[feat] Added Quest Sparsity Policy.

nanovllm/kvcache/sparse/__init__.py

@@ -19,7 +19,8 @@ Usage:
             return available_blocks[:5]  # Just first 5 blocks
 """
 
-from nanovllm.kvcache.sparse.policy import SparsePolicy, PolicyContext, SparsePolicyType
+from nanovllm.config import SparsePolicyType
+from nanovllm.kvcache.sparse.policy import SparsePolicy, PolicyContext
 from nanovllm.kvcache.sparse.full_policy import FullAttentionPolicy
 from nanovllm.kvcache.sparse.quest import QuestPolicy, QuestConfig, BlockMetadataManager
 

nanovllm/kvcache/sparse/policy.py

@@ -7,15 +7,11 @@ from CPU for each query chunk during chunked attention computation.
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
-from enum import Enum, auto
 from typing import List, Optional, Any
 import torch
 
-
-class SparsePolicyType(Enum):
-    """Built-in sparse attention policy types."""
-    FULL = auto()   # prefill + decode
-    QUEST = auto()  # decode only
+# Import SparsePolicyType from config to avoid circular imports
+from nanovllm.config import SparsePolicyType
 
 
 @dataclass