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

bugfix revise interface get cpu copy for npu mem pool to align with gpu (#29146)

Browse Source
This commit is contained in:
McZyWu
2026-06-29 19:30:14 +08:00
committed by GitHub
parent 489017b3d6
commit bb7d3440b5
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
+2 -2
View File
@@ -506,7 +506,7 @@ class NPUMLATokenToKVPool(MLATokenToKVPool):
out.append(layer_chunks) out.append(layer_chunks)
return out return out
def get_cpu_copy(self, indices): def get_cpu_copy(self, indices, mamba_indices=None):
torch.npu.synchronize() torch.npu.synchronize()
buf_of_layers = [] buf_of_layers = []
has_ik = self.index_head_dim is not None has_ik = self.index_head_dim is not None
@@ -524,7 +524,7 @@ class NPUMLATokenToKVPool(MLATokenToKVPool):
torch.npu.synchronize() torch.npu.synchronize()
return kv_cache_cpu return kv_cache_cpu
def load_cpu_copy(self, kv_cache_cpu, indices): def load_cpu_copy(self, kv_cache_cpu, indices, mamba_indices=None):
torch.npu.synchronize() torch.npu.synchronize()
chunk_size = self.cpu_offloading_chunk_size chunk_size = self.cpu_offloading_chunk_size
has_ik = self.index_head_dim is not None has_ik = self.index_head_dim is not None