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

fix mooncake store mla zero copy meta (#9678)

Browse Source
This commit is contained in:
huangtingwei
2025-08-27 15:43:16 +08:00
committed by GitHub
parent a3aee7c377
commit ae7428a8a7
1 changed files with 1 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
python/sglang/srt/mem_cache/memory_pool_host.py
View File

@@ -705,7 +705,6 @@ class MLATokenToKVPoolHost(HostKVCache):
raise ValueError(f"Unsupported layout: {self.layout}") raise ValueError(f"Unsupported layout: {self.layout}")
def get_buffer_meta(self, keys, indices): def get_buffer_meta(self, keys, indices):
local_rank = get_tensor_model_parallel_rank()
ptr_list = [] ptr_list = []
key_list = [] key_list = []
kv_buffer_data_ptr = self.kv_buffer.data_ptr() kv_buffer_data_ptr = self.kv_buffer.data_ptr()
@@ -719,7 +718,7 @@ class MLATokenToKVPoolHost(HostKVCache):
) )
ptr_list.append(k_ptr) ptr_list.append(k_ptr)
key_ = keys[index // self.page_size] key_ = keys[index // self.page_size]
key_list.append(f"{key_}_{local_rank}_k") key_list.append(f"{key_}_k")
element_size = ( element_size = (
self.layer_num self.layer_num
* self.dtype.itemsize * self.dtype.itemsize