[BugFix] Add block_size validation for mamba cache align mode (#34445)

Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>

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