fix: pad dummy run query_start_loc (#44603)
Signed-off-by: UranusSeven <109661872+UranusSeven@users.noreply.github.com>
This commit is contained in:
@@ -5787,6 +5787,9 @@ class GPUModelRunner(
|
|||||||
num_scheduled_tokens, self.query_pos.np
|
num_scheduled_tokens, self.query_pos.np
|
||||||
)
|
)
|
||||||
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
|
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
|
||||||
|
self.query_start_loc.np[num_reqs + 1 : num_reqs_padded + 1].fill(
|
||||||
|
cum_num_tokens[-1]
|
||||||
|
)
|
||||||
self.query_start_loc.copy_to_gpu()
|
self.query_start_loc.copy_to_gpu()
|
||||||
|
|
||||||
# Sync block table CPU->GPU so cleared rows from
|
# Sync block table CPU->GPU so cleared rows from
|
||||||
|
|||||||
Reference in New Issue
Block a user