ttft llm
2 articles
Estimating TTFT From FLOPs: The Prefill Compute Floor
Derive a time-to-first-token floor from model FLOPs, dense GPU TFLOPS and MFU, with the attention term, the sparsity trap and H100 worked examples.
October 28, 2026• 15 min read
ttft llmttft itldgx spark prefill speed
Chunked Prefill Tuning: Token Budget, TTFT and ITL
How max_num_batched_tokens and chunked-prefill-size trade TTFT against ITL in vLLM and SGLang, with defaults, worked arithmetic and a sweep protocol.
October 28, 2026• 15 min read
chunked prefill vllmmax_num_batched_tokens tuningchunked prefill ttft itl tradeoff