fuse allreduce and residual_rmsnorm (#8731)
This commit is contained in:
@@ -1294,6 +1294,7 @@ class RowParallelLinear(LinearBase):
|
||||
with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
|
||||
output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
|
||||
sm.tag(output_parallel)
|
||||
|
||||
if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
|
||||
output = tensor_model_parallel_all_reduce(output_parallel)
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user