[WIP] need refactor.

nanovllm/kvcache/sparse/quest.py

@@ -11,7 +11,7 @@ import logging
 import torch
 from dataclasses import dataclass
 from typing import List, Tuple, Optional
-from .policy import SparsePolicy, PolicyContext
+from .policy import AttentionPolicy, PolicyContext
 
 logger = logging.getLogger(__name__)
 
@@ -137,7 +137,7 @@ class QuestConfig:
     """Always include this many recent blocks (last N blocks), in addition to Top-K."""
 
 
-class QuestPolicy(SparsePolicy):
+class QuestPolicy(AttentionPolicy):
     """
     Quest-style Top-K block selection using min/max key bounds.
 
@@ -317,6 +317,25 @@ class QuestPolicy(SparsePolicy):
         if self.metadata is not None:
             self.metadata.reset()
 
+    def compute_prefill(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        layer_id: int,
+        softmax_scale: float,
+    ) -> torch.Tensor:
+        """
+        Quest does not support prefill - raises error.
+
+        Quest is a decode-only policy for selective block loading.
+        For prefill, use FullAttentionPolicy or XAttentionPolicy.
+        """
+        raise NotImplementedError(
+            "QuestPolicy does not support prefill. "
+            "Use FullAttentionPolicy or XAttentionPolicy for prefill."
+        )
+
     def __repr__(self) -> str:
         return (
             f"QuestPolicy(topk={self.config.topk_blocks}, "