SOTA2 Research TTFT measurement on context sequences 32K-256K (vLLM 0.18.0/0.20.0)
benchmark
Fields
id
sources/bench-sota2-wizwand-vllm-ttft-32k-256kPublisher
SOTA2 Research (formerly WizWand)
Kind
benchmark
Published
2026-06-05
Accessed
2026-10-03
notes
TTFT leaderboard for vLLM 0.18/0.20 across 32K-256K context. Dense baseline is FA2 (FlashAttention 2). Sparse numbers use a distilled-indexer sparse attention variant. Hardware: GPU 80G (specific GPU not stated) and NPU 64G (Ascend). Serving: vLLM 0.20 chunk 4K for GPU, vLLM 0.18 full for NPU. Concurrency and cache state not explicitly stated.
Notes
TTFT leaderboard for vLLM 0.18/0.20 across 32K-256K context. Dense baseline is FA2 (FlashAttention 2). Sparse numbers use a distilled-indexer sparse attention variant. Hardware: GPU 80G (specific GPU not stated) and NPU 64G (Ascend). Serving: vLLM 0.20 chunk 4K for GPU, vLLM 0.18 full for NPU. Concurrency and cache state not explicitly stated.
Cross-references
This record references no other records.