[v1][KVCacheManager] Avoid full cache hit by controlling max_length (#17999)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

tests/v1/core/test_specialized_manager.py

@@ -17,8 +17,9 @@ def get_sliding_window_manager(sliding_window_spec, block_pool):
 
 
 def test_sliding_window_possible_cached_prefix():
+    block_size = 2
     sliding_window_spec = SlidingWindowSpec(
-        block_size=2,
+        block_size=block_size,
         num_kv_heads=1,
         head_size=1,
         dtype=torch.float32,
@@ -44,7 +45,9 @@ def test_sliding_window_possible_cached_prefix():
                     i: block_pool.blocks[i + 10]
                 }
 
-        computed_blocks = manager.find_longest_cache_hit(block_hash_list)
+        computed_blocks = manager.find_longest_cache_hit(
+            block_hash_list,
+            len(block_hash_list) * block_size)
         assert len(computed_blocks) == expect_length
 
         assert all(block == block_pool.null_block

vllm/v1/core/kv_cache_manager.py

@@ -146,21 +146,16 @@ class KVCacheManager:
             assert self.prefix_cache_stats is not None
             self.prefix_cache_stats.requests += 1
 
-        if len(block_hashes) * self.block_size == request.num_tokens:
-            # When prompt length is divisible by the block size and all
-            # blocks are cached, we need to recompute the last token. This
-            # have to be achieved by re-computing an entire block because
-            # allocate_slots() assumes num_computed_tokens is always a
-            # multiple of the block size. To achieve this, remove the last
-            # block hash from the block_hashes for find_longest_cache_hit
-            # This limitation can potentially be removed in the future to
-            # slightly improve the performance.
-            last_block_hash = block_hashes.pop()
-        else:
-            last_block_hash = None
+        # NOTE: When all tokens hit the cache, we must recompute the last token
+        # to obtain logits. Thus, set max_cache_hit_length to prompt_length - 1.
+        # This can trigger recomputation of an entire block, rather than just
+        # the single last token, because allocate_slots() requires
+        # num_computed_tokens to be block-size aligned. Removing this limitation
+        # could slightly improve performance in the future.
+        max_cache_hit_length = request.num_tokens - 1
 
-        computed_blocks = (
-            self.single_type_manager.find_longest_cache_hit(block_hashes))
+        computed_blocks = self.single_type_manager.find_longest_cache_hit(
+            block_hashes, max_cache_hit_length)
         # NOTE(woosuk): Since incomplete blocks are not eligible for
         # sharing, `num_computed_tokens` is always a multiple of
         # `block_size`.
@@ -171,12 +166,6 @@ class KVCacheManager:
             self.prefix_cache_stats.queries += request.num_tokens
             self.prefix_cache_stats.hits += num_computed_tokens
 
-        if last_block_hash is not None:
-            # Add back the last block hash if it was removed.
-            # NOTE: Because block_hashes is cached in req_to_block_hashes,
-            # we shouldn't modify it directly.
-            block_hashes.append(last_block_hash)
-
         return KVCacheBlocks(computed_blocks), num_computed_tokens
 
     def allocate_slots(

vllm/v1/core/single_type_kv_cache_manager.py

@@ -187,17 +187,19 @@ class SingleTypeKVCacheManager(ABC):
         raise NotImplementedError
 
     @abstractmethod
-    def find_longest_cache_hit(
-            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+    def find_longest_cache_hit(self, block_hashes: list[BlockHashType],
+                               max_length: int) -> list[KVCacheBlock]:
         """
-        Get the longest cache hit prefix of the blocks. If no cache hit is 
-        found, return an empty list. if eagle is enabled, drop the last matched 
-        block to force recompute the last block to get the required hidden 
-        states for eagle drafting head. Need to be customized for each attention
-        type.
+        Get the longest cache hit prefix of the blocks that is not longer than 
+        `max_length`. If no cache hit is found, return an empty list. 
+        If eagle is enabled, drop the last matched block to force recompute the 
+        last block to get the required hidden states for eagle drafting head. 
+        Need to be customized for each attention type.
 
         Args:
             block_hashes: The block hashes of the request.
+            max_length: The maximum length of the cache hit prefix.
+
         Returns:
             A list of cached blocks with skipped blocks replaced by null block.
             For example, sliding window manager should return a list like
@@ -226,10 +228,12 @@ class SingleTypeKVCacheManager(ABC):
 
 class FullAttentionManager(SingleTypeKVCacheManager):
 
-    def find_longest_cache_hit(
-            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+    def find_longest_cache_hit(self, block_hashes: list[BlockHashType],
+                               max_length: int) -> list[KVCacheBlock]:
         computed_blocks: list[KVCacheBlock] = []
-        for block_hash in block_hashes:
+        max_num_blocks = max_length // self.block_size
+        for i in range(max_num_blocks):
+            block_hash = block_hashes[i]
             # block_hashes is a chain of block hashes. If a block hash is not
             # in the cached_block_hash_to_id, the following block hashes are
             # not computed yet for sure.
@@ -276,19 +280,20 @@ class SlidingWindowManager(SingleTypeKVCacheManager):
             self.sliding_window_contiguous_blocks += 1
         self._null_block = block_pool.null_block
 
-    def find_longest_cache_hit(
-            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+    def find_longest_cache_hit(self, block_hashes: list[BlockHashType],
+                               max_length: int) -> list[KVCacheBlock]:
         # TODO: reduce i by sliding_window_contiguous_blocks when cache miss, to
-        # optimize the time complexity from O(len(block_hashes)) to
-        # O(len(block_hashes) / sliding_window_contiguous_blocks +
+        # optimize the time complexity from O(max_num_blocks) to
+        # O(max_num_blocks / sliding_window_contiguous_blocks +
         # sliding_window_contiguous_blocks),
         # which is good for low cache hit rate scenarios.
-        computed_blocks = [self._null_block] * len(block_hashes)
+        max_num_blocks = max_length // self.block_size
+        computed_blocks = [self._null_block] * max_num_blocks
         num_contiguous_blocks = 0
 
         match_found = False
         # Search from right to left and early stop when a match is found.
-        for i in range(len(block_hashes) - 1, -1, -1):
+        for i in range(max_num_blocks - 1, -1, -1):
             if cached_block := self.block_pool.get_cached_block(
                     block_hashes[i]):
                 computed_blocks[i] = cached_block