All pull requests

Pull requests

• setup: the disk check counts only what step 6 still writes

  #897·bsvinay opened Oct 5, 2026·

  ·

• ple: accept Q5_1 and Q8_0 n-gram tables

  #896·anon761 opened Oct 5, 2026·

  ·

• hip: support Radeon gfx1151 APUs

  #895·DTU-travelpals opened Oct 5, 2026·

  ·

• serve: decode chunked request bodies (fixes empty-body 400 behind some proxies)

  #894·oscar-investmatic opened Oct 5, 2026·

  ·
  1

• sycl: read doorbell flags with an uncached L1+L3 hint (the GPU never saw the host's store on an Arc Pro B60)

  #889·joeyjoe02 opened Oct 5, 2026·

  ·

• hip: chunk-parallel GDN recurrence (WY form), two parity-gated arms - opt-in STRATA_GDN_WY (measured on gfx1100)

  #888·xyzzing opened Oct 5, 2026Contributor·

  ·

• layer split: score every four-way placement instead of guessing it

  #887·gopinath87607 opened Oct 5, 2026·

  ·

• serve: skip completed empty assistant turns in chat history

  #886·Arthur031221 opened Oct 5, 2026·

  ·
  1

• Perf/duplex transfers only

  #885·CC-David-CC opened Oct 5, 2026Contributor·

  ·

• sm_60: PTX vmad for the __dp4a fallback (bit-exact, ~2.2x in isolation)

  #883·willowmerestudios opened Oct 5, 2026·

  ·

• bench: community report, Flash-Next IQ2_XS vs a dense 27B as coding agents (RTX 4090 + 32 GB RAM)

  #882·T-Crypt opened Oct 5, 2026·

  ·

• layer split: the stage weight trim works under auto too 4X gain in the prefill in 4 way ring

  #880·gopinath87607 opened Oct 5, 2026·

  ·

• # Support sm_60, sm_70, and sm_86 with CUDA 12.6 or 12.8

  #878·FearL0rd opened Oct 5, 2026·6 tasks done·

  ·
  1

• Add a prometheus metrics endpoint

  #877·zeevo opened Oct 5, 2026·

  ·

• Asynchronous adaptive expert tier for the resident RAM mode (--adapt-async 1, opt-in)

  #876·Hardin22 opened Oct 5, 2026Contributor·

  ·
  1

• Fix first-kernel-launch crash on Windows (ROCm/Adrenalin WDDM startup race)

  #873·unimo-chan opened Oct 5, 2026·

  ·

• serve: opt-in recovery from repeated reasoning passages

  #869·huangserva opened Oct 5, 2026·

  ·

• docs: Arc Pro B60 rows and notes, and the B60's PCI id (e211) in setup_intel.py

  #868·LocalXPU opened Oct 5, 2026·

  ·
  2

• sycl: keep cudaStreamQuery's answer in the per-layer ring waits

  #866·LocalXPU opened Oct 5, 2026·

  ·
  1

• ngram: add opt-in Q8_0 PLE table support

  #865·CC-David-CC opened Oct 5, 2026Contributor·

  ·

• perf: opt-in resident expert exchange buffer rotation

  #864·CC-David-CC opened Oct 5, 2026Contributor·

  ·

• cpu kernels: gathered decode on the cores where it is faster, AVX-VNNI rows, IQ3_S one-token kernel

  #863·Hardin22 opened Oct 5, 2026Contributor·

  ·

• serve: "strata_checkpoint": false lets a one-shot request skip its conversation checkpoint (#830)

  #861·brenoperucchi opened Oct 5, 2026·

  ·

• Unsloth UD-Q6_K_XL: Q8_0 PLE rows and Q6_K gate/up experts

  #860·DaRealDaHoodie opened Oct 5, 2026·

  ·

• Layer split: pipelined verify windows for one conversation (--pipeline-windows, opt-in)

  #859·Hardin22 opened Oct 4, 2026Contributor·

  ·