liteinfer benchmarks

Each config is scored against the one it improves on (vs base), against the start of its lineage (vs first), and against vLLM at the same batch width (vs vLLM). Above 1.00x is better. Each row names the config it improves on; vLLM rows are references, not competitors. Bars show standing on the headline metric of each table.

model Llama-3.2-1B-InstructISL 128OSL 256prompts 50 / 200decoding greedy, forced lengthlast run 2026-08-29T19:12:54+00:00

latency · Llama-3.2-1B-Instruct · ISL=128 OSL=256

One request in flight at a time. Headline metric: ITL p50. 50 prompts, forced output length 256. Prompt set 7790c497f52d.

configITL p50ITL p95TTFT p50TTFT p95E2E p50vs basevs firstvs vLLM
liteinfer-nocacheNo KV cache: every step re-feeds the whole sequence · reference point16.516.514.816.14,215.60.32x
liteinfer-eagerKV cache via transformers DynamicCache · improves on liteinfer-nocache13.713.915.716.33,521.21.20x1.20x0.38x
liteinfer-native-eagerKV cache as plain tensors, no DynamicCache · improves on liteinfer-eager13.713.914.816.13,518.51.00x1.20x0.38x
liteinfer-pagedPaged KV cache: fixed-size blocks from a pool · improves on liteinfer-native-eager15.315.319.321.13,910.70.90x1.08x0.34x
liteinfer-eager-b4Static batching, B=4, eager cache · improves on liteinfer-eager14.014.115.817.73,596.60.98x1.17x0.37x
liteinfer-native-eager-b4Static batching, B=4, native eager cache · improves on liteinfer-native-eager13.913.914.916.73,561.10.99x1.18x0.38x
liteinfer-paged-b4Static batching, B=4, paged cache · improves on liteinfer-paged15.015.019.020.93,832.21.02x1.10x0.35x
liteinfer-continuousContinuous batching, up to 32 concurrent sequences · improves on liteinfer-paged-b414.914.919.021.03,807.61.01x1.11x0.35x
vllmvLLM, one sequence at a time · reference point5.25.222.727.41,354.8
vllm-b4vLLM, up to 4 concurrent sequences · reference point5.25.222.627.01,353.9
vllm-continuousvLLM, up to 32 concurrent sequences · reference point5.25.223.232.21,356.2

throughput · Llama-3.2-1B-Instruct · ISL=128 OSL=256

Every request offered at once. Headline metric: output tok/s. 200 prompts, forced output length 256. Prompt set d97d71495aff.

configtok/sreq/swall (s)vs basevs firstvs vLLM
liteinfer-nocacheNo KV cache: every step re-feeds the whole sequence · reference point60.50.2846.70.32x
liteinfer-eagerKV cache via transformers DynamicCache · improves on liteinfer-nocache73.40.3697.91.21x1.21x0.39x
liteinfer-native-eagerKV cache as plain tensors, no DynamicCache · improves on liteinfer-eager72.70.3704.60.99x1.20x0.39x
liteinfer-pagedPaged KV cache: fixed-size blocks from a pool · improves on liteinfer-native-eager66.80.3766.70.92x1.10x0.35x
liteinfer-eager-b4Static batching, B=4, eager cache · improves on liteinfer-eager281.61.1181.83.84x4.66x0.39x
liteinfer-native-eager-b4Static batching, B=4, native eager cache · improves on liteinfer-native-eager277.61.1184.43.82x4.59x0.38x
liteinfer-paged-b4Static batching, B=4, paged cache · improves on liteinfer-paged252.91.0202.43.79x4.18x0.35x
liteinfer-continuousContinuous batching, up to 32 concurrent sequences · improves on liteinfer-paged-b41,268.35.040.45.01x20.98x0.28x
vllmvLLM, one sequence at a time · reference point188.40.7271.8
vllm-b4vLLM, up to 4 concurrent sequences · reference point724.02.870.7
vllm-continuousvLLM, up to 32 concurrent sequences · reference point4,466.617.411.5