All pull requests

Pull requests

• [lite] Consolidate DS4.1 FP4 training, resync and strict GRPO recipe

  #254·ISEEKYAN opened Oct 15, 2026Owner·

  ·

• lite: batch deterministic FP8 and sparse-attention launches

  #253·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• lite: use native ordered RoPE for deployment index queries

  #252·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• lite: use native FP32 MXFP4 sparse-index scores

  #251·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• lite: match deployment SWA KV RoPE before FP8 encoding

  #250·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• lite: stage atomic Muon updates with FP32 host state

  #249·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• lite: preserve frozen Engram across incremental resync

  #248·ISEEKYAN opened Oct 11, 2026Owner·

  ·

• Consume DS4.1 W4A8 and deployment math after explicit prerequisites

  #247·ISEEKYAN opened Oct 6, 2026Owner·

  ·

• fix(verl): support multinode DP with remote headless engines

  #246·ISEEKYAN opened Oct 6, 2026Owner·

  ·

• fix(lite): preserve packed GRPO head and owned DDP residency

  #245·ISEEKYAN opened Oct 6, 2026Owner·

  ·

• feat(lite): add opt-in deployment math with owned reference VJPs

  #244·ISEEKYAN opened Oct 6, 2026Owner·

  ·

• [lite] Make weight-only QAT cover TE GroupedLinear expert weights

  #243·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• [lite] Opt-in W4A8 routed experts (dynamic FP8 activations x MXFP4 weights)

  #242·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• [lite] add packed document objectives and router replay for V4

  #241·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• [lite] Add block32 FP8 Linear and native FP32 wgrad providers, with DeepSeek-V4 CSA as consumer

  #240·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• [lite] Make the isort and Black pre-commit hooks converge

  #239·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• [lite] Add bounded checkpoint rows and binding inventory, with Qwen3-MoE as consumer

  #238·ISEEKYAN opened Oct 3, 2026Owner·

  ·

• Handle already-closed DCP streams when publishing staged checkpoints

  #237·aoshen02 opened Oct 2, 2026·

  ·
  1

• Restore configured mLite optimizer offload after checkpoint loading

  #236·aoshen02 opened Oct 2, 2026·

  ·
  1

• Fix mLite checkpoint collisions between EP-local experts

  #235·aoshen02 opened Oct 2, 2026·

  ·
  1

• Fix FSDP2 Adam checkpoint restore layout and mmap lifetime

  #234·aoshen02 opened Oct 2, 2026·

  ·
  1

• feat(ds41): add bounded deployment resync export and receiver

  #231·ISEEKYAN opened Sep 20, 2026Owner·

  ·

• Fix generic verl connector policies and add a DS4.1 SFT recipe

  #230·ISEEKYAN opened Sep 19, 2026Owner·

  ·

• Rebuild DeepSeek-V4.1 model support with explicit topology and core CSA compatibility

  #227·ISEEKYAN opened Sep 18, 2026Owner·

  ·

• Add compact Qwen3 ChunkedEP overlap for NVIDIA dev

  #225·ISEEKYAN opened Sep 14, 2026Owner·

  ·
  10