All pull requests

Pull requests

• hip: the fused-int8 MoE kernels on gfx1100 rocWMMA - opt-in STRATA_PF_FUSED arm + parity fixture (engagement: parity, REJECT-at-parity recorded)

  #999·xyzzing opened Oct 5, 2026Contributor·

  ·

• serve: an engine that exits after an unread ERR line says why (#997 #890)

  #998·ischencheng opened Oct 5, 2026·

  ·

• hip: accept gfx1010 and gfx1011 (RDNA1) in the arch gate

  #996·AcckiyGerman opened Oct 5, 2026·

  ·

• Community benchmark: RTX 4070 Ti SUPER 16 GiB (Shin-BlackMamba, upstream 6f32ec07)

  #995·lfontanez opened Oct 5, 2026·

  ·

• Isolate concurrent image positions and validate vision records

  #994·ilumn opened Oct 5, 2026·

  ·

• bench: community report, RTX 5070 Ti + Ryzen 9 5900X, 64 GB DDR4-3600 (IQ2_XS / IQ3_XXS / IQ3_S, six-arm A/B)

  #993·obiscr opened Oct 5, 2026·

  ·

• security-hardening

  #992·aelnaby opened Oct 5, 2026·

  ·

• tools: bench_eviction — the R4.1 sweep: compulsory-miss vs LRU vs LFU-decay vs oracle on real routing traces

  #991·ZhongUncle opened Oct 5, 2026·

  ·

• serve: /props reports chat_template_caps so Zed offers tools (#986)

  #987·1it opened Oct 5, 2026·

  ·
  1

  9

• Added bazzite / fedora / multi gpu support

  #982·runningman84 opened Oct 5, 2026·

  ·

• hip: a rocBLAS solution table for the FP16 prompt GEMMs on gfx103x (the GDN projection 7x below 1,152 tokens; short prompts -10 to -14%)

  #981·xjc10 opened Oct 5, 2026·

  ·

• Fix hard-coded path in sycl/tools/build.sh

  #972·subsevenx2001 opened Oct 5, 2026·

  ·

• serve: Q4 at 297.3 tok/s non-MTP (N=8); MTP +30.4% decode (N=2)

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

  ·

• serve: recover tool calls the model writes next to the template's form, and return the rest as content

  #966·signalnine opened Oct 5, 2026·

  ·

• serve: pin Claude Code's per-request billing stamp, so the conversation cache reuses agent prompts

  #965·signalnine opened Oct 5, 2026·

  ·

• feat(server): stop unchanged completed tool-call loops

  #963·a00012025 opened Oct 5, 2026·

  ·

• serve: prefix snapshots on disk - a new chat restores its system prompt instead of reading it again

  #960·konijiwa110 opened Oct 5, 2026·

  ·

• #583: the fused layout shrinks the MoE buffers only when every layer can take the fused path

  #958·Krypto-Whitehat opened Oct 5, 2026·

  ·

• serve: an API key with spaces around it or outside Latin-1 works, and a refused key is not called a missing one (#725)

  #956·BGonnermann opened Oct 5, 2026·

  ·

• bench: community Arc Pro B65 Gen4 x16 IQ2_XS results

  #955·timnevits opened Oct 5, 2026·

  ·

• cpu: allow configuring expert-pool tasks per phase

  #949·Unmaple opened Oct 5, 2026·

  ·

• fix: avoid CPU doorbell waits in fully resident batch decoding

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

  ·

• launcher: pick a model and its settings in one window (experimental)

  #946·pipeob0 opened Oct 5, 2026Contributor·

  ·
  2

• setup: the low-RAM and KV-streaming decisions as functions

  #945·pipeob0 opened Oct 5, 2026Contributor·

  ·

• hip: gfx11 WMMA prompt attention and selection scorer, gfx1151 hipBLASLt table

  #944·nsbradley88 opened Oct 5, 2026·

  ·