evnisv/spec_gemma_4_benchmark_results

Gemma 4 31B speculative-decoding benchmark results This repository contains reproducible benchmark artifacts for google/gemma-4-31B-it on one NVIDIA H100 NVL, comparing baseline decoding, native Gemma MTP, and DFlash. Raw prompts, generated responses, per-request timings, server logs, commands, metadata, and acceptance records are retained under results/. H100 quick start Use an H100 with at least 80 GB VRAM. Target weights are BF16; the published matrix uses FP8… See the full description on the dataset page: https://huggingface.co/datasets/evnisv/spec_gemma_4_benchmark_results.

<iframe src="https://silytics.si/embed/item/b7039cbe-12a5-4c03-9da1-e84110e53ab9" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.