correctness!

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

Author: robertgshaw2-redhat

tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh

@@ -4,7 +4,7 @@ set -xe
 # Models to run
 MODELS=(
     "Qwen/Qwen3-0.6B"
-    # "deepseek-ai/deepseek-vl2-tiny"
+    "deepseek-ai/deepseek-vl2-tiny"
 )
 
 # Number of prefill and decode instances to create

vllm/v1/core/kv_cache_manager.py

@@ -301,16 +301,20 @@ def allocate_slots(
         if not self.enable_caching:
             return KVCacheBlocks(new_blocks)
 
-        if not delay_cache_blocks:
+        if delay_cache_blocks:
+            # P/D: delay caching the blocks if we need to wait for the
+            # KVs to be recved from remote, but update num_cached_block
+            # with the prefix cache hits to avoid double caching later.
+            assert request.request_id not in self.num_cached_block
+            self.num_cached_block[request.request_id] = len(
+                new_computed_block_list)
+        else:
             self.cache_blocks(
                 request=request,
                 num_tokens=num_tokens,
                 num_computed_tokens=num_computed_tokens,
                 new_computed_block_list=new_computed_block_list,
             )
-        else:
-            self.num_cached_block[request.request_id] = len(
-                new_computed_block_list)
 
         return KVCacheBlocks(new_blocks)