SO, LET'S SUPPORT 85+ MODELS WITH 35+ RUNTIMES AND INFERENCE BACKENDS

Every architecture is different under the hood

bert_flash

qwen2_flash

modernbert_flash

colbert

cross_encoder

clip_siglip_florence2

splade_flash

sglang

No universal engine

An agentic workflow

Re-implementation

15 of 35 adapters use Flash Attention 2 with variable-length sequence packing