All pull requests

Pull requests

• Muse Spark: opt-in streaming reference runtime and CPU/CUDA parity

  #2769·BColsey opened Oct 16, 2026·

  ·

• Experimental gfx803 (RX 580) Q36 Linux path

  #2765·minkiboo opened Oct 16, 2026·4 of 6 tasks·

  ·

• Community benchmark: Qwen3.8-Flash-Next Coder IQ1_M on a 5070ti laptop (12GB) GPU + 32 GB RAM

  #2763·Aoyi-Feng opened Oct 16, 2026·

  ·

• what only the decode needs leaves VRAM while a prompt runs, and the prestage ring takes the loan (WIP - porting mw00/project-maya#223)

  #2761·gopinath87607 opened Oct 16, 2026Contributor·

  ·

• Add a single-file chat HUD

  #2759·matrix9neonebuchadnezzar2199-sketch opened Oct 16, 2026·

  ·

• sycl: STRATA_QFUSE - commit one-token rounds (the commit skipped them, so QFUSE gave wrong output)

  #2758·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl: STRATA_RESIDENT_TOUCH=1 - re-touch the pinned resident experts after reclaim (xe shrinker backs them up to shmem)

  #2757·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl/A770 prefill: QSA attention 128 queries a launch on Alchemist (bitwise); parity test hashes the FP32 kernel

  #2756·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl/A770: commit all GDN layers' conv history + state replay in two launches (STRATA_GDN_COMMIT_BATCH; bitwise)

  #2755·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl/A770: MTP shared-expert branch off by default on Alchemist (a non-linear draft graph resets ~50 events a replay)

  #2754·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl: Q4_K/Q5_K wide GEMV - pass dp4a's ones word as a value IGC cannot fold (it expanded each dp4a into ~24 ops)

  #2753·aslater3 opened Oct 16, 2026Contributor·

  ·

• sycl: Q6_K GEMV - subtract 32 from the 6-bit values with three integer ops instead of a saturating byte subtract

  #2752·aslater3 opened Oct 16, 2026Contributor·

  ·

• cpu pool: STRATA_POOL_FUSED=1 - a native layer's gate/up and down without the barrier between them; pool_layer_bench

  #2751·aslater3 opened Oct 16, 2026Contributor·

  ·

• kq_cpu_bench: quantize i-quant blobs with ggml_quantize_chunk (it crashed on them); KQ_PHASE=gu|down

  #2750·aslater3 opened Oct 16, 2026Contributor·

  ·

• Add community benchmark: 4x AMD Radeon gfx1201, EPYC 7452, UD-Q5_K_XL

  #2749·TechwriterSSchmidt opened Oct 16, 2026·2 tasks done·

  ·
  3

• Community benchmark: Apple M5 MacBook Air, 32 GB, Q2_0 (Metal, experts streamed)

  #2748·yourfate opened Oct 16, 2026·

  ·

• sycl: use packed word expansion for Q2_0 expert down on Battlemage

  #2746·xiangyuT opened Oct 16, 2026·

  ·

• Qwen3.6: opt-in int8 K/V cache and single staging buffer

  #2742·ggeorgovassilis opened Oct 16, 2026·

  ·

• Add a Nemotron-H MoE engine path (Nemotron 3.5 Super VL 120B-A12B)

  #2740·Rhonstin opened Oct 16, 2026Contributor·

  ·

• bench: mixed RTX 3090, RTX 3060 and dual V100 Q4_K_M community report

  #2739·techrian opened Oct 16, 2026·

  ·

• batch-mtp: split the solo MTP budget across batch slots — multi-GPU multi-stream tuning (agent concurrency)

  #2738·superlee-terry opened Oct 16, 2026·

  ·

• bench: community report - GTX 1080 Ti (sm_61): eight-knob screen, long context 32K-250K with needles, 1h soak

  #2737·MrKop-git opened Oct 16, 2026·

  ·

• STRATA_SLOT_COPY_DIRECT=1 (opt-in, CUDA): main session <-> batch slot copies go state to state, without a host image

  #2728·andyhrc opened Oct 16, 2026·

  ·

• tools: add opt-in FP4-Local CSF checkpoint conversion

  #2727·hagope opened Oct 16, 2026Contributor·

  ·

• bench: RX 7900 XT (Windows HIP) community results + fix BENCH.bat on GBK Windows

  #2725·abddba1313-art opened Oct 16, 2026·

  ·