All pull requests

Pull requests

• [kv_transfer] lmcache_mp: lookup thread, transfer thread, same-step saves (042 experiment)

  #36·yhl-amd opened Oct 29, 2026Owner·

  ·

• [DSV4.1] DSpark probabilistic draft + block verification (ATOM_DSPARK_PROB_DRAFT=1)

  #34·yhl-amd opened Oct 26, 2026Owner·

  ·

• [DSV4.1] opt-in per-step timer in ModelRunner (ATOM_STEP_TIMER=1)

  #33·yhl-amd opened Oct 26, 2026Owner·

  ·

• docs(v41): run DeepSeek-V4.1-Flash on gfx1250 (MI455X)

  #32·yhl-amd opened Oct 23, 2026Owner·

  ·

• [gfx1250] DeepSeek-V4.1: FlyDSL FP8 GEMMs and paged MQA logits

  #30·yhl-amd opened Oct 23, 2026Owner·

  ·

• [DSV4] Bounded replay over LMCache MP: load the PAGE KV alone, then replay

  #29·yhl-amd opened Oct 11, 2026Owner·4 tasks done·

  ·

• [DSV4] Bounded replay: rebuild window state from the compressed history

  #27·yhl-amd opened Oct 11, 2026Owner·5 of 6 tasks·

  ·

• [DP] Balance lockstep prefill steps by estimated chunk cost

  #26·yhl-amd opened Oct 11, 2026Owner·7 tasks done·

  ·

• [offload] Take the lmcache_mp lookup off the DPA scheduler path

  #25·yhl-amd opened Oct 1, 2026Owner·3 tasks done·

  ·
  1

• feat(offload): support Kimi-K3 with lmcache_mp

  #24·yhl-amd opened Sep 30, 2026Owner·6 tasks done·

  ·
  1

• [DSV4] Run HCA fp8 decode through aiter's persistent V4-NM kernel

  #23·yhl-amd opened Sep 29, 2026Owner·3 tasks done·

  ·

• MegaMoE backend: DP pad-row masking, configurable decode capacity

  #22·yhl-amd opened Sep 29, 2026Owner·

  ·

• epx EP backend (--all2all-backend epx) with epx-tuned fused_moe

  #21·yhl-amd opened Sep 29, 2026Owner·

  ·

• fix(offload): unregister lmcache_mp caches when the engine exits

  #19·yhl-amd opened Sep 29, 2026Owner·4 tasks done·

  ·
  1

• fix(offload): stop refusing native-state MP restores on a byte budget

  #18·yhl-amd opened Sep 29, 2026Owner·5 tasks done·

  ·
  1

• feat(offload): support DeepSeek-V4.1 with lmcache_mp

  #17·yhl-amd opened Sep 29, 2026Owner·6 tasks done·

  ·
  1

• feat(offload): serve the FP4 sparse indexer to lmcache_mp on MLA

  #16·yhl-amd opened Sep 29, 2026Owner·6 tasks done·

  ·
  1

• fix(offload): publish lmcache_mp PAGE-only loads into the GPU prefix index

  #15·yhl-amd opened Sep 29, 2026Owner·4 tasks done·

  ·
  1

• chore: test empty pull request

  #11·yhl-amd opened Aug 27, 2026Owner·

  ·
  1

• feat: add GLM-5.2 LMCache MP offload support

  #10·yhl-amd opened Aug 19, 2026Owner·

  ·

• Add xgrammar tool-call constrained decoding for DeepSeek-V4 (fix agentic eval 0%→63%)

  #9·yhl-amd opened Jul 30, 2026Owner·

  ·

• [Frontend] DeepSeek-V4 native OpenAI/Anthropic/Responses API + DSML tool parser

  #8·yhl-amd opened Jul 30, 2026Owner·

  ·
  3

• Refactor DSV4 KV cache control and data planes

  #6·yhl-amd opened Jul 29, 2026Owner·

  ·

• fix: make GLM sparse indexer decode top-k deterministic

  #4·yhl-amd opened Jul 14, 2026Owner·

  ·

• Review LMCache offload compatible connector and bounded staging

  #3·yhl-amd opened Jun 4, 2026Owner·

  ·
  2