WAN 2.2 Sampled for 28 Minutes, Then Died at VAE Decode — 48 Channels vs 16
A mismatched VAE ended a WAN 2.2 run after 28 minutes of sampling. The successful retry used different settings, so its timing is not a controlled comparison.
A mismatched VAE ended a WAN 2.2 run after 28 minutes of sampling. The successful retry used different settings, so its timing is not a controlled comparison.
My first exception concerned multi-file refactors. GPU contention complicated the local-coding policy later. The chronology matters when explaining why the rule changed.
Four keyboards, one preferred layout, and different places to store mappings. My K380 switching preference did not carry to every board; Air60 HE is wired-only.
A resident whisper server reduced the recorded STT stage from about 9.9 to 1.0 seconds. End-to-end latency and answer-quality tradeoffs were not measured.
I built an AI-assisted phone bridge around existing CLI sessions. A restart loop, argument parsing, keyboard handling, and stale PWA state exposed different boundaries.
Two local embedding models ranked the same 24 bookmarks identically — 8 of 8 correct each. Their similarity scores sat in almost non-overlapping ranges, and one hardcoded 0.75 cutoff silently erased every result from one of them.
A cloud model wrote 1,900 lines in one session and burned my five-hour limit — then the next session remembered none of it. Three layers of context files plus hooks, and what each one is actually for.
I swept expert CPU offload on two PCs. The tables show throughput and a silent speed cliff; they do not isolate the driver’s memory policy.
Same model, same build, same test — one on an RTX 2080 Ti (11GB, CUDA), one on an RX 9070 XT (16GB, Vulkan). The result split: 1.9x prompt processing one way, 1.3x generation the other.
Real llama-bench numbers for Qwen3-Coder-30B-A3B on an AMD RX 9070 XT (16GB, RDNA4) using llama.cpp with Vulkan — including the full n-cpu-moe sweep that took generation from 21 to 47 tok/s.