mirror of
https://github.com/ml-explore/mlx-examples.git
synced 2025-10-24 06:28:07 +08:00
fix deepseek sharding (#1242)
This commit is contained in:
@@ -386,6 +386,7 @@ class DeepseekV2Model(nn.Module):
|
||||
self.num_layers = layers_per_rank
|
||||
self.layers = self.layers[: self.end_idx]
|
||||
self.layers[: self.start_idx] = [None] * self.start_idx
|
||||
self.num_layers = len(self.layers) - self.start_idx
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
|
@@ -397,12 +397,11 @@ class DeepseekV3Model(nn.Module):
|
||||
layers_per_rank = (
|
||||
len(self.layers) + self.pipeline_size - 1
|
||||
) // self.pipeline_size
|
||||
start = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
|
||||
self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
|
||||
self.end_idx = self.start_idx + layers_per_rank
|
||||
self.num_layers = layers_per_rank
|
||||
self.layers = self.layers[: self.end_idx]
|
||||
self.layers[: self.start_idx] = [None] * self.start_idx
|
||||
self.num_layers = len(self.layers) - self.start_idx
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
|
Reference in New Issue
Block a user