mlp_only_layers is more flexible than decoder_sparse_step (#30552)

* force back to commit ba40a21 and fix workflow errors

* match the review suggestions

* fix ci errors

* fix CI

* fix ci, format code

* fix ci, ruff format

* fix ci, ruff format again

* Update src/transformers/models/qwen2_moe/configuration_qwen2_moe.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/qwen2_moe/configuration_qwen2_moe.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/qwen2_moe/configuration_qwen2_moe.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* solve this warning: Default Argument Value is mutable

---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/qwen2_moe/configuration_qwen2_moe.py

@@ -91,6 +91,10 @@ class Qwen2MoeConfig(PretrainedConfig):
             allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
         router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
             The aux loss factor for the total loss.
+        mlp_only_layers (`List[int]`, *optional*, defaults to `[]`):
+            Indicate which layers use Qwen2MoeMLP rather than Qwen2MoeSparseMoeBlock
+            The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
+            If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
 
     ```python
     >>> from transformers import Qwen2MoeModel, Qwen2MoeConfig
@@ -135,6 +139,7 @@ class Qwen2MoeConfig(PretrainedConfig):
         norm_topk_prob=False,
         output_router_logits=False,
         router_aux_loss_coef=0.001,
+        mlp_only_layers=None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -164,6 +169,7 @@ class Qwen2MoeConfig(PretrainedConfig):
         self.norm_topk_prob = norm_topk_prob
         self.output_router_logits = output_router_logits
         self.router_aux_loss_coef = router_aux_loss_coef
+        self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers
 
         super().__init__(
             tie_word_embeddings=tie_word_embeddings,

src/transformers/models/qwen2_moe/modeling_qwen2_moe.py

@@ -17,7 +17,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" PyTorch Qwen2MoE model."""
+"""PyTorch Qwen2MoE model."""
+
 import inspect
 import math
 import warnings
@@ -861,7 +862,9 @@ class Qwen2MoeDecoderLayer(nn.Module):
 
         self.self_attn = QWEN2MOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
 
-        if config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0:
+        if (layer_idx not in config.mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
             self.mlp = Qwen2MoeSparseMoeBlock(config)
         else:
             self.mlp = Qwen2MoeMLP(config, intermediate_size=config.intermediate_size)