serving_overlay/vllm/model_executor/models/gemma4.py is adapted from vLLM 0.19.1. Copyright contributors to the vLLM project; copyright 2025 The vLLM team and Google Inc. / HuggingFace Inc. team. The local changes support PLE reconstruction and the Gemma 4 checkpoint variants used in this comparison.

The upstream source is licensed under Apache 2.0. The accompanying patch contains the changes to that source.

data/request-sets/gsm8k-fewshot-train-first-8.jsonl and data/request-sets/gsm8k-full-reference-1319.jsonl contain examples from OpenAI's GSM8K dataset. Copyright (c) 2021 OpenAI. The MIT license applies to those examples.

data/request-sets/piqa-validation-complete-1838.jsonl contains the PIQA validation set by Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. The authors' dataset page specifies the Academic Free License 3.0.

data/request-sets/wikitext-2-raw-v1-test.txt and data/calibration/e4b-calibration/calibration_text.txt contain WikiText-2 text from Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher's WikiText dataset. The calibration text uses the training split at revision b08601e04326c79dfdd32d625aee71d232d685c3.

The dataset card attributes the text to Wikipedia and describes CC BY-SA 4.0 in its licensing section; its metadata also lists CC BY-SA 3.0 and GFDL. The texts retain the WikiText-2 content; request and calibration files select the splits and positions specified by the configurations.

The configurations refer to Google's Gemma 4 checkpoints, which are released under the Gemma 4 license. Download checkpoints from the model repositories named in configs/models/.