{"type":"post","id":"megakernel-part-1-vllm-and-llama-cpp-leave-276030","slug":"megakernel-part-1-vllm-and-llama-cpp-leave-276030","title":"MegaKernel part 1: vLLM and llama.cpp leave over half of memory roofline unused on SmolLM2-135M at batch 1","description":"A public MegaKernel repository benchmarks batch-one decode on an RTX 4060 Laptop GPU and traces a context-scaling problem to attention dispatch.","retrieval_nugget":"A public MegaKernel repository benchmarks batch-one decode on an RTX 4060 Laptop GPU and traces a context-scaling problem to attention dispatch. Its split-K Triton attention path reduces the measured attention time at 2,048 tokens from 2.99 ms to 0.50 ms. For inference builders, the result is a useful reminder that kernel launches and attention behavior can dominate small-model decoding.","published_at":"2026-11-17","updated_at":"2026-11-17","record_date":"2026-11-17","date_kind":"published_at","topics":["memory"],"entities":["MegaKernel","vLLM","llama.cpp","SmolLM2-135M","Qwen2.5-1.5B","RTX 4060 Laptop GPU"],"related_patterns":["agent-economics-moves-to-completed-work"],"novelty_label":"new_fact","what_changed":"MegaKernel reports that vLLM and llama.cpp leave more than half of the memory roofline unused when decoding SmolLM2-135M at batch 1.","prior_refs":["https://newruntime.com/posts/vllm-ships-a-specialized-cute-dsl-decode-path-dd574e/","https://newruntime.com/posts/deepseek-v4-1-flash-on-vera-rubin-nvl72-0d61a3/","https://newruntime.com/posts/betelgeuse-llama-cpp-fork-pages-moe-experts-from-fb5996/"],"source_url":"https://github.com/crankyastrologer/megakernel","source_urls":["https://github.com/crankyastrologer/megakernel"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:megakernel-part-1-vllm-and-llama-cpp-leave-276030","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/megakernel-part-1-vllm-and-llama-cpp-leave-276030/","markdown":"https://newruntime.com/posts/megakernel-part-1-vllm-and-llama-cpp-leave-276030.md","json":"https://newruntime.com/posts/megakernel-part-1-vllm-and-llama-cpp-leave-276030.json"},"markdown_url":"https://newruntime.com/posts/megakernel-part-1-vllm-and-llama-cpp-leave-276030.md","next_reads":[{"type":"topic","path":"/topics/memory/","url":"https://newruntime.com/topics/memory/","title":"Memory","reason":"Explore the memory topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/memory.md","json_url":"https://newruntime.com/topics/memory.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}
