DFlash drafter for Muse Glimmer-30B: 5 layers, blockwise 16-token spec decoding
TimDarcet · x · 2026-10-21
TimDarcet says Muse Glimmer-30B fits in memory but remains bandwidth-bound, so the team added a DFlash speculative-decoding drafter: 5 draft layers, block size 16, sliding-window attention (2048) on all layers, 32 query / 8 KV heads (GQA), and 128k sequence length.
Related event: Meta Engineers Train a 30B LLM from Scratch and Squeeze It onto a 32GB GPU(4 posts)→
More from Models
- X trends: SophontAI's Holon-S1 predicts gene activity from cancer slides; decision models rise — iScienceLuvr · 2026-10-21
- OpenAI's MultiLangBench: PHP scores highest (91.3%) at building a database engine from scratch — ricklamers · 2026-10-21
- 704-vote robotics Elo ranks Claude Opus 5.5 first at 1379; no model completes the task — qinzytech · 2026-10-21
- DeepSeek V4 called 'god-tier infra, 4th-tier ML': 7 tricks stacked, 20 patches to fix them — teortaxesTex · 2026-10-21
- GPT-6 Astra and Claude Fable 5.1 write $10,000 research papers — experts reject all — sayashk · 2026-10-21
- GitHub Copilot's HydraFusion Orchestrates Multiple Models Per Task — WirelessLife · 2026-10-21