EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[HiCache] Cleaning the deprecated host memory state (#10778)

Browse Source
This commit is contained in:
Zhiqiang Xie
2025-09-24 23:43:53 -07:00
committed by GitHub
parent c1f39013b7
commit 3d40794fcf
3 changed files with 15 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/mem_cache/hiradix_cache.py
View File

@@ -305,7 +305,7 @@ class HiRadixCache(RadixCache):
def _evict_backuped(self, node: TreeNode):
# evict a node already written to host
num_evicted = self.cache_controller.evict_device(node.value, node.host_value)
num_evicted = self.cache_controller.evict_device(node.value)
assert num_evicted > 0
self.evictable_size_ -= num_evicted
node.value = None
@@ -576,8 +576,6 @@ class HiRadixCache(RadixCache):
written_indices,
hash_value[: min_completed_tokens // self.page_size],
)
if len(written_indices):
self.cache_controller.mem_pool_host.update_prefetch(written_indices)
self.cache_controller.mem_pool_host.free(host_indices[:matched_length])
self.cache_controller.append_host_mem_release(
@@ -775,7 +773,6 @@ class HiRadixCache(RadixCache):
# change the reference if the node is evicted
# this often happens in the case of KV cache recomputation
node.value = value[:prefix_len]
self.token_to_kv_pool_host.update_synced(node.host_value)
self.evictable_size_ += len(node.value)
else:
self._inc_hit_count(node, chunked)
@@ -785,7 +782,6 @@ class HiRadixCache(RadixCache):
new_node = self._split_node(node.key, node, prefix_len)
if new_node.evicted:
new_node.value = value[:prefix_len]
self.token_to_kv_pool_host.update_synced(new_node.host_value)
self.evictable_size_ += len(new_node.value)
else:
self._inc_hit_count(new_node, chunked)