MixtralSparseMoeBlock: add gate jitter (#29865)

This commit adds gate jitter to MixtralSparseMoeBlock's input data
before passing it through the MoE layer, if turned on.

src/transformers/models/mixtral/configuration_mixtral.py

@@ -92,6 +92,8 @@ class MixtralConfig(PretrainedConfig):
             allow the model to output the auxiliary loss. See [here]() for more details
         router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
             The aux loss factor for the total loss.
+        router_jitter_noise (`float`, *optional*, defaults to 0.0):
+            Amount of noise to add to the router.
 
     ```python
     >>> from transformers import MixtralModel, MixtralConfig
@@ -133,6 +135,7 @@ class MixtralConfig(PretrainedConfig):
         num_local_experts=8,
         output_router_logits=False,
         router_aux_loss_coef=0.001,
+        router_jitter_noise=0.0,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -159,6 +162,7 @@ class MixtralConfig(PretrainedConfig):
         self.num_local_experts = num_local_experts
         self.output_router_logits = output_router_logits
         self.router_aux_loss_coef = router_aux_loss_coef
+        self.router_jitter_noise = router_jitter_noise
         super().__init__(
             pad_token_id=pad_token_id,
             bos_token_id=bos_token_id,

src/transformers/models/mixtral/modeling_mixtral.py

@@ -837,9 +837,14 @@ class MixtralSparseMoeBlock(nn.Module):
 
         self.experts = nn.ModuleList([MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
 
+        # Jitter parameters
+        self.jitter_noise = config.router_jitter_noise
+
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         """ """
         batch_size, sequence_length, hidden_dim = hidden_states.shape
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
         hidden_states = hidden_states.view(-1, hidden_dim)
         # router_logits: (batch * sequence_length, n_experts)
         router_logits = self.gate(hidden_states)

tests/models/mixtral/test_modeling_mixtral.py

@@ -42,7 +42,6 @@ if is_torch_available():
 
 
 class MixtralModelTester:
-    # Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.__init__
     def __init__(
         self,
         parent,
@@ -69,6 +68,7 @@ class MixtralModelTester:
         num_choices=4,
         pad_token_id=0,
         scope=None,
+        router_jitter_noise=0.1,
     ):
         self.parent = parent
         self.batch_size = batch_size
@@ -94,6 +94,7 @@ class MixtralModelTester:
         self.num_choices = num_choices
         self.pad_token_id = pad_token_id
         self.scope = scope
+        self.router_jitter_noise = router_jitter_noise
 
     # Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.prepare_config_and_inputs
     def prepare_config_and_inputs(self):
@@ -137,6 +138,7 @@ class MixtralModelTester:
             pad_token_id=self.pad_token_id,
             num_experts_per_tok=2,
             num_local_experts=2,
+            router_jitter_noise=self.router_jitter_noise,
         )
 
     # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Mixtral