Issues

• nightly: Apple performance guard is red

  bug
  geisten/geistlib#602 · Filter by author github-actions opened on Oct 4, 2026

• nightly: pi5 is red

  bug
  geisten/geistlib#601 · Filter by author github-actions opened on Oct 4, 2026
  1

• research(decision): shared-state multi-question scoring with hybrid-session isolation

  bonsaidecision-inferenceenhancement
  geisten/geistlib#589 · Filter by author geisten opened on Oct 3, 2026

• research(decision): evaluate trained classifier heads and checkpoint-bound calibration

  decision-inferenceenhancement
  geisten/geistlib#588 · Filter by author geisten opened on Oct 3, 2026

• eval(decision): Bonsai 2 quality, calibration and latency without generated reasoning

  bonsaidecision-inferenceenhancement
  geisten/geistlib#587 · Filter by author geisten opened on Oct 3, 2026
  3

• Tracking: modular decision inference in geistlib

  bonsaidecision-inferenceenhancement
  geisten/geistlib#583 · Filter by author geisten opened on Oct 3, 2026
  1

• perf(metal): pin_prefix on Gemma 4 E4B is slower than re-prefilling (4.7 s pin, no gain after)

  bug
  geisten/geistlib#578 · Filter by author geisten opened on Oct 3, 2026
  2

• memory: Gemma 4 E4B Q4_K_M peaks at 11.7 GiB on cpu_x86 and 10.6 GiB on Metal (4.6 GiB file, load_from_memory)

  bug
  geisten/geistlib#577 · Filter by author geisten opened on Oct 3, 2026
  2

• metal: one command sequence per session thread, so concurrent sessions overlap on one GPU

  enhancement
  geisten/geistlib#576 · Filter by author geisten opened on Oct 3, 2026
  1

• cpu_neon: native W*A8 int8 GEMM for prefill, replacing dequant + BLAS sgemm (needs i8mm hardware)

  enhancement
  geisten/geistlib#500 · Filter by author geisten opened on Sep 27, 2026
  2

• vulkan: scratch pool spills out of the 256 MB BAR heap into host memory — prefill collapses for chunks above 64

  enhancement
  geisten/geistlib#488 · Filter by author geisten opened on Sep 26, 2026
  3

• vulkan: shader and sequencing efficiency backlog (attention, reductions, staging reuse, first-submit latency) — unmeasured

  enhancement
  geisten/geistlib#475 · Filter by author geisten opened on Sep 26, 2026
  2

• vulkan: robustness debts from the backend review — sticky flush failure, probe/entry drift, fallback visibility, limits and constants

  enhancement
  geisten/geistlib#474 · Filter by author geisten opened on Sep 26, 2026
  3

• vulkan: efficiency backlog — VkPipelineCache/lazy pipelines, registry lookup, upload staging ring, fused act-quant, DeltaNet reductions

  enhancement
  geisten/geistlib#469 · Filter by author geisten opened on Sep 26, 2026
  2

• transformer/vulkan: arena-skip predicate, lazy embed-table resolve and resident mmap after upload (weights_device_copy follow-ups)

  enhancement
  geisten/geistlib#468 · Filter by author geisten opened on Sep 26, 2026
  1

• vulkan: tune Q4_0/Q5_K/TQ2_0 and DeltaNet kernels; benchmark qwen3.5/3.8 (0.8B, 4B, 27B)

  enhancement
  geisten/geistlib#467 · Filter by author geisten opened on Sep 26, 2026
  5

• vulkan: models larger than device memory (27B Q4_0 on an 11 GiB card) — no host spill

  enhancement
  geisten/geistlib#466 · Filter by author geisten opened on Sep 26, 2026
  1

• vulkan: one per-dtype descriptor table; shared shader and test helpers (cleanup follow-up to #458)

  enhancement
  geisten/geistlib#465 · Filter by author geisten opened on Sep 26, 2026
  2

• perf(mac): Apple CPU path trails llama.cpp — 0.53× prefill / 0.72× decode on bare-metal M1 Max

  bug
  geisten/geistlib#456 · Filter by author geisten opened on Sep 25, 2026
  3

• bench: populate a same-protocol model × system performance matrix

  enhancement
  geisten/geistlib#364 · Filter by author geisten opened on Aug 31, 2026
  7