/explore
|
You are here 
|
mercurialsolo.github.io | ||
| | | | |
www.spheron.network
FlashInfer is the kernel library backing vLLM, SGLang, TensorRT-LLM, and MLC-Engine. Step-by-step setup guide, 29-69% ITL gains, H100/H200/B200 benchmarks, and cost-per-token math. |
|
| | | | | ||
| | | | |
www.spheron.network
B300 vs B200 direct comparison: FP4 throughput, 288 GB vs 192 GB memory, real cost-per-token math at Spheron pricing, and a decision checklist for Blackwell GPU rental in 2026. |
|
| | | | | ||
| | | | |
www.spheron.network
Flash attention explained: why standard attention is memory-bound, how tiling avoids HBM round-trips, and how it lets one GPU serve longer context or more requests. |
|
| | | | | ||
| | | | |
blog.pamelafox.org
Recently, as part of my work on Azure OpenAI code samples, I've been experimenting with different ways of streaming data from a server i... |
|
| | | |||
