All pull requests

Pull requests

• [Engine] Metal: fix garbage output with Unsloth low-bit files (BF16 ple_conv1d read as F16)

  #5387·qChauncey opened Nov 6, 2026·

  ·

• [Benchmark] RX 7900 XTX / Qwen3.8-Flash-Next / IQ3_XXS / strata-qwen38-thinking-v1-draft.1

  #5386·GoldenLeafBird opened Nov 6, 2026Contributor·

  ·
  2

• [Benchmark] RTX 4080 SUPER 16 GiB, i9-14900KF, 32 GiB RAM: Swift 1.5 IQ2_XS at 128K

  #5384·diegoarodriguezr opened Nov 6, 2026·

  ·

• [Engine] dsv41: apply a request's logit_bias in the DeepSeek-V4.1 engine

  #5382·srkim-business opened Nov 6, 2026·

  ·

• [GUI] show saved KV format for packaged Qwen3.8 models

  #5374·zapabob opened Nov 6, 2026Contributor·

  ·

• [TUI] keep the values the Settings view saved across launches; --set idle-unload writes its config key instead of an engine flag

  #5371·shikaku2 opened Nov 6, 2026·

  ·
  1

• [Benchmark] Apple M2 Ultra / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1

  #5368·5ugAv opened Nov 6, 2026·

  ·
  2

• benchmark: 2x Tesla V100-SXM2-16GB (PCIe carriers) / Huihui abliterated-v2 / UD-Q4_K_XL / strata-qwen38-thinking-v1-draft.1

  needs-format
  #5367·PoisonousHumanMineral opened Nov 6, 2026·

  ·
  1

• [Engine] Keep concurrent request accounting and prompt progress correct

  #5364·cennac opened Nov 6, 2026·

  ·

• [Engine] gfx906 (MI50/MI60/Radeon VII) performance branch — fastest known 4xMI50 Flash-Next numbers

  #5361·rpcribari opened Nov 6, 2026·

  ·
  2

• [GUI] recover chat revisions after an index write failure

  #5359·zapabob opened Nov 5, 2026Contributor·

  ·

• [Engine] DeepSeek: accept an MXFP4 down beside a low-bit gate / up (Unsloth UD-Q2_K_XL)

  #5357·R-Okauchi opened Nov 5, 2026·

  ·

• [GUI] xe sysfs reader in the shared telemetry so the SYCL container's Monitor shows temp, power and PCIe (#5353)

  #5356·james-martinez opened Nov 5, 2026·

  ·
  1

• [Benchmark] RTX 5060 Ti / Flash-Next IQ3_S and IQ3_XXS 262K: three standard-v1 runs

  #5352·stoeggich opened Nov 5, 2026·

  ·
  2

• Codex/ascend910 npu monitor live

  needs-format
  #5350·tyloryang opened Nov 5, 2026·

  ·
  1

• feat: native Flint 40B MoE backend with Batched TC GEMM & 256K 3-Tier KV Ring Buffer

  needs-format
  #5347·Andgihat opened Nov 5, 2026·

  ·
  1

• [Other] serve: preserve multi-turn reasoning by default and skip empty historical think blocks

  #5346·bsorensen110 opened Nov 5, 2026Contributor·3 of 4 tasks·

  ·
  4

• [Benchmark] 0.2.11 UD-IQ4_XS-v31 vs UD-IQ4_XS on two mismatched Blackwell cards

  #5345·jkenner111 opened Nov 5, 2026·

  ·

• [Engine] add fractional total-VRAM ceiling (--vram-frac) and setup scaling

  #5344·atomicmilkshake opened Nov 5, 2026·

  ·
  2

• [Benchmark] Tesla V100 SXM2 32GB / Qwen3.8-Flash-Next / GSQ-RCO IQ3_XXS / strata-qwen38-thinking-v1-draft.1

  #5343·ildargiv opened Nov 5, 2026·

  ·

• [Engine] credit a reserved huge-page pool when the expert arena weighs the RAM (#5317)

  #5341·qdotme opened Nov 5, 2026·

  ·

• [GUI] Monitor: per-card tiles for a split run (GPU load, VRAM, temp, Power, PCIe)

  #5339·ilfarpro opened Nov 5, 2026·

  ·

• [Benchmark] 3x AMD Instinct MI100, MTP-only spec 3, Qwen3.8-Flash-Next Q4_K_S

  #5338·nsaxploit opened Nov 5, 2026·

  ·
  4

• [Benchmark] Quadro P5000 + 2x Xeon X5650 (no AVX): IQ3_S vs Q2_0, standard-v1

  #5332·modulismm opened Nov 5, 2026·

  ·

• [Benchmark] RTX 5070 DeepSeek V4.1 Flash Huihui speed and recall results

  #5330·SiNaPsEr0x opened Nov 5, 2026·

  ·
  2