All issues

Search results

[Experiment] Routing-count coldness badly overestimates expert prunability — OLMoE int8, M5 Pro, a8f2ca6

Status: Open.

#1621 In JustVugg/colibri;

· KhalidAlnujaidi opened on Sep 19, 2026

[Bug]: DeepSeek V4 REAP checkpoint: every per_matrix expert record takes the buffered read path (~36 % of expert reads), bypassing F_NOCACHE/O_DIRECT

Status: Open.

#1615 In JustVugg/colibri;

· Freizeitminister opened on Sep 18, 2026

[Performance]: macOS planner trusts free + inactive, but the kernel compresses the expert cache — 32 GiB budget is slower than 16 GiB on a 48 GB Mac

Status: Open.

#1614 In JustVugg/colibri;

· Freizeitminister opened on Sep 18, 2026

[Feature]: heterogeneous 3-lane MoE scheduling, wall-clock telemetry & routing-history improvements (from the Peregrine prototype)

Status: Open.

#1609 In JustVugg/colibri;

· s-b-repo opened on Sep 18, 2026

coli serve: RAM auto-detect returns 0.0 GB and silently falls back to a tiny expert cache (no warning)

Status: Open.

#1601 In JustVugg/colibri;

· adirburke opened on Sep 18, 2026

coli run: SNAP env var never set for non-glm engines (olmoe confirmed) — always fails with 'started without a model'

Status: Open.

#1600 In JustVugg/colibri;

· adirburke opened on Sep 18, 2026

[Bug]: Qwen 3.8 Flash Next is too slow

Status: Open.

#1594 In JustVugg/colibri;

· MichaelFomenko opened on Sep 17, 2026

[Bug]: doctor reports "no tensor fills these core roles: token embedding, output head" for a healthy DeepSeek V4 container

Status: Open.

#1593 In JustVugg/colibri;

· pauloapoloni opened on Sep 17, 2026

[Feature]: How I can run Maxmini 2.7

Status: Open.

#1592 In JustVugg/colibri;

· neno-research opened on Sep 17, 2026

[Bug]: Linux, sibling engines: --auto-tier silently drops the VRAM tier that coli plan/coli doctor advertise (qwen36 CUDA build, 11.8 → 21 tok/s with --gpu auto)

Status: Open.

#1581 In JustVugg/colibri;

· troturier opened on Sep 17, 2026

[Bug]: qwen36 CUDA_DLL on Windows: tier selects 0 devices unless COLI_GPUS is set (available_device_count is queried before the DLL is loaded)

Status: Open.

#1577 In JustVugg/colibri;

· cyphrrr opened on Sep 16, 2026

glm53: KV prefix reuse lost after bare tool-call turns — template writes "\n<tool_call>", the model does not

Status: Open.

#1576 In JustVugg/colibri;

· vrinek opened on Sep 16, 2026