qwen3moe support two batch overlap (#6598)
This commit is contained in:
@@ -356,14 +356,14 @@ def model_forward_maybe_tbo(
|
||||
forward_batch: ForwardBatch,
|
||||
hidden_states: torch.Tensor,
|
||||
residual: Optional[torch.Tensor],
|
||||
zero_allocator: BumpAllocator,
|
||||
zero_allocator: Optional[BumpAllocator] = None,
|
||||
):
|
||||
inputs = dict(
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
forward_batch=forward_batch,
|
||||
residual=residual,
|
||||
zero_allocator=zero_allocator,
|
||||
**(dict(zero_allocator=zero_allocator) if zero_allocator is not None else {}),
|
||||
)
|
||||
operations_strategy = OperationsStrategy.init_new_tbo(
|
||||
layers, forward_batch.global_forward_mode
|
||||
@@ -401,7 +401,7 @@ def _model_forward_tbo_split_inputs(
|
||||
residual: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
forward_batch: ForwardBatch,
|
||||
zero_allocator: BumpAllocator,
|
||||
zero_allocator: Optional[BumpAllocator] = None,
|
||||
) -> List[Dict]:
|
||||
return [
|
||||
dict(
|
||||
@@ -412,7 +412,11 @@ def _model_forward_tbo_split_inputs(
|
||||
output_forward_batch=output_forward_batch,
|
||||
tbo_subbatch_index=tbo_subbatch_index,
|
||||
),
|
||||
zero_allocator=zero_allocator,
|
||||
**(
|
||||
dict(zero_allocator=zero_allocator)
|
||||
if zero_allocator is not None
|
||||
else {}
|
||||
),
|
||||
)
|
||||
for tbo_subbatch_index, output_forward_batch in enumerate(
|
||||
forward_batch.tbo_children
|
||||
|
||||
Reference in New Issue
Block a user