Repository Issue Activity (beta)

#1384 [Performance]: GLM-5.2 int4 gs64 on Azure D128ads v7, 4x local NVMe RAID0

#1376 [Feature]:

#1375 [Bug]: GLM 5.3-Flash puts Coli Tune OOM in 1.10.2

#1380 [Performance]: GLM-5.2 int4 gs64 on Azure E64pds v6 Arm64, 4x local NVMe RAID0

#1379 [Performance]: GLM-5.2 int4 gs64 on Azure E64ads v7, 4x local NVMe RAID0

#1378 [Feature]: Add Architecture for OpenAI-Compatible Tool Translation Proxy via Sidecar/Dual-Engine Routing

#1370 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

#689 [Bug]: deep speculative verify batches are not token-exact vs CPU on CUDA (near-tie flips + lower draft acceptance)

#1339 qwen36 CUDA tier: qt_issue overruns G.is_x with two or more GPUs (stride 8*D into a 32*D buffer)

#1341 qwen36 CUDA tier: int8 containers free the weights the tier still points at (regression from #1334)

#1340 qwen36 CUDA tier: qt_shutdown never signals cv_take, so pthread_join can hang if a group is open

#1331 [Bug]: CUDA VRAM expert tier silently no-ops on int8 checkpoints -- reserves budget, promotes zero experts, no error

#1359 [Bug]: release packer does not follow the imports of scripts it packages for subprocess launch

#1368 [Bug]: 1.10.1 Windows Coli Web GLM 5.3-Flash embed_tokens tensor

#1242 [Feature]: Add support for Qwen3.8-Flash-Next

#1365 [Bug]: coli doctor reports 2 missing core tensors on a healthy GLM-5.3-Flash: the required list is the GLM-5.2 layout

#1040 qwen36 GPU detection: 'none compute-capable' on NVIDIA (GTX 1660 Ti) → CPU fallback

#1352 Discord invite link expired

#683 I trained a small model for this engine from scratch, anyone can use it

#1351 [Bug]: single-GPU auto expert tier stops at ~56% of budget: prefix_est divides by the MTP int8 width since #766 (masks #687 on dev)

#687 [Bug]: CUDA_EXPERT_GB=auto fills all VRAM on single-GPU; lazy dense uploads then fail 60x and fall back to CPU silently

#708 [Experiment] expert-transition-history placement policy vs gate-momentum — controlled A/B for hypothesis #1

#797 Intel Optane Storage on PMEM 100 or PMEM 200

#441 [Performance]: PILOT prefetch runs at NVMe queue-depth 1, and DIRECT=1 re-reads WILLNEED-prefetched pages

#972 [Feature]: Add support for Mistral Large 3