[CI] Fix missing tp_size attribute on RoutedExperts (#46163)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
fxmarty-amd
2026-06-21 18:46:49 -06:00
committed by GitHub
co-authored by Andreas Karatzas
parent a19ff2218a
commit 485bbe1c6f
@@ -426,6 +426,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
device = get_tp_group().device
tp_rank = get_tensor_model_parallel_rank()
tp_size = layer.moe_config.moe_parallel_config.tp_size
loaded_weight = loaded_weight.to(device)
shard_size = layer.intermediate_size_per_partition
@@ -464,9 +465,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
)
if "w13_qzeros" in weight_name:
tensor = loaded_weight.view(layer.tp_size, -1, loaded_weight.size(1))[
tp_rank
]
tensor = loaded_weight.view(tp_size, -1, loaded_weight.size(1))[tp_rank]
if shard_id == "w1":
param.data[expert_id, : shard_size // 2] = tensor
else:
@@ -474,7 +473,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
return True if return_success else None
elif "w2_qzeros" in weight_name:
param.data[expert_id] = loaded_weight.view(
loaded_weight.size(0), layer.tp_size, -1
loaded_weight.size(0), tp_size, -1
)[:, tp_rank]
return True if return_success else None
else: