fix gpt_oss pp>1 with ep (#46441)
Signed-off-by: mayuyuace <qiming1.zhang@intel.com>
This commit is contained in:
@@ -1078,7 +1078,7 @@ class GptOssModel(nn.Module, EagleModelMixin):
|
||||
head_start = tp_rank * heads_per_rank
|
||||
|
||||
ep_size = get_ep_group().world_size
|
||||
ep_rank = get_ep_group().rank
|
||||
ep_rank = get_ep_group().rank_in_group
|
||||
num_experts = self.config.num_local_experts
|
||||
experts_per_rank = num_experts // ep_size
|
||||
ep_rank_start = ep_rank * experts_per_rank
|
||||
|
||||
Reference in New Issue
Block a user