[Feature] Detect all2all peer fault with fault tolerance backend and prevent corrupted output (#43637)
Signed-off-by: fangyuchu <fangyuchu@qq.com> Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
Tyler Michael Smith
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
e840f0d3f5
commit
245888ff77
@@ -56,7 +56,7 @@ if current_platform.is_cuda_alike():
|
||||
)
|
||||
|
||||
|
||||
def _get_ep_all2all_manager(eep_stage: bool = False) -> Any:
|
||||
def get_ep_all2all_manager(eep_stage: bool = False) -> Any:
|
||||
if eep_stage:
|
||||
from vllm.distributed.elastic_ep.standby_state import get_standby_ep_group
|
||||
|
||||
@@ -146,7 +146,7 @@ def maybe_make_prepare_finalize(
|
||||
"Detected DP deployment with no --enable-expert-parallel. "
|
||||
"Falling back to AllGather+ReduceScatter dispatch/combine."
|
||||
)
|
||||
all2all_manager = _get_ep_all2all_manager(eep_stage)
|
||||
all2all_manager = get_ep_all2all_manager(eep_stage)
|
||||
return make_moe_prepare_and_finalize_naive_dp_ep(
|
||||
is_sequence_parallel=moe.moe_parallel_config.is_sequence_parallel,
|
||||
num_dispatchers=all2all_manager.world_size,
|
||||
@@ -155,7 +155,7 @@ def maybe_make_prepare_finalize(
|
||||
else:
|
||||
return make_moe_prepare_and_finalize_no_dp_ep(use_monolithic)
|
||||
|
||||
all2all_manager = _get_ep_all2all_manager(eep_stage)
|
||||
all2all_manager = get_ep_all2all_manager(eep_stage)
|
||||
|
||||
prepare_finalize: FusedMoEPrepareAndFinalize | None = None
|
||||
|
||||
|
||||
Reference in New Issue
Block a user