bugfix:fix unifiedcache c128 radix cache management (#39426)

python/sglang/srt/hardware_backend/npu/dsv4/c128_sidecar_component.py

@@ -329,6 +329,14 @@ class C128SidecarComponent(TreeComponent):
         ].clone()
         return cache_len + 1 if self.tree_core.is_eagle and cache_len > 0 else cache_len
 
+    def floor_cache_len(self, cache_len: int) -> int:
+        logical_len = cache_len
+        if self.tree_core.is_eagle and logical_len > 0:
+            logical_len -= 1
+        group_tokens = 128 * self.allocator.c128_attn_allocator.page_size
+        floored = logical_len // group_tokens * group_tokens
+        return floored + 1 if self.tree_core.is_eagle and floored > 0 else floored
+
     def apply_component_action(self, action: ComponentAction) -> None:
         if isinstance(action, FreeComponentDeviceSlot):
             for page_ids in action.indices:

python/sglang/srt/mem_cache/unified_cache/components/base.py

@@ -384,6 +384,11 @@ class TreeComponent(ABC):
         - Mamba: performs the copy-on-write into a per-request slot."""
         return result
 
+    def floor_cache_len(self, cache_len: int) -> int:
+        """Constrain the combined effective cache length after every
+        component's `prepare_for_caching_req` truncation has been min'd."""
+        return cache_len
+
     def update_component_on_insert_overlap(
         self,
         node: UnifiedTreeNode,

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -995,6 +995,8 @@ class UnifiedRadixCache(BasePrefixCache):
                 )
                 if cl is not None:
                     effective_cache_len = min(effective_cache_len, cl)
+            for comp in self._components_tuple:
+                effective_cache_len = comp.floor_cache_len(effective_cache_len)
 
             # Truncate if needed; the tail free is deferred and batched with
             # the unaligned tail below so a shared boundary page is emitted once.
@@ -1068,7 +1070,12 @@ class UnifiedRadixCache(BasePrefixCache):
             )
             ranges = [(free_from, len(kv_indices))]
             if tail_free_start is not None:
-                ranges.append((tail_free_start, len(kv_indices_full)))
+                if free_from < len(kv_indices) and tail_free_start <= len(kv_indices):
+                    # The two halves touch at the truncation boundary and share
+                    # that page; free the union as one range.
+                    ranges[0] = (free_from, len(kv_indices_full))
+                else:
+                    ranges.append((tail_free_start, len(kv_indices_full)))
             self.free_kv_row(req.kv, ranges)
         else:
             self.free_kv_row(req.kv, [(req.kv.cache_protected_len, kv_len_to_handle)])
@@ -1131,6 +1138,9 @@ class UnifiedRadixCache(BasePrefixCache):
             if cl is not None:
                 effective_cache_len = min(effective_cache_len, cl)
 
+        for comp in self._components_tuple:
+            effective_cache_len = comp.floor_cache_len(effective_cache_len)
+
         radix_key = RadixKey(
             token_ids[:effective_cache_len],
             req.extra_key,