[NPU] Fix disaggregation metadata buffer bootstrap_room_dtype for npu backend (#19423)
This commit is contained in:
@@ -98,10 +98,13 @@ class MetadataBuffers:
|
|||||||
custom_mem_pool: torch.cuda.MemPool = None,
|
custom_mem_pool: torch.cuda.MemPool = None,
|
||||||
):
|
):
|
||||||
self.custom_mem_pool = custom_mem_pool
|
self.custom_mem_pool = custom_mem_pool
|
||||||
|
bootstrap_room_dtype = torch.uint64
|
||||||
device = "cpu"
|
device = "cpu"
|
||||||
if is_npu():
|
if is_npu():
|
||||||
# For ascend backend, output tokens are placed in the NPU and will be transferred by D2D channel.
|
# For ascend backend, output tokens are placed in the NPU and will be transferred by D2D channel.
|
||||||
device = "npu"
|
device = "npu"
|
||||||
|
# TODO: Fix me when npu backend supports torch.uint64
|
||||||
|
bootstrap_room_dtype = torch.int64
|
||||||
elif self.custom_mem_pool:
|
elif self.custom_mem_pool:
|
||||||
# TODO(shangming): Fix me (use 'cuda') when nvlink_transport of Mooncake is bug-free
|
# TODO(shangming): Fix me (use 'cuda') when nvlink_transport of Mooncake is bug-free
|
||||||
device = "cpu"
|
device = "cpu"
|
||||||
@@ -144,7 +147,7 @@ class MetadataBuffers:
|
|||||||
)
|
)
|
||||||
# Request validation: store bootstrap_room to detect metadata corruption
|
# Request validation: store bootstrap_room to detect metadata corruption
|
||||||
self.bootstrap_room = torch.zeros(
|
self.bootstrap_room = torch.zeros(
|
||||||
(size, 8), dtype=torch.uint64, device=device
|
(size, 8), dtype=bootstrap_room_dtype, device=device
|
||||||
)
|
)
|
||||||
|
|
||||||
def get_buf_infos(self):
|
def get_buf_infos(self):
|
||||||
|
|||||||
Reference in New Issue
Block a user