fused_moe: treat NIXL EP as batched experts (#40412)
Signed-off-by: Itay Alroy <ialroy@nvidia.com>
This commit is contained in:
@@ -993,6 +993,10 @@ class FusedMoEParallelConfig:
|
||||
def use_batched_activation_format(self):
|
||||
return self.use_deepep_ll_kernels or self.use_nixl_ep_kernels
|
||||
|
||||
@property
|
||||
def needs_round_robin_routing_tables(self):
|
||||
return self.use_deepep_ll_kernels or self.use_nixl_ep_kernels
|
||||
|
||||
@property
|
||||
def use_ag_rs_all2all_kernels(self):
|
||||
return (
|
||||
@@ -1294,3 +1298,7 @@ class FusedMoEConfig:
|
||||
@property
|
||||
def use_nixl_ep_kernels(self):
|
||||
return self.moe_parallel_config.use_nixl_ep_kernels
|
||||
|
||||
@property
|
||||
def needs_round_robin_routing_tables(self):
|
||||
return self.moe_parallel_config.needs_round_robin_routing_tables
|
||||
|
||||
@@ -180,8 +180,7 @@ def determine_expert_placement_strategy(
|
||||
return "linear"
|
||||
if (
|
||||
moe_parallel_config.use_all2all_kernels
|
||||
and not moe_parallel_config.use_deepep_ll_kernels
|
||||
and not moe_parallel_config.use_nixl_ep_kernels
|
||||
and not moe_parallel_config.needs_round_robin_routing_tables
|
||||
):
|
||||
logger.warning(
|
||||
"Round-robin expert placement currently only supports "
|
||||
@@ -687,8 +686,7 @@ class FusedMoE(PluggableLayer):
|
||||
# Currently routing_tables only needed for round-robin expert placement
|
||||
# with DeepEP-ll or NIXL EP all2all backends.
|
||||
if self.expert_placement_strategy != "round_robin" or (
|
||||
not self.moe_parallel_config.use_deepep_ll_kernels
|
||||
and not self.moe_parallel_config.use_nixl_ep_kernels
|
||||
not self.moe_parallel_config.needs_round_robin_routing_tables
|
||||
):
|
||||
return None
|
||||
|
||||
|
||||
@@ -884,7 +884,7 @@ def make_mxfp4_moe_kernel(
|
||||
experts,
|
||||
shared_experts=(
|
||||
shared_experts
|
||||
if moe_config.moe_parallel_config.use_deepep_ll_kernels
|
||||
if moe_config.moe_parallel_config.use_batched_activation_format
|
||||
else None
|
||||
),
|
||||
inplace=(
|
||||
|
||||
@@ -168,10 +168,7 @@ def select_nvfp4_moe_backend(
|
||||
NvFp4MoeBackend.EMULATION,
|
||||
]
|
||||
|
||||
# NOTE(rob): this is kind of a hack. We need to peak into
|
||||
# the prepare-finalize selection to determine if we are using
|
||||
# the batched or standard expert format.
|
||||
use_batched = config.moe_parallel_config.use_deepep_ll_kernels
|
||||
use_batched = config.moe_parallel_config.use_batched_activation_format
|
||||
activation_format = (
|
||||
mk.FusedMoEActivationFormat.BatchedExperts
|
||||
if use_batched
|
||||
|
||||
Reference in New Issue
Block a user