Issues

• [Feature]: retain model and conversation state while releasing GPU resources

  Indras-Mirror/Xolotl#74 · Filter by author NektoDron opened on Oct 30, 2026
  1

• Vulkan implementation for multi-brands setups

  Indras-Mirror/Xolotl#64 · Filter by author notGabeNewell opened on Oct 27, 2026
  1

• [Showcase / Feedback] RTX 5090 (Blackwell sm_120a) + CUDA 13.1 Native Build & 10.6M Token Production Benchmark (Avg 30 tok/s on Windows 11)

  Indras-Mirror/Xolotl#61 · Filter by author polygizer opened on Oct 27, 2026
  2

• V100 (sm_70), single card: DeepSeek-V4.1-1017 V4 decodes at ~9 tok/s, and the expert tier looks launch/sync-bound

  Indras-Mirror/Xolotl#60 · Filter by author metat-hub opened on Oct 26, 2026
  3

• Benchmark: 2x RTX 3090 (vast.ai) - Qwen3.8-Flash-Next IQ3_S + DeepSeek-V4.1-Flash-1017 V4, plus a CPU-pinning gotcha with two instances

  Indras-Mirror/Xolotl#47 · Filter by author alchemicwebstudio opened on Oct 25, 2026
  2

• Xolotl qwen4exp MXFP4 — 1M-context multi-needle report (2026-10-24)

  Indras-Mirror/Xolotl#44 · Filter by author biurquezjoselina-design opened on Oct 24, 2026
  1

• # issue: GSQ-RCO IQ2_XXS first GPU run — works, but decode stays on the MMVQ fallback (2.4 tok/s vs V3's 15.8)

  Indras-Mirror/Xolotl#40 · Filter by author biurquezjoselina-design opened on Oct 24, 2026
  3

• ds41_generate fails to build with STRATA_ENABLE_HIP=ON (cuda_stream, ds41_apply_mmq_cc_policy)

  Indras-Mirror/Xolotl#34 · Filter by author phly95 opened on Oct 23, 2026
  2

• AC922 (ppc64le, 6x V100-16GB, NVLink2): does the Buster sm_70 DS41 path need CPU expert kernels? (ppc64le support)

  Indras-Mirror/Xolotl#33 · Filter by author penink-WH opened on Oct 23, 2026
  15

• [Bug]: --remote-device on Windows: conversation pool OOM at load, sliced arena pin needs a cap of 48 GiB or less (auto = 97 GiB), then slower than one card

  Indras-Mirror/Xolotl#32 · Filter by author Xiaomi2008 opened on Oct 22, 2026
  3

• First report from a new machine: **RTX 5090 D v2 24 GB** (sm_120, 24427 MiB, PCIe x16 Gen5), i7-14700KF, 128 GB RAM (DDR5-4800), Windows, model + Engram on local NVMe (NTFS). Built from source today.

  Indras-Mirror/Xolotl#14 · Filter by author biurquezjoselina-design opened on Oct 19, 2026
  8

• Project Maya fork

  Indras-Mirror/Xolotl#12 · Filter by author merbanan opened on Oct 19, 2026
  1

• GLM 5.3 Flash performance optimization investigation findings

  Indras-Mirror/Xolotl#4 · Filter by author jazir555 opened on Oct 18, 2026
  1

• Many-core gen4 host: STRATA_TIER_BATCH_PCIE=0 (CPU pool tier) is ~+60% D4.1 serve decode — and --pcie never shapes the DSpark serve path

  Indras-Mirror/Xolotl#3 · Filter by author chimpera opened on Oct 18, 2026
  2

• Deepseek 4.1 Flash 1017 and multi gpu

  Indras-Mirror/Xolotl#2 · Filter by author sgmihai opened on Oct 17, 2026
  11