diff --git a/vllm/v1/kv_cache_interface.py b/vllm/v1/kv_cache_interface.py index fa58297c42..66d1582e21 100644 --- a/vllm/v1/kv_cache_interface.py +++ b/vllm/v1/kv_cache_interface.py @@ -435,6 +435,17 @@ class SlidingWindowSpec(AttentionSpec): @property def real_page_size_bytes(self) -> int: + # Mirror ``FullAttentionSpec.real_page_size_bytes`` for NVFP4 KV cache. + if self.kv_quant_mode.is_nvfp4: + last_dim = nvfp4_kv_cache_full_dim( + self.head_size + ) + nvfp4_kv_cache_full_dim(self.head_size_v) + return ( + self.block_size + * self.num_kv_heads + * last_dim + * get_dtype_size(self.dtype) + ) return ( self.block_size * self.num_kv_heads