All pull requests

Pull requests

• Bound Qwen3.6 Windows CUDA placement by dedicated VRAM headroom

  #3128·kbc20f11 opened Oct 19, 2026·

  ·

• test+guard: a call for a tool the request never offered stays text (#3075)

  #3126·ctr-137 opened Oct 19, 2026·

  ·

• Metal backend: fix the build on macOS 27.2 / Apple clang 21 and make fused_gr_persist_init a no-op

  #3122·mnott opened Oct 19, 2026·

  ·

• conversation cache: opt-in overwrite growth for snapshot buffers

  #3121·JumpEggKiller opened Oct 19, 2026·

  ·

• Community benchmark: RTX 5070 Ti + Core Ultra 7 270K Plus, IQ3_S at 262K on 0.1.45 - speed, and the quant yardstick against Q8_0

  #3120·Rui-Shi opened Oct 19, 2026·

  ·

• serve: temperature 0 stays greedy - no model-card penalties on a greedy request (#2940)

  #3119·christopherrobertbrooks-tech opened Oct 19, 2026Contributor·

  ·
  1

• fix(serve): a cached image whose .sve was deleted is encoded again; a 500 instead of a dropped connection

  #3118·kai-chop opened Oct 19, 2026·

  ·

• gpustat: the Intel VRAM readings on a host with two cards

  #3117·beyondcr opened Oct 19, 2026·

  ·

• Optimize strata m5 max metal rope scaling 3378619349143257926

  #3114·WillkTech opened Oct 19, 2026·

  ·
  1

• bench: RTX 4090 reports for 0.1.45 (IQ3_XXS + IQ3_S, up to the 204800 limit): STRATA_DEQUANT_WIDE regression at 4K, isolated by A/B; full test suite

  #3109·Dmitry-B opened Oct 18, 2026Contributor·

  ·

• perf(hip): optimize exact QSA block scoring and top-k selection on gfx12

  #3108·bsorensen110 opened Oct 18, 2026Contributor·

  ·
  5

• Add request-scoped prefill VRAM loans

  #3105·c3c opened Oct 18, 2026Contributor·

  ·

• Extend native-IQ prefill dispatch to SM75/Turing

  #3104·c3c opened Oct 18, 2026Contributor·

  ·

• batch-groups: STRATA_SPLIT_MTP_BATCH beside STRATA_PIPED_CONCURRENT (opt-in STRATA_PIPED_CONCURRENT_SMB=1)

  #3103·Vop3p opened Oct 18, 2026Contributor·

  ·

• perf(hip): enable wave32 QSA decode attention and MMVQ row reuse on gfx12

  #3102·bsorensen110 opened Oct 18, 2026Contributor·

  ·
  4

• Community report: Apple M1 Max 64 GB (Metal streaming) — tuning matrix neutral

  #3100·DarkJoll opened Oct 18, 2026·

  ·

• fix: strata_bench defines the port it prints (#3096)

  #3098·ctr-137 opened Oct 18, 2026·

  ·
  1

  1

• batch: STRATA_PW_WITH_BATCH=1 keeps --pipeline-windows for a lone request beside --batch slots (+20% single-stream decode on a 2-card split, opt-in)

  #3095·Vop3p opened Oct 18, 2026Contributor·

  ·

• Gfx906 exact pipeline j48 20261018

  #3094·0FL01 opened Oct 18, 2026Contributor·

  ·

• fix: strata_bench crashed reading CMakeLists.txt on GBK-default Windows

  #3092·gg320324492-lgtm opened Oct 18, 2026·

  ·

• monitor: show every parallel request + fixed-height slot dots (#465)

  #3091·superlee-terry opened Oct 18, 2026·

  ·

• bench: community report, Radeon VII 16 GB (gfx906), Qwen3.6-35B-A3B UD-Q4_K_XL

  #3089·Kaidora opened Oct 18, 2026·

  ·

• expert source: batch-prefetch a layer's CPU experts for a plain mapped experts.bin (STRATA_MMAP_PREFETCH, #3065)

  #3086·As7656431 opened Oct 18, 2026·

  ·

• Enable IQ prefill path on sm_75 (RTX 20-series / Turing)

  #3085·dd777mine opened Oct 18, 2026·

  ·
  3

• qwen35: FP16 K/V cache by default, int8 with --kv-q8 (262144 fits a 16 GiB card)

  #3084·benthic-io opened Oct 18, 2026·

  ·