• (fsdp2 dev. perf) Add sequence packing for EAGLE3, DFlash, and DFlash2

  #928·yushengsu-thu opened Oct 4, 2026·4 tasks done·

  ·
  1

• (fsdp2 dev. perf) Add training.checkpoint_async: background checkpoint writes

  #925·yushengsu-thu opened Oct 4, 2026·

  ·

• (fsdp2 dev. perf) Add training.fp8_linear: torchao Float8Linear with float8 all-gather

  #923·yushengsu-thu opened Oct 4, 2026·

  ·

• (fsdp2 dev. perf) Add training.compile_blocks: per-block torch.compile before fully_shard

  #922·yushengsu-thu opened Oct 4, 2026·

  ·

• [exp. do not merge] Compare FSDP, native TorchTitan and GraphTrainer

  #921·yushengsu-thu opened Oct 3, 2026·4 tasks done·

  ·

• [exp. do not merge] Add native TorchTitan training runtime for DFlash-family drafts

  #920·yushengsu-thu opened Oct 2, 2026·4 tasks done·

  ·

• fix(data): forward loss_mask_filter on the raw-conversation prompt path

  #919·li-lizhe opened Oct 2, 2026·

  ·
  1

  1

• feat(training): add H-Spec (mamba_attn_hybrid) training with sglang offline capture

  #918·curnane-lab opened Oct 2, 2026Collaborator·6 tasks·

  ·

• Add DSpine draft model and training support

  #917·maocheng23 opened Oct 2, 2026Collaborator·3 of 4 tasks·

  ·

• [opt] Add opt-in length-aware online scheduling and offline bucketing

  #916·yushengsu-thu opened Oct 2, 2026·3 of 4 tasks·

  ·

• [verifying] Add optional FSDP2 training backend

  #915·yushengsu-thu opened Oct 2, 2026·

  ·
  3

• fix(data): restore the deepseek-r1-distill chat template

  #914·yushengsu-thu opened Oct 1, 2026·

  ·
  1

• [Feature] Refactor examples/configs filename

  #911·jiapingW opened Sep 30, 2026Collaborator·5 tasks·

  ·
  5