biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
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 16:22:17 +01:00
committed by GitHub
parent 5d16d0fa62
commit e50c454672
27 changed files with 957 additions and 529 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/model_executor/models/lfm2_moe.py
View File

@@ -105,7 +105,7 @@ class Lfm2MoeSparseMoeBlock(nn.Module):
self.routed_scaling_factor = config.routed_scaling_factor
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()
self.n_routed_experts = config.num_experts
@@ -707,7 +707,7 @@ class Lfm2MoeForCausalLM(
# Set MoE hyperparameters
self.expert_weights = []
self.moe_layers: list[FusedMoE] = []
self.moe_layers = []
example_layer = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
@@ -737,22 +737,6 @@ class Lfm2MoeForCausalLM(
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
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,