[Feature] Support sequence parallel without the need for DP, 1.9%~5.0% E2E Throughput Improvement (#47070)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Signed-off-by: gcanlin <canlinguosdu@gmail.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>
This commit is contained in:
@@ -1038,7 +1038,7 @@ class FusedMoEParallelConfig:
|
||||
|
||||
@property
|
||||
def use_all2all_kernels(self):
|
||||
return self.dp_size > 1 and self.use_ep
|
||||
return self.use_ep and (self.dp_size > 1 or self.is_sequence_parallel)
|
||||
|
||||
@property
|
||||
def use_deepep_ht_kernels(self):
|
||||
|
||||
@@ -725,8 +725,8 @@ class MoERunner(MoERunnerInterface):
|
||||
@property
|
||||
def do_naive_dispatch_combine(self) -> bool:
|
||||
return (
|
||||
self.moe_config.dp_size > 1 and not self._quant_method.supports_internal_mk
|
||||
)
|
||||
self.moe_config.dp_size > 1 or self.moe_config.is_sequence_parallel
|
||||
) and not self._quant_method.supports_internal_mk
|
||||
|
||||
def _maybe_dispatch(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user