Cast fp32 routing weights to bf16 in the Kimi-K3 fused finalize (#38588)
This commit is contained in:
@@ -371,6 +371,8 @@ def finalize_all_reduce_push_norm(
|
|||||||
|
|
||||||
state = _get_state()
|
state = _get_state()
|
||||||
assert state is not None
|
assert state is not None
|
||||||
|
if expert_weights.dtype != torch.bfloat16:
|
||||||
|
expert_weights = expert_weights.to(torch.bfloat16)
|
||||||
return mod.finalize_all_reduce_push_norm(
|
return mod.finalize_all_reduce_push_norm(
|
||||||
state.world_size,
|
state.world_size,
|
||||||
out,
|
out,
|
||||||
|
|||||||
Reference in New Issue
Block a user