Support TP in attention for two batch overlap (#6634)
This commit is contained in:
@@ -1613,6 +1613,9 @@ class DeepseekV2Model(nn.Module):
|
||||
forward_batch=forward_batch,
|
||||
hidden_states=hidden_states,
|
||||
residual=residual,
|
||||
input_data_scatter_mode=self.layers[
|
||||
normal_num_layers - 1
|
||||
].layer_scatter_modes.layer_output_mode,
|
||||
zero_allocator=zero_allocator,
|
||||
)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user