[BugFix] Support EP/DP + EPLB with MTP (#25311)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
This commit is contained in:
Ilya Markov
2025-11-05 15:22:17 +00:00
committed by GitHub
co-authored by Sage Moore Tyler Michael Smith Lucas Wilkinson
parent 5d16d0fa62
commit e50c454672
27 changed files with 957 additions and 529 deletions
+2 -18
View File
@@ -98,7 +98,7 @@ class MixtralMoE(nn.Module):
self.hidden_size = hidden_size
self.ep_group = get_ep_group().device_group
self.ep_rank = self.ep_group.rank()
self.ep_rank = get_ep_group().rank_in_group
self.ep_size = self.ep_group.size()
# Expert Parallelism Load balancing settings.
@@ -546,7 +546,7 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
)
self.expert_weights = []
self.moe_layers: list[FusedMoE] = []
self.moe_layers = []
example_moe = None
for layer in self.model.layers:
@@ -572,22 +572,6 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
self.num_expert_groups = 1
self.num_shared_experts = 0
def set_eplb_state(
self,
expert_load_view: torch.Tensor,
logical_to_physical_map: torch.Tensor,
logical_replica_count: torch.Tensor,
) -> None:
for layer_idx, layer in enumerate(self.moe_layers):
# Register the expert weights.
self.expert_weights.append(layer.get_expert_weights())
layer.set_eplb_state(
moe_layer_idx=layer_idx,
expert_load_view=expert_load_view,
logical_to_physical_map=logical_to_physical_map,
logical_replica_count=logical_replica_count,
)
def update_physical_experts_metadata(
self,
num_physical_experts: int,