vllm

[Core] Default to using per_token quantization for fp8 when cutlass is supported.

#8651

Merged