All pull requests

Pull requests

• docs: visual community RTX PRO 6000 concurrency results (64K/128K)

  #1104·CC-David-CC opened Oct 6, 2026Contributor·

  ·

• fix: batch host paths honor refreshed residency (all_resident_ latch vs ar_on())

  #1102·win10ogod opened Oct 6, 2026·

  ·

• prefill: let Stager threads sleep instead of yield-spinning (Linux + --mmap-experts: prompts ~10x faster on DGX Spark, ~1 core instead of ~20)

  #1101·uncle-daddy-jp opened Oct 6, 2026·

  ·
  1

• s2_qpn8: opt-in QPN8 expert GEMV on Volta m8n8k4 (M >= 4)

  #1100·ATIVX928 opened Oct 6, 2026·

  ·

• layer split: the hand-off over peer access on NVLink (3.3 -> 48 GB/s, +0.9% e2e)

  #1099·ATIVX928 opened Oct 6, 2026·

  ·

• prefill/gdn: the Volta prompt recurrence chain-split (+7-15%) and the opt-in chunked path

  #1098·ATIVX928 opened Oct 6, 2026·

  ·

• V100: bit-exact GEMV / MoE operator fusions (experimental build only)

  #1097·ATIVX928 opened Oct 6, 2026·

  ·

• qsa_prompt_attn: the Volta kernel's latency pipeline (32K int8 12.1 -> 10.2 ms, +1.2% e2e)

  #1096·ATIVX928 opened Oct 6, 2026·

  ·

• kv_q8: the gather decodes 8 int8 per thread (V100 experimental build, +11%)

  #1095·ATIVX928 opened Oct 6, 2026·

  ·

• Monitor: Load/Unload split button + idle-unload slider, load busy guard, honest unload

  #1093·miskahm opened Oct 6, 2026Contributor·

  ·

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

  #1092·pipeob0 opened Oct 6, 2026Contributor·

  ·

• qsa_select: FP32 tiled block scores below sm_80 (+22% 128K / +44% 248K prefill on a 2080 Ti)

  #1091·konijiwa110 opened Oct 6, 2026Contributor·

  ·

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

  #1090·konijiwa110 opened Oct 6, 2026Contributor·

  ·

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

  #1089·ZhongUncle opened Oct 6, 2026·

  ·

• cpu: allow configuring expert-pool tasks per phase

  #1087·Unmaple opened Oct 6, 2026·

  ·

• strata-vision: flash attention off when the CPU encoder's ggml has AVX-512 (3.3x faster there)

  #1086·shefowl opened Oct 6, 2026·

  ·

• hip/gfx906: __dp4a falls back to a 4-iteration loop although gfx906 has v_dot4_i32_i8

  #1084·xxDoman opened Oct 6, 2026Contributor·

  ·

• hip/gfx906: the tree does not build (three STRATA_USE_HIP gates miss STRATA_HIP_GFX906)

  #1083·xxDoman opened Oct 6, 2026Contributor·

  ·

• fix(build): qualify size_t in vmm.hpp (GCC 12 rejects the unqualified name)

  #1081·aleesposito85 opened Oct 6, 2026·

  ·

• bench: RX 6800M (gfx1031) community report - self-built HIP engine on Windows

  #1078·Hugua700 opened Oct 6, 2026·

  ·