vllm.models.hy_v4_flash.nvidia.indexpool_cache
¶
Engine-owned pool pages and speculative tail rings for Flash IndexPool.
Classes:
IndexPoolMetadataBuilder
¶
Bases: AttentionMetadataBuilder[IndexPoolMetadata]
Methods:
-
build_for_cudagraph_capture–Reserve scoring width for replay beyond the dummy's short context.
Source code in vllm/models/hy_v4_flash/nvidia/indexpool_cache.py
build_for_cudagraph_capture(common_attn_metadata)
¶
Reserve scoring width for replay beyond the dummy's short context.