← sources

dev.to: One Word, Two GPUs: Why Gemma 4 Answers the Same Question Differently on a Laptop and a Tesla T4

source: dev-ecosystems (dev-ecosystems:devto)published 2026-11-06 16:31ingested 2026-11-12 02:33

retained text

Same GGUF file, same llama.cpp build, greedy decoding, and a GTX 1650 Ti answers $20 where a Tesla T4 answers 2000. One GSM8K question followed down to the word where the two cards part, the order in which each card adds its numbers, and the one stock setting that makes them agree.

tags: gemma, llamacpp, cuda, nvidia

reactions=26 | comments=7

document fetched 2026-11-12 02:33