[compile] Fix fake kernel return dtype (#50444)
Signed-off-by: Richard Zou <zou3519@gmail.com> Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
@@ -794,7 +794,7 @@ def unified_kv_cache_update(
|
||||
layer_slot_mapping,
|
||||
)
|
||||
|
||||
return torch.empty(0, device=kv_cache.device, dtype=kv_cache.dtype)
|
||||
return key.new_empty(0)
|
||||
|
||||
|
||||
def unified_kv_cache_update_fake(
|
||||
|
||||
Reference in New Issue
Block a user