All pull requests

Pull requests

• [Engine] dsv41: build ds4_gguf_cpu.cpp's AVX-VNNI tier only where CMake's check passes (GCC 11 + binutils 2.35 on RHEL/Alma/Rocky 9 stopped at vpdpbusd)

  #4970·yazun opened Nov 2, 2026·

  ·

• [Engine] dsv41: the chunked prompt pass DMAs a pinned expert copy from every NUMA part

  #4969·yazun opened Nov 2, 2026·

  ·
  1

• [GUI] Monitor: name the model, add lifetime averages, and the speed min and max

  #4968·noahark opened Nov 2, 2026Contributor·

  ·

• [Engine] serve: warn when systemd-oomd is running, and name it when the engine is ended by signal 9

  #4967·cakescats opened Nov 2, 2026·

  ·

• [Engine] Opt-in solo/batch MTP switching with physical single-GPU controls

  needs-format
  #4966·cennac opened Nov 2, 2026·

  ·
  1

• benchmark: RTX 5070 Ti 16 GB + Core Ultra 9 285K / Qwen3.8-Flash-Next NVFP4-NVFP4-BF16

  needs-format
  #4963·utombords opened Nov 1, 2026·

  ·
  1

• [Engine] Metal batch slots: each slot rotates by its own M-RoPE table (pictures beside other requests)

  #4961·oceanio opened Nov 1, 2026·

  ·
  1

• [Engine] Metal batch slots: the slot window beats the serve watchdog

  #4960·oceanio opened Nov 1, 2026·

  ·

• [Benchmark] Intel Arc Pro B70 32 GB / Qwen3.8-Flash-Next Coder / IQ1_M / strata-qwen38-thinking-v1-draft.1

  needs-format
  #4957·rwkeyes opened Nov 1, 2026Contributor·

  ·
  1

• [GUI] Plan mode: a plan card for every Plan-mode turn, including the requirements and the steps

  #4956·MohSakhaei opened Nov 1, 2026·

  ·

• [Engine] Windows: disable EcoQoS background throttling for engine and benchmark processes

  #4952·utombords opened Nov 1, 2026·

  ·
  2

• [Engine] CPU: add AVX-VNNI vpdpbusd path for 256-bit NVFP4 expert rows

  #4951·utombords opened Nov 1, 2026·

  ·
  2

• Monitor: --monitor-aggregate shows speed and prefill summed over every stream

  needs-format
  #4947·elight opened Nov 1, 2026·

  ·
  1

• [Engine] Table-lookup Q2_0 rows for Sandy / Ivy Bridge (STRATA_Q2_LUT=1, opt-in)

  #4944·hiyazakite opened Nov 1, 2026·

  ·

• [Engine] Interleaved MMVQ (native_mmvq_il) for IQ3_S/IQ3_XXS and HIP/gfx12 (1202) tables

  #4942·bsorensen110 opened Nov 1, 2026Contributor·

  ·
  5

• [Benchmark] RTX 5060 Ti / Flash-Next IQ3_S 262K: first standard-v1 run on this card under native Windows

  #4940·stoeggich opened Nov 1, 2026·

  ·

• [Engine] reasoning_loop_recovery "wrap" and "nudge": end a reasoning loop with an answer, not an empty reply

  #4939·nomishbhardwaj opened Nov 1, 2026·

  ·

• [Engine] Update: DFlash2 batched multi-slot drafting, co-draft body/head, and concurrent prefill fixes

  #4937·bsorensen110 opened Nov 1, 2026Contributor·

  ·
  13

• [Other] serve: one unknown key in .shared-settings.json no longer clears the rest (#4933)

  #4934·dan64 opened Nov 1, 2026·

  ·
  1

• [Other] serve: preserve_thinking / preserve_reasoning reach the chat template

  #4932·dan64 opened Nov 1, 2026·

  ·
  1

• [Engine] STRATA_LOOKUP (opt-in): long prompt-lookup windows for a request that copies its own context

  #4928·lyesrock opened Nov 1, 2026Contributor·

  ·

• benchmark: RTX 5060 Ti / Qwen3.8-Flash-Next / UD-Q4_K_XL / strata-qwen38-thinking-v1-draft.1

  needs-format
  #4924·xss1992 opened Nov 1, 2026·

  ·
  1

• [Benchmark] Community report: M2 Ultra 64 GB headless — expert arena cap 41.2 -> 49.6 GB, +15-24%

  #4922·yoffy opened Nov 1, 2026·

  ·

• [Benchmark] GTX 1070 Ti 8 GB (Pascal sm_61) / Flash-Next + Swift 1.5 / IQ2_XS / strata-qwen38-thinking-v1-draft.1

  #4920·jaredrsommer opened Nov 1, 2026·

  ·
  2

• [Benchmark] 2x RX 7900 XT (gfx1100) on Windows / Qwen3.8-Flash-Next GSQ-RCO IQ3_S / community suite (custom flags)

  #4918·ozoooooo001 opened Nov 1, 2026·

  ·