Gemma 4 FPGA Artifact Bundle
A reproducible deployment bundle for running a distilled Gemma-based language model on a Xilinx Kria KV260 FPGA board, including the Verilog hardware sources, bitstream, tokenizer and integrity checks needed for edge inference.
Gemma 4 FPGA Artifact Bundle is a deployment bundle for running Gemma-based language-model inference on a Xilinx Kria KV260 FPGA, chosen because its Zynq UltraScale+ chip combines an ARM control-plane running Linux with FPGA programmable logic for parallel compute, and because its camera and image-signal-processing support make camera-to-inference pipelines realistic on the same board. The goal stated in the repository is deterministic, reproducible edge inference with strict artifact-integrity checks and a portable bring-up process for compatible hardware targets.
The bundle includes a compiled bitstream and split binary weight files, a model contract and pinned tokenizer with recorded hashes for integrity verification, and the underlying Verilog and SystemVerilog hardware sources, formal verification files and constraints. The deployed release corresponds to a 36-core hardware profile; the runtime model is the project's own smaller distilled model rather than the full-size teacher model, run through a bounded working set with hardware-managed streaming rather than requiring the whole model resident in memory at once.