colibri · Issues· 143 open

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

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

[Bug]: RTX 3080 sm_86, WSL2 — "hybrid batched block failed in MoE" on every prompt

[Bug]: Qwen 3.8 Flash Next is too slow

Qwen3.8-Flash-Next: is a GPU backend (CUDA/HIP) planned for the qwen38 engine? (offer: gfx1151 / ROCm 7.2.4 validation + benchmarks)

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

[Feature]: How I can run Maxmini 2.7

[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`)

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

Vulkan is never torn down: no engine calls coli_vk_shutdown (Metal is, in kimi_k3)

[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)

Perplexity vs Ollama's Q4_K_M on the same tokens: our int4 experts lose 2.5–3.4 %, the int8 trunk buys nothing — two questions about the format

[Bug]: GPU doesn't engage

[Feature]:Add Dark mode switch button

DeepSeek V4: incoherent GPU output on GB10 (sm_121, unified memory) from the routed-expert VRAM mirror cache