vllm

[Attention Backend] TurboQuant: 2-bit KV cache compression with 4x capacity

#38479

Merged