litert-community

/

Phi-4-mini-instruct

like 19

Follow

LiteRT Community (FKA TFLite) 8.66k

Thanks for publishing this bundle — Phi-4-mini through litert-lm 0.16.0 on a Galaxy S26 GPU delegates fully and generates correctly.

This PR adds one subsection under Performance: measured Galaxy S26 (SM8850) rows for Phi-4-mini-instruct_multi-prefill-seq_q8_ekv4096.litertlm, GPU (OpenCL) against CPU (XNNPACK), conditions inline. Both backends generated a correct answer before the numbers were quoted; the GPU rows are full delegation (decode 1648/1648 ops on LITERT_CL). The section is labeled as a different runtime path from the S24 Ultra table above it, so the two are not conflated.

Happy to reshape to the house style if you prefer.