A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 19:00 UTC
Patch SlidingWindowSpec.real_page_size_bytes for nvfp4 kv (#42464)
Signed-off-by: Shiyang Chen <shiychen@nvidia.com>
sychen52 committed
May 13, 2026 at 01:46 UTC
a8c13d28374702278210d33eec348ed007abfece
1 file changed
+11
vllm/v1/kv_cache_interface.py
+11
@@ -435,6 +435,17 @@ class SlidingWindowSpec(AttentionSpec):
435
436
@property
437
def real_page_size_bytes(self) -> int:
438
+ # Mirror ``FullAttentionSpec.real_page_size_bytes`` for NVFP4 KV cache.
439
+ if self.kv_quant_mode.is_nvfp4:
440
+ last_dim = nvfp4_kv_cache_full_dim(
441
+ self.head_size
442
+ ) + nvfp4_kv_cache_full_dim(self.head_size_v)
443
+ return (
444
+ self.block_size
445
+ * self.num_kv_heads
446
+ * last_dim
447
+ * get_dtype_size(self.dtype)
448
+ )
449
return (
450
self.block_size
451
* self.num_kv_heads