From 98661fe012c5c467252d4df8411d2f46190e9268 Mon Sep 17 00:00:00 2001 From: Itay Etelis <92247226+Etelis@users.noreply.github.com> Date: Tue, 5 May 2026 14:54:29 +0300 Subject: [PATCH] [Bugfix][KVConnector] Support DCP/PCP in OffloadingConnector (#41549) Signed-off-by: Itay Etelis Co-authored-by: Itay Etelis --- vllm/v1/kv_offload/base.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/vllm/v1/kv_offload/base.py b/vllm/v1/kv_offload/base.py index dc9d95499c..3d403ea508 100644 --- a/vllm/v1/kv_offload/base.py +++ b/vllm/v1/kv_offload/base.py @@ -331,12 +331,20 @@ class OffloadingSpec(ABC): assert kv_transfer_config is not None self.extra_config = kv_transfer_config.kv_connector_extra_config + parallel_config = vllm_config.parallel_config + context_parallel_factor = ( + parallel_config.decode_context_parallel_size + * parallel_config.prefill_context_parallel_size + ) + # block size used by vLLM for hashing request tokens for the sake # of enabling prefix caching - self.hash_block_size = vllm_config.cache_config.block_size + self.hash_block_size = ( + vllm_config.cache_config.block_size * context_parallel_factor + ) # gpu block size per group self.gpu_block_size: tuple[int, ...] = tuple( - kv_cache_group.kv_cache_spec.block_size + kv_cache_group.kv_cache_spec.block_size * context_parallel_factor for kv_cache_group in kv_cache_config.kv_cache_groups )