Fix NPUMHATokenToKVPool missing k_data_ptrs/v_data_ptrs (#35820)
This commit is contained in:
@@ -128,12 +128,16 @@ class MHATokenToKVPoolHost(HostKVCache):
|
|||||||
)
|
)
|
||||||
if self.mtp_draft_device_pools:
|
if self.mtp_draft_device_pools:
|
||||||
device_pools = (self.device_pool, *self.mtp_draft_device_pools)
|
device_pools = (self.device_pool, *self.mtp_draft_device_pools)
|
||||||
self.packed_device_k_data_ptrs = torch.cat(
|
if not _is_npu:
|
||||||
[pool.k_data_ptrs for pool in device_pools]
|
self.packed_device_k_data_ptrs = torch.cat(
|
||||||
)
|
[pool.k_data_ptrs for pool in device_pools]
|
||||||
self.packed_device_v_data_ptrs = torch.cat(
|
)
|
||||||
[pool.v_data_ptrs for pool in device_pools]
|
self.packed_device_v_data_ptrs = torch.cat(
|
||||||
)
|
[pool.v_data_ptrs for pool in device_pools]
|
||||||
|
)
|
||||||
|
else:
|
||||||
|
self.packed_device_k_data_ptrs = None
|
||||||
|
self.packed_device_v_data_ptrs = None
|
||||||
self.packed_device_k_buffers = [
|
self.packed_device_k_buffers = [
|
||||||
buffer for pool in device_pools for buffer in pool.k_buffer
|
buffer for pool in device_pools for buffer in pool.k_buffer
|
||||||
]
|
]
|
||||||
|
|||||||
Reference in New Issue
Block a user