Issues

• 0.1.42 -> 0.1.43: decode -3.4% on RTX 4090 48 GB, IQ3_S, --resident-experts (MTP accept unchanged, prefill flat)

  Niko1221/Strata#2074 · Filter by author Tlyer233 opened on Oct 12, 2026

• [Feature]: GLM-5.3-Flash (glm5next) support — the glm53f fork proves it works in Strata; AMD/HIP would make it unique

  Niko1221/Strata#2067 · Filter by author artfix opened on Oct 12, 2026

• Engine hangs at layer 48 during decode when VRAM headroom drops below ~500 MiB

  Niko1221/Strata#2065 · Filter by author bang-imvt opened on Oct 12, 2026

• strata_bench.py mislabels mixed-GPU hosts and single-GPU runs on multi-GPU hosts

  Niko1221/Strata#2062 · Filter by author MinhHaDuong opened on Oct 12, 2026

• # Community benchmark on 890M - Minisforum X1 Pro Ryzen AI 9 HX 470/96GB/Linux

  Niko1221/Strata#2061 · Filter by author htbmixbox opened on Oct 12, 2026

• gfx1030 / Windows: prompts read 16–20% slower since 0.1.40.2 — bisected to ac52c2c7 (#1240)

  Niko1221/Strata#2055 · Filter by author did-technomancer opened on Oct 12, 2026

• update.sh: failed engine build leaves BUILD.json stamped with new source hash — engine never rebuilt (stale binary reports old version)

  Niko1221/Strata#2054 · Filter by author MrKop-git opened on Oct 12, 2026
  1

• Default multi-GPU recommendation not ideal for asymmetrical GPUs

  Niko1221/Strata#2052 · Filter by author MaKla89 opened on Oct 12, 2026

• t

  Niko1221/Strata#2051 · Filter by author Sahil0015-btech opened on Oct 12, 2026

• Deterministic deadlock at the completion handshake after the last block: host at layer step 48 with the GPU already rang - 5/5 stalls identical, on 0.1.41 and 0.1.43 (RTX 4080 / Windows / IQ3_XXS native pack / 256K)

  Niko1221/Strata#2049 · Filter by author gaoyonghui123 opened on Oct 12, 2026

• 2026-10-12-community-rtx-3060-qwen3-8-flash-next-iq3-s

  Niko1221/Strata#2046 · Filter by author Tee8088 opened on Oct 12, 2026

• Community Bench: 2026-10-12-community-2x-2080ti-22gb-qwen3-8-flash-next-gsq-rco-iq3 (Proxmox 9 VM)

  Niko1221/Strata#2044 · Filter by author sense1024 opened on Oct 12, 2026

• [Linux, SATA SSD] --ple-io boot auto-switch to mmap degrades PLE prefill ~7x: the 15K rows/s "direct is healthy" threshold is calibrated for NVMe, so it trips on ordinary SATA SSDs where direct is actually the faster mode

  Niko1221/Strata#2040 · Filter by author lanceyang093-blip opened on Oct 12, 2026
  1

• [0.1.43] CUDA layer split (2× RTX 3090): decode windows lose async launch, solo 131.4 → 39–76 tok/s, prefill −35%

  Niko1221/Strata#2038 · Filter by author lijiajun1997 opened on Oct 12, 2026

• serve: tool_call_recovery misses fenced JSON inside an explicit tool-call wrapper

  Niko1221/Strata#2037 · Filter by author bettercallcaleb opened on Oct 12, 2026

• [Linux] --calibrate combined with --model silently falls through to plain re-setup (setup.py:4958)

  Niko1221/Strata#2036 · Filter by author satwareAG-ironMike opened on Oct 12, 2026

• [Feature]: Optional Windows Tray for local server status and diagnostics

  Niko1221/Strata#2034 · Filter by author medking82 opened on Oct 12, 2026

• [Enhancement] Mark conversation cache MADV_COLD to use swap preferentially on Linux

  Niko1221/Strata#2033 · Filter by author Interpause opened on Oct 12, 2026
  1

• E5 2673v3 64gDDR31866 AMD MI50 32G成功跑通Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S

  Niko1221/Strata#2029 · Filter by author fude365 opened on Oct 12, 2026

• [Bug]: STRATA_SPEC_PROB: --spec-min-p gates on the drawn draft's q, which skews sampled tokens toward the draft head's favourite

  Niko1221/Strata#2028 · Filter by author KarlGrier opened on Oct 12, 2026

• [Bug]: CMake: four native-expert tests are registered only from the second configure (STRATA_NATIVE_EXPERTS is read before its option())

  Niko1221/Strata#2027 · Filter by author KarlGrier opened on Oct 12, 2026

• Strata v0.1.43 returns an empty final response when a specific string appears in the model's thinking output

  Niko1221/Strata#2026 · Filter by author kntroh opened on Oct 12, 2026
  6

• [Bug]: gfx1151: a 4K prompt stays under STRATA_PF_SWITCH_MIN_T=4096 (its chunk is 4,095 tokens) and is ~16% slower

  Niko1221/Strata#2021 · Filter by author libratechw opened on Oct 12, 2026

• SYCL resident RAM mode: resident_stage_swaps stages the exchange without copying the evicted expert back

  Niko1221/Strata#2014 · Filter by author kuhlsnu opened on Oct 12, 2026

• [Bug]: Windows Strix Halo (gfx1151): prebuilt HIP engine never becomes ready, sticky "unspecified launch failure" during serve init

  Niko1221/Strata#2013 · Filter by author 400msstack opened on Oct 12, 2026