https://sparsenotes.com/tags/llm-systems/ Recent content in llm-systems on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 25 Jul 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/07/papers/windowed-mtp-draft-kv-tax/ Sat, 25 Jul 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/07/papers/windowed-mtp-draft-kv-tax/ A training-free, lossless fix for speculative decoding at million-token context: apply StreamingLLM-style sliding attention to the MTP draft head only, cutting per-step draft cost by 28–44% while the full-attention target still verifies every token.