Seminário do discente Guilherme Lopes, dia 13/02/2026, às 14:00.

Seminário do discente Guilherme Lopes, dia 13/02/2026, às 14:00.

Título: Lost in Quantization: Disproportionate Degradation of Morphologically Rich Languages in INT8 vs. FP8 Inference

Abstract: Quantization is essential for scaling LLM inference, yet its impact on non-English languages remains under-explored. We investigate the sensitivity of Meta-Llama-3-8B to INT8 and FP8 (E4M3) quantization, comparing English and Brazilian Portuguese (PT-BR). Contrary to the hypothesis that morphological richness increases fragility, INT8 with outlier separation remains robust for both languages. However, naive FP8 casting degrades English perplexity significantly more (+18\%) than PT-BR (+3.9\%). Activation analysis reveals that English exhibits extreme outlier spikes ($>35$) that exceed the dynamic range of E4M3, whereas PT-BR activations are more concentrated within representable bounds.

Data: 13/02/26 às 14h