MLA-Whisper across 11 languages, and what the compression actually costs
中文摘要
作者将 MLA-Whisper 的 KV 缓存压缩技术扩展至 11 种语言和 6 种脚本，弥补了原研究仅限英语且未发布模型的空白。
English Summary
MLA-Whisper's KV cache compression was extended to 11 languages and six scripts, filling the gap left by the original English-only research and missing models.
Original Excerpt
The Whisper-MLA paper compressed the decoder’s KV cache in English and released no models. I extended it to 11 languages and six scripts… Continue reading on Medium »