
ARGODRIVE
github.com/argonautlabsai/argodrive- Category
- Developer Tools
- Rank
- No. 2264Tools index
- Pricing
- Open Source
- Type
- TOOL
- GitHub
- 3 stars
- Latest release
- v0.2.0-beta.3
- Date
About
ARGODRIVE is a layout, balancer, and instrumentation toolkit for running mixture-of-experts language models directly from SSDs on memory-constrained machines, such as a 518GB model on a 128GB laptop. It found that the standard prefill sweep reads every expert of every routed layer even though a 512-token chunk only routes to about half of them, and fixing that gap alone yields a measured 1.73x speedup in prompt processing on one internal SSD without adding replicas.
Tags
mixture-of-expertsllm-inferencessdopen-sourcebenchmarkinglocal-llm
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.