All pull requests

Pull requests

• K2 Horizon: cached norm unit weight; bit-exact fast decode QMV for dense JANGH

  #567·jjang-ai opened Oct 8, 2026Contributor·

  ·
  1

• Preserve native Qwen XML tool string bytes

  #547·jjang-ai opened Oct 3, 2026Contributor·

  ·

• Keep Qwen3.5 warm text prefill at the cache offset

  #546·jjang-ai opened Oct 3, 2026Contributor·

  ·

• Schedule JANGH prefill tiles within expert boundaries

  #531·jjang-ai opened Sep 29, 2026Contributor·

  ·
  3

• Add validated JANGH packed decode primitives and no-copy bank admission

  #528·jjang-ai opened Sep 28, 2026Contributor·

  ·
  1

• Preserve routed activation and fused SwiGLU contracts

  #527·jjang-ai opened Sep 28, 2026Contributor·

  ·

• Correct explicit sliding windows over compiled rotating caches

  #524·jjang-ai opened Sep 28, 2026Contributor·

  ·
  1

• DFlash 2 for Spark2.5: capture, exact sliding-window rollback, bundled drafters, throughput governor

  #521·jjang-ai opened Sep 27, 2026Contributor·

  ·

• Fix MLXDistributedTP collective results on multi-rank groups

  #520·jjang-ai opened Sep 26, 2026Contributor·

  ·

• Keep the batched post-answer cache boundary when tools were only offered

  #478·jjang-ai opened Sep 17, 2026Contributor·3 of 4 tasks·

  ·

• MTP: measure initial admission and park on measured AR loss

  #470·jjang-ai opened Sep 12, 2026Contributor·

  ·
  1

• Native MTP: measured admission, direct AR parking, and Flash cache checkpoints

  #468·jjang-ai opened Sep 12, 2026Contributor·

  ·
  3

• MTP: preserve productive depth across AR calibration

  #467·jjang-ai opened Sep 11, 2026Contributor·

  ·

• Respect fixed native MTP depth ceilings and use monotonic timing

  #460·jjang-ai opened Sep 9, 2026Contributor·

  ·
  15

• Fix Qwen4Exp long-prefill QSA allocation

  #369·jjang-ai opened Aug 30, 2026Contributor·

  ·

• Pin Stop cache safety and terminal ordering contracts

  #340·jjang-ai opened Aug 29, 2026Contributor·

  ·

• Stream DSV4 affine routed experts from exact SSD regions

  #338·jjang-ai opened Aug 29, 2026Contributor·

  ·
  5

• Preserve nested JSON Schema in Zaya tool declarations

  #336·jjang-ai opened Aug 29, 2026Contributor·

  ·
  4

• DiskCache: a store too big for the cap took the whole cache with it

  #293·jjang-ai opened Aug 22, 2026Contributor·

  ·

• DeepSeek-OCR cache salt, and pin what a growing VL conversation can reuse

  #292·jjang-ai opened Aug 22, 2026Contributor·

  ·

• NemotronLabs VoiceChat 11B: full-duplex speech-to-speech runtime

  #285·jjang-ai opened Aug 21, 2026Contributor·

  ·

• Add opt-in DSV4 FP32 output-head cache with teardown lifecycle and token-ID trace

  #214·mimeding opened Aug 7, 2026Contributor·5 of 9 tasks·

  ·

• Fix nested LFM/DSML tool arguments and quantized LFM loads

  #180·jjang-ai opened Jul 24, 2026Contributor·

  ·

• Fix Zaya XML parameter resynchronization

  #162·jjang-ai opened Jul 20, 2026Contributor·

  ·
  1

• [WIP] MiniMax-M3 runtime: sparse MSA cache + indexer decode

  #75·jjang-ai opened Jun 18, 2026Contributor·

  ·