All pull requests

Pull requests

• MUSA: gate the hybrid KV admission fast path to Qwen3.5-0.8B/2B

  #317·lllszzz opened Oct 30, 2026·

  ·

• fix(musa): default DeepSeek-OCR vision attention to Triton

  #316·Elon-yf opened Oct 30, 2026·

  ·

• [MUSA] fa3: per-attention-window AOT scheduler metadata (opt-in)

  #314·yeahdongcn opened Oct 29, 2026Collaborator·4 of 6 tasks·

  ·
  1

• MUSA: route Mamba2 SSD through MATE's flashinfer.mamba on v0.32.0-dev

  #312·yeahdongcn opened Oct 28, 2026Collaborator·

  ·
  10

• tactics: one table for measured kernel numbers, keyed on a pattern

  #310·yeahdongcn opened Oct 28, 2026Collaborator·

  ·
  1

• musa: take the arch facts from torchada and gate on them

  #308·yeahdongcn opened Oct 26, 2026Collaborator·

  ·
  36

• [MUSA] Add CAR v2 fusion with Qwen Gemma normalization support

  #304·Elon-yf opened Oct 23, 2026·

  ·
  26

• feat(musa): enable DFlash2 and DSpark2 by sizing the separate attention pool from layout strides

  #303·lllszzz opened Oct 22, 2026·

  ·
  3

• fix(startup): pre-build both rope store-hint variants; drop the stale cache switch

  #300·yeahdongcn opened Oct 22, 2026Collaborator·

  ·
  5

• feat(multiarch): make the MUSA architecture a fact - device spec, arch-keyed JIT cache, build stamp

  #299·yeahdongcn opened Oct 22, 2026Collaborator·

  ·
  14

• feat(tactics): data-driven tactic tables with a fail-closed lookup

  #297·yeahdongcn opened Oct 22, 2026Collaborator·

  ·
  5

• kernels: serve W4A16 linear layers with MATE's Marlin-class kernel (mp_31)

  #287·yeahdongcn opened Oct 18, 2026Collaborator·

  ·
  12

• feat(musa): adapt v0.28.0 for torch post2 and triton 3.6

  #265·lllszzz opened Sep 28, 2026·

  ·
  1

• MUSA: route the Mamba2 SSD to FlashInfer for --mamba-backend flashinfer

  #257·yeahdongcn opened Sep 23, 2026Collaborator·

  ·

• MUSA: add DFlash2 for Qwen3.8-27B adaptation patch

  #230·lllszzz opened Sep 10, 2026·

  ·

• [MUSA] Add vLLM-Omni 0.28.0 image target

  #220·yeahdongcn opened Sep 2, 2026Collaborator·

  ·

• [Model] Enable GLM-5.3-Flash on MUSA

  #209·lllszzz opened Aug 31, 2026·

  ·

• [Model] Enable Tencent Hy4-preview on MUSA (v0.28.0)

  #205·yeahdongcn opened Aug 29, 2026Collaborator·

  ·
  1

• perf(musa): select AOT and JIT tiles by active MP

  #202·yeahdongcn opened Aug 28, 2026Collaborator·

  ·
  6

• [Model] Enable Qwen3.8-Flash-Next on MUSA

  #200·Joey-gvwal opened Aug 27, 2026Collaborator·

  ·
  1

• MUSA: enable CAR-RMSNorm by default under contract

  #199·Elon-yf opened Aug 27, 2026·

  ·
  18

• [Draft] feat(musa): unify RuntimePlan and add in-tree engine autotuning

  #185·yeahdongcn opened Aug 17, 2026Collaborator·

  ·
  18

• MUSA: switch Ubuntu 22.04 image to system Python 3.12

  #119·yeahdongcn opened Jul 24, 2026Collaborator·

  ·
  1