Evidence for a LiteRT-LM GPU bug report (MiniCPM-V-4 text decoder, Pixel 11 Pro)
Attachments for a Buganizer report: on a Pixel 11 Pro (PowerVR, OpenCL), LiteRT-LM 0.19.0's GPU backend answers off the prompt when an int8-weight decoder's prompt is prefilled through prefill_64 or prefill_256; prefill_16 and an int4 build of the same decoder are correct.
c_attachment_tables.mdβ table 1: the 12 text-only prompts and each arm's reply; table 2: files and environment; table 3: delegation lines.x12_prompts.jsonβ the 12 prompts (a passage about a living room, cut at six lengths, with one of two questions).
The model files themselves (4.6 GB and 3.0 GB) are not here; they can be shared on request.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support