minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

hicache storage backend mooncake support ascend hixl (#20016)

Browse Source
This commit is contained in:
lawtherWu
2026-04-14 20:51:06 +08:00
committed by GitHub
parent 074c2a476d
commit 454228e071
2 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py
+2 -1
View File
@@ -497,7 +497,8 @@ class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
"page_first", "page_first",
"page_first_direct", "page_first_direct",
"page_head", "page_head",
], "mooncake store storage backend only support page first or page first direct layout" "page_first_kv_spilt",
], "mooncake store storage backend only support page first, page first direct, page head and page_first_kv_spilt layout"
buffer = self.mem_pool_host.kv_buffer buffer = self.mem_pool_host.kv_buffer
try: try:
super().register_buffer(buffer) super().register_buffer(buffer)
python/sglang/srt/model_executor/model_runner.py
+3 -3
View File
@@ -449,12 +449,12 @@ class ModelRunner(ModelRunnerKVCacheMixin):
if self.device == "cpu": if self.device == "cpu":
self.init_threads_binding() self.init_threads_binding()
# Initialize MooncakeTransferEngine
self.init_shared_mooncake_transfer_engine()
# Get available memory before model loading # Get available memory before model loading
pre_model_load_memory = self.init_torch_distributed() pre_model_load_memory = self.init_torch_distributed()
# Initialize MooncakeTransferEngine
self.init_shared_mooncake_transfer_engine()
# Init forward stream for overlap schedule # Init forward stream for overlap schedule
self.forward_stream = torch.get_device_module(self.device).Stream() self.forward_stream = torch.get_device_module(self.device).Stream()