All pull requests

Pull requests

• bench: community report — Strix Halo laptop (Radeon 8060S, 70 W), v0.1.41, UD-IQ4_XS

  #1716·nsbradley88 opened Oct 9, 2026Contributor·

  ·

• sycl/Arc: stepped verify window (opt-in STRATA_VERIFY_STEPPED) - CPU experts served between graph segments

  #1713·aslater3 opened Oct 9, 2026·

  ·

• sycl: opt-in grouped XMX expert GEMM for the prompt path (STRATA_PF_XMX), with a parity test

  #1711·aslater3 opened Oct 9, 2026·

  ·

• sycl: experimental A770 fast router and paired mapped transfers

  #1710·duncanmcqueen opened Oct 9, 2026·

  ·

• docs: improve Japanese README wording and readability

  #1708·Moge800 opened Oct 9, 2026·

  ·

• Add GPU energy monitoring and cache-aware token cost analytics

  #1707·hash13 opened Oct 9, 2026Contributor·

  ·

• prefill: opt-in active-layer cache, P4 108.1 -> 115.8 tok/s

  #1704·CC-David-CC opened Oct 9, 2026Contributor·

  ·

• bench: validate 128K-1M live, RAM and disk cache reuse

  #1702·CC-David-CC opened Oct 9, 2026Contributor·

  ·

• cpu: q2_avx1 - the Q2_0 rows and the Q8_1 quantizer for CPUs with AVX but no AVX2 (fixes #1699)

  #1701·nashcap opened Oct 9, 2026·

  ·
  1

• generate, mtp: near the context's end the spec window and its drafts stay inside it

  #1700·midagedev opened Oct 9, 2026Contributor·

  ·

• feat: experimental RoPE to YaRN 4x saved-cache migration

  #1698·CC-David-CC opened Oct 9, 2026Contributor·

  ·

• cpu: AVX1 128-bit expert kernels (iq_avx1) - the i-quant rows on CPUs with no AVX2

  #1697·nashcap opened Oct 9, 2026·

  ·
  1

• Resident RAM mode on Windows: the start-up fill and the RAM copy read experts.bin unbuffered (ready 42 -> 10 s with 21 GiB free)

  #1696·sergqwer opened Oct 9, 2026·

  ·

• bench: P4 prefill 78.8 -> 114.8 tok/s with existing auto sizing; expert reuse roofline

  #1695·CC-David-CC opened Oct 9, 2026Contributor·

  ·

• mcp + docs: let the swift family take IQ3_S too (follow-up to #1688)

  #1694·superlee-terry opened Oct 9, 2026·

  ·

• setup: the page-file warning advises a fixed size, and says why when Windows is set to grow the file (#60)

  #1693·sergqwer opened Oct 9, 2026·

  ·

• bench: existing RoPE vs YaRN at 512K and 1M, with setup guide

  #1692·CC-David-CC opened Oct 9, 2026Contributor·

  ·

• kv-grow: kvg_trim uploads the residency table with res_put, which waits for the copy

  #1690·sergqwer opened Oct 9, 2026·

  ·

• setup: let the swift family take IQ3_S (#1651)

  #1688·nekomario28 opened Oct 9, 2026·

  ·
  1

• docs: STRATA_PREFILL_CPU_SHARE=auto is not repeatable run to run (#1684)

  #1687·nekomario28 opened Oct 9, 2026·

  ·

• bench: community report, RTX 5080 16 GB + RTX 4060 Ti 8 GB (helper cache), IQ3_S at 4K / 32K / 128K on 0.1.41

  #1681·KarlGrier opened Oct 9, 2026·

  ·

• Windows: cap conversation-memory admission at the commit limit too (#1607)

  #1679·nekomario28 opened Oct 9, 2026·

  ·
  1

• serve: keep-alive while a tool call's held-back argument is generated (#1666)

  #1678·nekomario28 opened Oct 9, 2026·

  ·
  1

• setup: pass CC/CXX/CUDAHOSTCXX to the CUDA build's cmake (#1645)

  #1677·nekomario28 opened Oct 9, 2026·

  ·
  1

• verify: initialize remote expert helper state in both batch paths

  #1676·saikiran-rs opened Oct 9, 2026Contributor·

  ·