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

18
vllm/model_executor/models/openpangu.py
View File

@@ -1009,7 +1009,7 @@ class OpenPanguMoEModel(OpenPanguModelBase, MixtureOfExperts):
self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
self.num_expert_groups = 1
self.moe_layers: list[SharedFusedMoE] = []
self.moe_layers = []
example_moe = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
@@ -1031,22 +1031,6 @@ class OpenPanguMoEModel(OpenPanguModelBase, MixtureOfExperts):
self.n_shared_experts = example_moe.n_shared_experts
self.num_redundant_experts = example_moe.n_redundant_experts
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,