SageAttention is research-only: absent from vLLM and SGLang LLM serving paths

SageAttention is a quantization-based attention optimization that is NOT deployed in production engines. It is absent from the vLLM source tree and in SGLang it appears only under multimodal_gen, not the LLM serving path. SageAttention quantizes attention computations to int8/fp8 to reduce compute cost, but it is research-only. The key distinction: SageAttention is not in vLLM (absent from the tree), not in SGLang's LLM path (only in multimodal_gen), and not in any production serving engine. This is a verified fact from source code inspection, not a README claim.

SageAttention is research-only. It is NOT in vLLM (absent from the tree). In SGLang, it appears only under multimodal_gen, not the LLM serving path. This is verified from source code, not from README claims. The practical implication: do not expect SageAttention to be available in production serving engines. For quantized attention, use FA-3's FP8 path on Hopper or FA-4's FP8 path on Blackwell, or FlashMLA's fp8/fp4 KV cache on Blackwell. Sources: vLLM GitHub, SGLang GitHub, SageAttention paper.