[claudesquad] update from 'int-minference-1' on 08 Jan 26 23:22 CST

nanovllm/kvcache/sparse/full_policy.py

@@ -25,6 +25,7 @@ class FullAttentionPolicy(SparsePolicy):
     # Full attention supports both prefill and decode
     supports_prefill = True
     supports_decode = True
+    requires_block_selection = False  # Load all blocks, no selective loading
 
     def select_blocks(
         self,

nanovllm/kvcache/sparse/minference.py

@@ -30,6 +30,7 @@ class MInferencePolicy(SparsePolicy):
 
     supports_prefill = True
     supports_decode = False  # MInference is prefill-only sparse strategy
+    requires_block_selection = False  # MInference only affects attention computation, not KV load
 
     def __init__(
         self,

nanovllm/kvcache/sparse/policy.py

@@ -77,6 +77,12 @@ class SparsePolicy(ABC):
     supports_prefill: bool = True
     supports_decode: bool = True
 
+    # Whether this policy requires selective block loading during decode
+    # If True: OffloadEngine will call select_blocks() before loading KV from CPU
+    # If False: OffloadEngine will load all blocks (select_blocks ignored for load)
+    # Example: MInference=False (only affects attention), Quest=True (affects load)
+    requires_block_selection: bool = False
+
     def initialize(
         self,
         num_layers: int,

nanovllm/kvcache/sparse/quest.py

@@ -158,6 +158,7 @@ class QuestPolicy(SparsePolicy):
     # Quest is decode-only
     supports_prefill = False
     supports_decode = True
+    requires_block_selection = True  # Quest affects KV load strategy (selective block loading)
 
     def __init__(self, config: QuestConfig):
         """