All pull requests

Pull requests

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

  needs-format
  #3233·bsorensen110 opened Oct 19, 2026Contributor·

  ·
  1

• [Benchmark] 3x RTX 2080 Ti 22 GB (sm_75): v0.1.45 vs v0.2.0-beta, UD-Q4_K_XL at 262K, layer split 16,32

  #3232·Adamyno opened Oct 19, 2026·

  ·

• macOS: conversation cache parks again (Metal checkpoint validation, memory guard)

  needs-format
  #3231·mnott opened Oct 19, 2026·

  ·
  1

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

  #3230·bsorensen110 opened Oct 19, 2026Contributor·

  ·

• [Engine] loop guard: catch a young sentence loop, splice medium effort on recovery

  #3229·Yunado opened Oct 19, 2026Contributor·

  ·

• [Other] setup: install the FP4-Local converter's packages before a prepare

  #3228·Yunado opened Oct 19, 2026Contributor·

  ·
  1

• [Engine] perf(hip): optimize GDN verify kernels with LDS-only barriers and enable batched commit under STRATA_GDN_SPLIT=1

  #3226·bsorensen110 opened Oct 19, 2026Contributor·

  ·
  1

• [Engine] perf(hip): enable fused_gr tsum on gfx12, fix q8_inject guard, and use LDS-only barriers

  needs-format
  #3225·bsorensen110 opened Oct 19, 2026Contributor·

  ·
  3

• [Engine] perf(hip): tune RDNA4 QSA prompt WMMA path (STRATA_HIP_WMMA stays opt-in)

  #3224·bsorensen110 opened Oct 19, 2026Contributor·6 of 7 tasks·

  ·
  2

• [Engine] perf(prefill): port wmma_gemm and gdn_chunked WMMA kernels to RDNA4 gfx12

  #3223·bsorensen110 opened Oct 19, 2026Contributor·4 of 5 tasks·

  ·
  1

• [Engine] perf(prefill): use single LDS-only barrier per stage in native_w11_kernel

  #3222·bsorensen110 opened Oct 19, 2026Contributor·3 of 4 tasks·

  ·
  1

• [Engine] perf(hip): vectorize HC_Q8 prefill dequant and enable gfx12 Q8_0 gr_upmix

  #3220·bsorensen110 opened Oct 19, 2026Contributor·4 of 5 tasks·

  ·
  3

• [Other] serve: support IPv6 bind addresses

  #3212·purrsec opened Oct 19, 2026·

  ·

• [Benchmark] RX 7900 GRE (gfx1100, 16 GB): IQ2_XS production serving measurements at 262K context

  #3210·lsunay opened Oct 19, 2026·

  ·

• [Engine] a reply that ends inside its thinking is continued and reported (#1814)

  #3207·47Hunter47 opened Oct 19, 2026·

  ·
  2

• [Benchmark] calibrate checks the output directory up front (#3187)

  #3205·ctr-137 opened Oct 19, 2026·

  ·
  1

• [Engine] OpenAI logprobs / top_logprobs / allowed_token_ids for the CUDA/HIP engine

  #3200·xiezhipeng-git opened Oct 19, 2026·

  ·

• [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