All pull requests

Pull requests

• [Engine] serve: per-layer expert routing metrics for /metrics ("routing_counts", opt-in)

  #3194·blange48 opened Oct 19, 2026Contributor·

  ·

• [Benchmark] RTX 5090: engine 0.1.45 vs 0.1.45.3, prompt reading +2.7 to +9.8 %

  #3189·IVoyt opened Oct 19, 2026·

  ·

• [Engine] arena: hand back the block for layers a GPU holds whole (opt-in, saves 21.7 GiB of RAM)

  needs-format
  #3185·weituotian opened Oct 19, 2026·

  ·
  1

• [Other] setup: say that --kv, --context and --vision are not applied when starting an installed model

  #3182·victoralcazardev opened Oct 19, 2026·

  ·

• [Other] setup: pass the family's pack args to the low-RAM experts.bin run (Gyro needs --compat-bf16)

  #3181·victoralcazardev opened Oct 19, 2026·

  ·

• Web app: use wide windows and fit narrow screens

  #3176·mnott opened Oct 19, 2026·

  ·

• [Benchmark] 2x MI50 16 GB (gfx906): IQ3_XXS at 128K, --pipeline-windows 2 gives +36% decode

  #3175·romeo5353535353-ui opened Oct 19, 2026·

  ·

• serve: start an elastic image encoder before encoding a tool-returned image (#3171)

  #3174·Lithivm opened Oct 19, 2026·

  ·
  1

• [Engine] kv: size the resident ring from the run's depth (opt-in, STRATA_KV_RING_ADAPT)

  #3168·xyzzing opened Oct 19, 2026Contributor·

  ·
  3

• serve: accept the three thinking-budget field names (pi's thinkingTokenBudgetField) as aliases

  #3167·shizhx opened Oct 19, 2026·

  ·

• hip: STRATA_SELECT_GRID=2 - the selection scorer's CTA order swaps for L2 (bitwise identical)

  #3164·vlupilin opened Oct 19, 2026·

  ·

• hip: STRATA_SELECT_SPLIT2=1 - the selection scorer's hi/mid bf16 pieces (3 wmma instead of 6)

  #3163·vlupilin opened Oct 19, 2026·

  ·

• tests: three suite gaps a live PC showed (no engine, setup, serve, or docs change)

  #3162·Buddychek opened Oct 19, 2026·

  ·

• [GUI] web: the Monitor headlines the PCIe link's current generation (#2137)

  #3158·ctr-137 opened Oct 19, 2026·

  ·

• [GUI] web: a tool call no longer widens the page on phones (#2971)

  #3156·ctr-137 opened Oct 19, 2026·

  ·

• gentoo-setup: setup.py builds the CUDA and HIP engines on Gentoo

  #3154·arabek opened Oct 19, 2026·

  ·
  1

• [Benchmark] [SYCL][Perf] Qwen3.8-Flash-Next IQ2_XS single-B70 E2E preview

  #3153·xiangyuT opened Oct 19, 2026·

  ·

• qsa: the default block-scores kernel without #2234's per-query keys (a gfx1151 verify window stalled on v0.1.45); the batch rows get their own kernel

  #3151·gt50 opened Oct 19, 2026Contributor·

  ·

• bench: 4x Tesla V100-SXM2 reports for 0.1.45 (IQ3_S, 262144): prompt reading unchanged, the 6-7% ru gap is measurement order, full test suite

  #3147·Dmitry-B opened Oct 19, 2026Contributor·

  ·

• report: RX 6700 XT (gfx1031 running as gfx1030) - built-in suite in two arms; STRATA_HIP_PROMPT_F16=1 reads prompts 1.76x-1.99x faster, decode unchanged

  #3146·ljmbot opened Oct 19, 2026·

  ·

• fix(HIP): pair gfx1031 with gfx1030 and run it as 10.3.0 - ROCm 7.1.1's rocBLAS ships no gfx1031 kernels

  #3145·ljmbot opened Oct 19, 2026·

  ·

• Dashboard: show the loaded model in the top bar and on the Model state card

  #3142·vrami-dev opened Oct 19, 2026·

  ·

• [GUI] web: retry a failed GET /mcp instead of hiding the MCP tools for the session (#3113)

  #3141·ctr-137 opened Oct 19, 2026·

  ·

• [Engine] serve/mcp: a cut-off HTTP answer is an error, not a TypeError on the waiting side

  #3139·ctr-137 opened Oct 19, 2026·

  ·
  1

• [Engine] serve: read llama.cpp's n_predict, including the prompt-only 0 (#2980)

  #3138·ctr-137 opened Oct 19, 2026·

  ·