All pull requests

Pull requests

• test: exercise uncached BPE paths in tokenizer parity checks

  #2238·mustafaerdinc opened Oct 13, 2026·

  ·

• bench: community report, RX 6900 XT (gfx1030) x1 and x2 on 0.1.43.1 - IQ3_S and UD-Q4_K_XL, stock and with the open gfx103x PRs

  #2237·Vop3p opened Oct 13, 2026Contributor·

  ·

• Batch decode: one launch for all rows in QSA attention and GDN recurrence (opt-in)

  #2234·yi-sudo opened Oct 13, 2026·

  ·
  1

• docs(bench): RTX 3060 12GB / 32GB RAM Coder IQ1_M with Budget12 64K profile

  #2233·wookie4lif3 opened Oct 13, 2026·

  ·

• HIP gfx1100: bit-exact decode and prompt-path speedups (+12% decode, KL 0)

  #2231·android-imdad opened Oct 13, 2026·

  ·

• [HIP] Keep mixed-gfx prefill GEMM modes per device (#690, #703)

  #2230·PJGV333 opened Oct 13, 2026·

  ·

• sycl: enable vision (--vision cpu) on intel

  #2229·rriggs opened Oct 13, 2026·

  ·

• STRATA_HC_PERSIST=1 (opt-in, HIP): the verifier's hyper-connection read as one persistent kernel (gfx1030: read 1.35-2.10x, per window -4.2 % / -2.8 %)

  #2226·Vop3p opened Oct 13, 2026Contributor·

  ·

• pcie: the measured PCIe share default (PcieAuto) on gfx1100 too — gate-measured on a native pack

  #2224·xyzzing opened Oct 13, 2026Contributor·

  ·

• verify: STRATA_NO_PCIE_GROUP=1 drops the empty PCIe-share stage with --pcie-frac 0

  #2223·rafatxf opened Oct 13, 2026Contributor·

  ·

• sm_75: fewer launches per decode layer: plan wait folded, fused quantize + interleave (bitwise)

  #2222·rafatxf opened Oct 13, 2026Contributor·

  ·

• sm_75: Q2_0 matvec with one warp per four rows (bitwise)

  #2221·rafatxf opened Oct 13, 2026Contributor·

  ·

• sm_75: QSA decode attention for INT8 KV off the shared-memory pipe (bitwise)

  #2220·rafatxf opened Oct 13, 2026Contributor·

  ·

• Count clean active file cache as reclaimable under a cgroup limit

  #2217·kuhlsnu opened Oct 13, 2026·

  ·

• bench: community report, one Arc A770 16 GB (xe) - main on the card and the A770 pull requests, same-day A/B

  #2216·aslater3 opened Oct 13, 2026Contributor·

  ·

• cpu: one token's IQ3_XXS/IQ3_S/IQ2_S gate/up rows in ggml's exact bits, faster (same output; STRATA_CPU_SINGLE_EXACT=0 keeps ggml's dot)

  #2215·aslater3 opened Oct 13, 2026Contributor·

  ·

• CUDA: keep CPU expert servicing active during instrumented graph launches

  #2209·CC-David-CC opened Oct 13, 2026Contributor·

  ·

• docs/STRIX_HALO.md: STRATA_PF_SWITCH_MIN_T compares the chunk, not the prompt (a 4,096-token prompt is a 4,089-token chunk); 2048 puts it on the switches

  #2208·libratechw opened Oct 13, 2026Contributor·

  ·

• sycl/A770: wide MMVQ kernels at 2-8 columns in large-GRF mode on Arc A-series (bit-identical; 1.3-7.4x kernel, decode +10-17%)

  #2207·aslater3 opened Oct 13, 2026Contributor·

  ·

• bench: community report, 2x RTX 3090 24 GB (NVLink), IQ3_S, 204K, eng…

  #2206·superlee-terry opened Oct 13, 2026·

  ·

• sycl: with --expert-parallel-device, card N computes its own experts of each prompt chunk

  #2205·recutita opened Oct 13, 2026Contributor·

  ·

• Monitor: token usage and cost per day, priced, with the folder's models merged or shown one by one

  #2204·pipti opened Oct 13, 2026·

  ·
  1

• sycl/A770: fused GR read with more, smaller work-groups on Arc A-series (same sums; decode +3-5%; on top of #2193)

  #2203·aslater3 opened Oct 13, 2026Contributor·

  ·

• --vram-elastic with --kv-grow: VRAM takes its chunks where the K/V grows into

  #2201·i-Vladimir-i opened Oct 13, 2026·

  ·

• sycl: prompt GDN recurrence with one barrier per token on Alchemist (bit-identical; A770 kernel 2.1x, prompt +3-7%)

  #2199·aslater3 opened Oct 13, 2026Contributor·

  ·