From 6f019e6e0a0cde34a33826bc08756480816448dd Mon Sep 17 00:00:00 2001 From: Harry Huang Date: Fri, 13 Feb 2026 10:18:07 +0800 Subject: [PATCH] [BugFix] Add block_size validation for mamba cache align mode (#34445) Signed-off-by: huanghaoyan.hhy --- vllm/config/vllm.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index e9f6b37ab..0310e8aed 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -1110,6 +1110,15 @@ class VllmConfig: self.scheduler_config.disable_hybrid_kv_cache_manager = False if self.cache_config.mamba_cache_mode == "align": + assert ( + self.cache_config.block_size + <= self.scheduler_config.max_num_batched_tokens + ), ( + "In Mamba cache align mode, block_size " + f"({self.cache_config.block_size}) must be <= " + "max_num_batched_tokens " + f"({self.scheduler_config.max_num_batched_tokens})." + ) if self.scheduler_config.long_prefill_token_threshold > 0: assert ( self.scheduler_config.long_prefill_token_threshold