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

HiCache Storage: generate hash when inserting new nodes (#9053)

Browse Source
This commit is contained in:
Zhiqiang Xie
2025-08-11 14:18:59 -07:00
committed by GitHub
parent f508cd3cb7
commit 9f78f391ae
5 changed files with 37 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
python/sglang/srt/managers/cache_controller.py
View File

@@ -169,12 +169,13 @@ class StorageOperation:
host_indices: torch.Tensor,
token_ids: List[int],
last_hash: Optional[str] = None,
hash_value: Optional[List[str]] = None,
):
self.host_indices = host_indices
self.token_ids = token_ids
self.last_hash = last_hash
self.completed_tokens = 0
self.hash_value = []
self.hash_value = hash_value if hash_value is not None else []
self.id = StorageOperation.counter
StorageOperation.counter += 1
@@ -702,12 +703,12 @@ class HiCacheController:
self,
host_indices: torch.Tensor,
token_ids: List[int],
last_hash: Optional[str] = None,
hash_value: Optional[List[str]] = None,
) -> int:
"""
Write KV caches from host memory to storage backend.
"""
operation = StorageOperation(host_indices, token_ids, last_hash)
operation = StorageOperation(host_indices, token_ids, hash_value=hash_value)
self.backup_queue.put(operation)
return operation.id
@@ -762,24 +763,6 @@ class HiCacheController:
if operation is None:
continue
last_hash = operation.last_hash
tokens_to_backup = operation.token_ids
backup_hit_count = 0
remaining_tokens = len(tokens_to_backup)
hash_value = []
while remaining_tokens >= self.page_size:
last_hash = self.get_hash_str(
tokens_to_backup[
backup_hit_count : backup_hit_count + self.page_size
],
last_hash,
)
backup_hit_count += self.page_size
hash_value.append(last_hash)
remaining_tokens -= self.page_size
operation.hash_value = hash_value
if self.is_mooncake_backend():
self.mooncake_page_backup(operation)
elif self.storage_backend_type == "hf3fs":
@@ -802,7 +785,6 @@ class HiCacheController:
self.ack_backup_queue.put(
(
operation.id,
operation.hash_value[: min_completed_tokens // self.page_size],
min_completed_tokens,
)
)