🎬 Gemma Advisor — PR0 Feasibility Harness (gemma-3-270m)

Measures download size, load/inference latency, peak memory, WebGPU support, and — most importantly — output correctness per dtype (the fp16/q4f16 WebGPU overflow bug). Run on each target device/browser and record the results table into the PR0 findings.

Served from a Hugging Face Static Space (https) — WebGPU works here. If you instead run this locally, serve over http (python -m http.server), not file:// (WebGPU + ES modules are blocked on the file origin).

Environment

Results

Model output (eyeball for correctness — garbage = dtype broken)

Log