Fix master CI for DeepSeek (#6447)
This commit is contained in:
@@ -141,6 +141,7 @@ class EPMoE(torch.nn.Module):
|
||||
top_k: int,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
layer_id: int,
|
||||
params_dtype: Optional[torch.dtype] = None,
|
||||
renormalize: bool = True,
|
||||
use_grouped_topk: bool = False,
|
||||
@@ -164,6 +165,7 @@ class EPMoE(torch.nn.Module):
|
||||
)
|
||||
self.tp_rank = get_tensor_model_parallel_rank()
|
||||
|
||||
self.layer_id = layer_id
|
||||
self.num_experts = num_experts
|
||||
assert self.num_experts % self.tp_size == 0
|
||||
self.num_experts_per_partition = self.num_experts // self.tp_size
|
||||
@@ -837,6 +839,7 @@ class DeepEPMoE(EPMoE):
|
||||
top_k: int,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
layer_id: int,
|
||||
params_dtype: Optional[torch.dtype] = None,
|
||||
renormalize: bool = True,
|
||||
use_grouped_topk: bool = False,
|
||||
@@ -856,6 +859,7 @@ class DeepEPMoE(EPMoE):
|
||||
top_k,
|
||||
hidden_size,
|
||||
intermediate_size,
|
||||
layer_id,
|
||||
params_dtype,
|
||||
renormalize,
|
||||
use_grouped_topk,
|
||||
|
||||
Reference in New Issue
Block a user