Speaker
Description
Quantization‑aware training (QAT) is essential for delivering low‑latency inference on the FPGAs that power high‑energy‑physics (HEP) experiments. QKeras has become the de‑facto Keras‑based framework for QAT in this community and is tightly coupled with the hls4ml toolchain, which converts Keras models into synthesizable HLS code. The transition from Keras 2 to Keras 3 introduced a new "ops" API and a redesigned model‑serialization format, breaking compatibility with existing QKeras models and halting the established hls4ml workflow.
We present QKerasV3, a modernized implementation that restores full compatibility with Keras 3 while preserving the quantization semantics required by hls4ml. The redesign rewrites all quantized layers to use the Keras 3 ops interface, introduces a robust custom‑object registration mechanism that guarantees correct model saving and loading, and adds strict configuration validation to detect mismatched bit‑widths, rounding modes, or saturation policies at construction time. The hls4ml export module has been updated to ingest Keras 3 models directly, generate the appropriate fixed‑point types, and produce HLS code that faithfully reproduces the behavior of the original QKeras layers.
Validation on representative HEP datasets shows that the generated HLS designs match the latency, resource utilization, and numerical accuracy of the legacy Keras 2 flow within expected tolerances. QKerasV3 therefore enables HEP users to adopt the latest Keras ecosystem while retaining a seamless path to FPGA deployment via hls4ml.
| Do you plan to submit a 4-page extended abstract on OpenReview (only for Presentations/Posters)? | No |
|---|