[Bugfix] get_num_blocks_to_allocate with null_block (#19031)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>

vllm/v1/core/block_pool.py

@@ -63,6 +63,7 @@ class BlockPool:
         # The ref_cnt of null_block is not maintained, needs special care to
         # avoid freeing it.
         self.null_block = self.free_block_queue.popleft()
+        self.null_block.is_null = True
 
         self.enable_kv_cache_events = enable_kv_cache_events
         self.kv_event_queue: list[KVCacheEvent] = []
@@ -252,7 +253,7 @@ class BlockPool:
         for block in blocks:
             # ref_cnt=0 means this block is in the free list (i.e. eviction
             # candidate), so remove it.
-            if block.ref_cnt == 0 and block != self.null_block:
+            if block.ref_cnt == 0 and not block.is_null:
                 self.free_block_queue.remove(block)
             block.incr_ref()
 
@@ -267,7 +268,7 @@ class BlockPool:
         for block in ordered_blocks:
             block.decr_ref()
             # null_block should not be added to the free list.
-            if block.ref_cnt == 0 and block != self.null_block:
+            if block.ref_cnt == 0 and not block.is_null:
                 self.free_block_queue.append(block)
 
     def reset_prefix_cache(self) -> bool:

vllm/v1/core/kv_cache_utils.py

@@ -125,6 +125,9 @@ class KVCacheBlock:
     prev_free_block: Optional["KVCacheBlock"] = None
     next_free_block: Optional["KVCacheBlock"] = None
 
+    # Whether the block is a null block that should never be cached.
+    is_null: bool = False
+
     def incr_ref(self):
         self.ref_cnt += 1
 

vllm/v1/core/single_type_kv_cache_manager.py

@@ -83,8 +83,9 @@ class SingleTypeKVCacheManager(ABC):
         # free queue and ref_cnt == 0), it will be changed from a free block
         # to a computed block when the request is allocated, so we also count
         # it as needed to be allocated.
-        num_evictable_computed_blocks = sum(blk.ref_cnt == 0
-                                            for blk in new_computed_blocks)
+        num_evictable_computed_blocks = sum(
+            blk.ref_cnt == 0 and not blk.is_null
+            for blk in new_computed_blocks)
         return ((num_new_blocks + num_evictable_computed_blocks) *
                 self.num_kv_cache_groups)