All pull requests

Pull requests

• [Other] calibrate: keep the PCIe/draft-floor winner only when it beats this run's own noise instead of the fixed >3%

  #3522·dsenchankau opened Oct 20, 2026·

  ·

• [Other] tests: stop the wizard-discovery test from reading the checkout's own run configs

  #3521·dsenchankau opened Oct 20, 2026·

  ·

• [Benchmark] name the AMD card and its ROCm runtime in the standard report

  #3519·l33tm4st3r opened Oct 20, 2026Contributor·

  ·
  1

• [Engine] conversation cache: the retention pin follows the system-prompt root, not chain item 0

  #3517·chaog992 opened Oct 20, 2026Contributor·

  ·

• [Benchmark] community report: 2x RTX 3090, Swift-1.5 Q6_K, Strata 2.1 (standard-v1)

  needs-format
  #3515·ada3216 opened Oct 20, 2026·

  ·
  2

• [Engine] tests: the trellis kernel units join the isa guard, and the TQ tests skip on an upstream ggml (#3409)

  #3511·chaog992 opened Oct 20, 2026Contributor·

  ·
  1

• [Engine] arena: attach a complete shared arena instead of re-copying (STRATA_ARENA_REUSE=1, opt-in)

  #3509·Jerry-Lee661 opened Oct 20, 2026·

  ·

• [Benchmark] community report: 3x Tesla P100 16 GB (sm_60), Flash-Next IQ3_S at 262,144 tokens

  #3506·maxlawrie opened Oct 20, 2026·

  ·

• [Benchmark] One R9700 with IQ2_XS: one client, two clients, and four levers that did not help

  #3505·Faks opened Oct 20, 2026·

  ·

• [Engine] add GET /v1/models/{id} so Visual Studio can use Strata

  #3500·MajMcCloud opened Oct 20, 2026·

  ·

• [Engine] hip: STRATA_SELECT_GRID=2 - the selection scorer's CTA order swaps for L2 (bitwise identical)

  #3499·vlupilin opened Oct 20, 2026·

  ·

• [Engine] hip: STRATA_SELECT_SPLIT2=1 - the selection scorer's hi/mid bf16 pieces (3 wmma instead of 6)

  #3498·vlupilin opened Oct 20, 2026·

  ·

• [Benchmark] 2x NVIDIA TITAN RTX / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1

  #3495·shrisha108 opened Oct 20, 2026Contributor·

  ·

• [Benchmark] Paired stats for bench comparisons: the test says, and a size note next to it

  #3494·victorgabr opened Oct 20, 2026·

  ·

• [Engine] serve: a slot's live view keeps the request's own clocks after it moves from the solo path (elapsed_s, tok_s)

  #3492·andyhrc opened Oct 20, 2026·

  ·

• [Engine] prefill: a padded activation is gfx11 only, and Gemm::native copies it unpadded instead of exiting

  #3484·John-Pi opened Oct 20, 2026·

  ·

• [Engine] stage buffers: fence a queued copy per stream, not per buffer (#1237)

  #3483·John-Pi opened Oct 20, 2026·

  ·

• [Engine] HIP: opt-in Q4_0-KV prompt attention on RDNA4 matrix cores (STRATA_HIP_WMMA_Q4)

  #3482·shenmintao opened Oct 20, 2026·

  ·

• [Benchmark] RX 9060 XT 16GB / Qwen3.8-Flash-Next / IQ2_XS / strata-qwen38-thinking-v1-draft.1 (custom track)

  #3481·shenmintao opened Oct 20, 2026·

  ·

• [Engine] HIP: verify-window routing for non-native routers (256 experts) on every HIP card

  #3476·pugolini opened Oct 20, 2026Contributor·

  ·

• [Benchmark] RX 9070 XT / 64 GB: IQ3 and Q4 performance and coding results

  #3475·visualizedigital opened Oct 20, 2026·

  ·

• [Other] setup: FP4-Local conversion no longer dies at step 6 on Python 3.15 (torch pin 2.9.1 -> 2.14.1)

  #3474·beplee opened Oct 20, 2026·

  ·

• [Engine] Zen 1/2: build ggml-cpu without BMI2 pdep (IQ1_S/IQ1_M CPU dots 16-19x faster, same results)

  #3472·Licenser opened Oct 20, 2026·

  ·
  2

• [Other] tools: the busbar gateway, one OpenAI address in front of several providers

  #3466·Buddychek opened Oct 20, 2026·

  ·

• [Engine] Metal: opt-in --dflash DFlash2 drafter in place of the MTP layer

  #3457·hotocoo opened Oct 20, 2026·

  ·