Add Tensor G4 build (seq256, mixed-precision)
#5
by xThr45hx - opened
Adds the missing Tensor G4 entry to the SoC matrix.
This repo currently covers google.tensor_g5, google.tensor_g6, mediatek.mt6991 / mt6993,
and qualcomm.sm8550 / sm8650 / sm8750 / sm8850 across seq256 / 512 / 1024 / 2048 —
every supported SoC except google.tensor_g4.
File: embeddinggemma-300M_seq256_mixed-precision.google.tensor_g4.tflite (196,993,056 B)
Mixed-precision AOT build for Tensor G4 (Pixel 9 series), named to match the existing convention
in this repo. Carries a compiled DarwiNN DGC payload.
Only seq256 here for now — the other three sequence lengths compile the same way if that would
be useful.
Caveat: as with any AOT DGC, this needs a compatible GoogleTensor dispatch library on device.