improve

vllm/v1/core/kv_cache_manager.py

@@ -70,10 +70,7 @@ def __init__(
         self.req_to_blocks: Dict[str, List[KVCacheBlock]] = {}
 
         # Prefix cache metrics.
-        self.prefix_caching_metrics: PrefixCachingMetrics = {
-            "query_total": 0,
-            "query_hit": 0,
-        }
+        self.prefix_caching_metrics = PrefixCachingMetrics()
 
     @property
     def usage(self) -> float:
@@ -85,21 +82,14 @@ def usage(self) -> float:
         return 1.0 - (self.free_block_queue.num_free_blocks /
                       self.num_gpu_blocks)
 
-    def get_and_reset_prefix_cache_hit_rate(self) -> float:
-        """Get the overall hit rate of prefix caching and reset
-        the metrics.
+    @property
+    def prefix_cache_hit_rate(self) -> float:
+        """Get the prefix caching hit rate.
 
         Returns:
-            The hit rate of prefix caching (between 0.0 and 1.0).
+            The prefix caching hit rate.
         """
-        hit_rate = 0.0
-        if self.prefix_caching_metrics["query_total"] > 0:
-            hit_rate = self.prefix_caching_metrics[
-                "query_hit"] / self.prefix_caching_metrics["query_total"]
-
-        self.prefix_caching_metrics["query_hit"] = 0
-        self.prefix_caching_metrics["query_total"] = 0
-        return hit_rate
+        return self.prefix_caching_metrics.hit_rate
 
     def get_computed_blocks(
             self, request: Request) -> Tuple[List[KVCacheBlock], int]:
@@ -136,8 +126,10 @@ def get_computed_blocks(
             else:
                 break
 
-        self.prefix_caching_metrics["query_total"] += len(block_hashes)
-        self.prefix_caching_metrics["query_hit"] += len(computed_blocks)
+        self.prefix_caching_metrics.add_request_query(
+            num_queries=len(block_hashes),
+            num_hits=len(computed_blocks),
+        )
 
         # NOTE(woosuk): Since incomplete blocks are not eligible for
         # sharing, `num_computed_tokens` is always a multiple of
@@ -367,6 +359,9 @@ def reset_prefix_cache(self) -> bool:
         for block in self.block_pool:
             block.reset_hash()
 
+        # Reset the prefix caching metrics.
+        self.prefix_caching_metrics.reset()
+
         logger.info("Successfully reset prefix cache")
         return True
 

vllm/v1/core/kv_cache_utils.py

@@ -1,7 +1,8 @@
 """KV-Cache Utilities."""
+from collections import deque
 from collections.abc import Sequence
 from dataclasses import dataclass
-from typing import Any, List, NamedTuple, Optional, Tuple, TypedDict
+from typing import Any, List, NamedTuple, Optional, Tuple
 
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
@@ -27,14 +28,52 @@ class BlockHashType(NamedTuple):
     extra_keys: Optional[Any] = None
 
 
-class PrefixCachingMetrics(TypedDict):
-    """Metrics for prefix caching."""
+class PrefixCachingMetrics:
+    """Metrics for prefix caching with a hit rate of the most recent N requests.
 
-    query_total: int
-    """The total number of queries."""
+    Args:
+        interval: The number of the most recent requests to aggregate.
+            Defaults to 1000.
+    """
+
+    def __init__(self, interval: int = 1000):
+        self.interval = interval
+        self.aggregated_query_total = 0
+        self.aggregated_query_hit = 0
+        self.request_queries: deque[Tuple[int, int]] = deque()
 
-    query_hit: int
-    """The number of queries that hit the prefix cache."""
+    def add_request_query(self, num_queries: int, num_hits: int):
+        """Add a request to the metrics. This function is called when
+        a new request is being scheduled and is looking for computed blocks.
+        When there are more than `interval` requests, the oldest request
+        is removed from the metrics.
+
+        Args:
+            num_queries: The number of queries in the request.
+            num_hits: The number of hits in the request.
+        """
+
+        self.request_queries.append((num_queries, num_hits))
+        if len(self.request_queries) > self.interval:
+            old_num_queries, old_num_hits = self.request_queries.popleft()
+            self.aggregated_query_total -= old_num_queries
+            self.aggregated_query_hit -= old_num_hits
+
+        self.aggregated_query_total += num_queries
+        self.aggregated_query_hit += num_hits
+
+    def reset(self):
+        """Reset the metrics."""
+        self.aggregated_query_total = 0
+        self.aggregated_query_hit = 0
+        self.request_queries.clear()
+
+    @property
+    def hit_rate(self) -> float:
+        """Calculate the hit rate for the past N requests."""
+        if self.aggregated_query_total == 0:
+            return 0.0
+        return self.aggregated_query_hit / self.aggregated_query_total
 
 
 @dataclass

vllm/v1/core/scheduler.py

@@ -546,7 +546,7 @@ def make_stats(self) -> SchedulerStats:
             num_waiting_reqs=len(self.waiting),
             gpu_cache_usage=self.kv_cache_manager.usage,
             gpu_prefix_cache_hit_rate=self.kv_cache_manager.
-            get_and_reset_prefix_cache_hit_rate(),
+            prefix_cache_hit_rate,
         )