Issues

• [Engine] V0.1.58 Flash Next UD-IQ4_XS fails to locate GGUF shard in Models directory

  Niko1221/Strata#5499 · Filter by author phoenix-thedev opened on Nov 7, 2026
  2

• [Engine] split engine 0.2.11: the prompt-path auto pick counts the arena's unallocated expert slots as free VRAM, so 3x RTX 3090 24 GB fails to start ("VRAM short" -> "pool 3 has 271 slots, a window needs 320")

  Niko1221/Strata#5496 · Filter by author chaerle opened on Nov 7, 2026

• [App], [Engine] Strata can't find the correct engine

  needs-format
  Niko1221/Strata#5491 · Filter by author Retina-sys opened on Nov 7, 2026
  3

• [Engine] Decode throughput drops after several requests; reloading restores 100+ tok/s even with 300K+ context

  Niko1221/Strata#5490 · Filter by author PPicku opened on Nov 7, 2026

• [GUI] error message do not clear

  Niko1221/Strata#5489 · Filter by author Gazoo opened on Nov 7, 2026

• [Engine] dsv41 on Apple-silicon Metal: routed experts go to the CPU lane, the one-GPU decode path turns itself off, DeepSeek-V4.1-Flash runs 10-40x slower than expected

  Niko1221/Strata#5488 · Filter by author Qfbxz opened on Nov 7, 2026

• [Engine] Rope yarn 4x anchored at native + 1M capacity: GPU command buffer never completes (deterministic, Metal, M3 Max)

  Niko1221/Strata#5487 · Filter by author Qfbxz opened on Nov 7, 2026

• [GUI] Monitor: on a Mac "GPU load" is time busy, not compute in use; VRAM HIT RATE is always a dash

  Niko1221/Strata#5484 · Filter by author wenxinzhang opened on Nov 7, 2026

• [TUI] Path is a system folder, and Strata should not keep model files there

  Niko1221/Strata#5483 · Filter by author mitya12342 opened on Nov 7, 2026
  2

• [Benchmark] RX 7900 XTX (gfx1100) Windows / Unsloth UD-IQ3_XXS thinking / engine 0.1.58 hip-rdna3, standard-v1 ×3

  Niko1221/Strata#5482 · Filter by author rabbit718 opened on Nov 7, 2026

• [Engine] Windows/HIP: silent data corruption in hipMemcpy2D uploads from read-only file mappings, plus two engine findings (DP pinned-KV waste blocking 1M startup; VRAM reserve coupling into prefill chunk size)

  Niko1221/Strata#5480 · Filter by author bigbrid2023 opened on Nov 7, 2026
  2

• [Engine] Windows RTX 3090 / 0.1.56: GPU lost during 620–1362-token batched prefill, also without lookup, correlated WHEA-17

  Niko1221/Strata#5478 · Filter by author carabinshely opened on Nov 7, 2026

• [Engine] q36: flash attention for the prompt

  needs-format
  Niko1221/Strata#5465 · Filter by author bahooo22 opened on Nov 7, 2026
  1

• [Engine] q36: report need/free MiB on expert-buffer and staging OOM

  Niko1221/Strata#5464 · Filter by author bahooo22 opened on Nov 7, 2026

• [Engine] 4-card layer split: the expert cache per card falls as cards are added, and --prefill 1024 frees VRAM that the cache does not take

  Niko1221/Strata#5462 · Filter by author sean640112 opened on Nov 7, 2026
  1

• [Engine] GET /v1/models/{model_id} (OpenAI model retrieve) is still missing in 0.2.11 - blocks Visual Studio 2026 BYOK custom models

  Niko1221/Strata#5460 · Filter by author tengdahuang opened on Nov 7, 2026
  2

• [Engine] gfx1100 defaults row (STRATA_TSUM=1) from paired RDNA3 measurements - scope agreement

  needs-format
  Niko1221/Strata#5459 · Filter by author xyzzing opened on Nov 7, 2026
  1

• [GUI] Monitor: the Mac's gpu_note is never shown, so GPU temp, Power and PCIe are a bare or empty "–"

  Niko1221/Strata#5457 · Filter by author thomasgroeger opened on Nov 7, 2026

• [GUI] Long URL in a sent message is clipped on the left

  needs-format
  Niko1221/Strata#5456 · Filter by author LonelyFellas opened on Nov 7, 2026
  1

• [Engine] Vision Encoder fails with CUDA kernel error on heterogeneous dual GPUs (4090 D + 3060 Ti)

  Niko1221/Strata#5455 · Filter by author qew4 opened on Nov 7, 2026
  2

• [Engine] 0.1.55+: the new 12-token verify window faults deterministically at startup on sm_86 with an IQ3_S native pack — 0.1.54 (5-token windows) runs fine on the same box

  Niko1221/Strata#5449 · Filter by author ccmspringming opened on Nov 7, 2026
  1

• [Engine] serve: an asked include_usage never gets the usage chunk of its own; spec-following clients count 0 tokens

  Niko1221/Strata#5448 · Filter by author wornet-che opened on Nov 7, 2026
  2

• [Benchmark] DeepSeek-V4.1-Flash, 2x Quadro RTX 8000 (sm_75 source build), 21.6 tok/s

  Niko1221/Strata#5447 · Filter by author jiaoaoshirenjinbu opened on Nov 7, 2026
  2

• Bug Report: DeepSeek V4.1 Chunked Prefill Stalls on 4× RTX 3080, Plus Possible VRAM Reallocation Issue with --restart

  needs-format
  Niko1221/Strata#5444 · Filter by author xenonfear128 opened on Nov 7, 2026
  1

• Vision CPU reader (300-token cap) reads dense Japanese pages far worse than the GPU reader (CER 12.1% vs 1.66%)

  needs-format
  Niko1221/Strata#5443 · Filter by author masatokus opened on Nov 7, 2026
  1