EAGLE cache fix for HiCache (#11215)

python/sglang/srt/managers/schedule_policy.py

@@ -583,6 +583,7 @@ class PrefillAdder:
                 req.prefix_indices = torch.cat([req.prefix_indices, new_indices])
                 req.extend_input_len = len(req.fill_ids) - len(req.prefix_indices)
                 prefix_len = len(req.prefix_indices)
+                req.last_matched_prefix_len = prefix_len
 
             input_tokens = self.ceil_paged_tokens(req.extend_input_len)
 

python/sglang/srt/managers/scheduler.py

@@ -762,6 +762,7 @@ class Scheduler(
                     hicache_storage_prefetch_policy=server_args.hicache_storage_prefetch_policy,
                     model_name=server_args.served_model_name,
                     storage_backend_extra_config=server_args.hicache_storage_backend_extra_config,
+                    is_eagle=self.spec_algorithm.is_eagle(),
                 )
                 self.tp_worker.register_hicache_layer_transfer_counter(
                     self.tree_cache.cache_controller.layer_done_counter

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -44,6 +44,7 @@ class HiRadixCache(RadixCache):
         hicache_storage_prefetch_policy: Optional[str] = "best_effort",
         model_name: Optional[str] = None,
         storage_backend_extra_config: Optional[str] = None,
+        is_eagle: bool = False,
     ):
 
         if hicache_io_backend == "direct":
@@ -135,6 +136,7 @@ class HiRadixCache(RadixCache):
             page_size,
             disable=False,
             eviction_policy=eviction_policy,
+            is_eagle=is_eagle,
         )
 
     def _parse_storage_backend_extra_config(
@@ -658,6 +660,7 @@ class HiRadixCache(RadixCache):
 
     def match_prefix(self, key: RadixKey, **kwargs):
         empty_value = torch.empty((0,), dtype=torch.int64, device=self.device)
+        key.token_ids = self.key_convert_fn(key.token_ids)
         if self.disable or len(key) == 0:
             return MatchResult(
                 device_indices=empty_value,
@@ -820,9 +823,15 @@ class HiRadixCache(RadixCache):
         return new_node
 
     def insert(self, key: RadixKey, value=None, chunked=False):
+        key.token_ids = self.key_convert_fn(key.token_ids)
+
         if len(key) == 0:
             return 0
 
+        if self.is_eagle and value is not None:
+            # Make sure the value len equal to the EAGLE bigram key len
+            value = value[: len(key)]
+
         node = self.root_node
         child_key = self.get_child_key_fn(key)
         total_prefix_length = 0