Fix code quality - apply black and ruff formatting

Author: akacmazz

src/transformers/models/mixtral/modular_mixtral.py

@@ -86,7 +86,9 @@ def load_balancing_loss_func(
 
     if isinstance(gate_logits, tuple):
         compute_device = gate_logits[0].device
-        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+        concatenated_gate_logits = torch.cat(
+            [layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0
+        )
 
     routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
 
@@ -102,20 +104,24 @@ def load_balancing_loss_func(
         router_prob_per_expert = torch.mean(routing_weights, dim=0)
     else:
         batch_size, sequence_length = attention_mask.shape
-        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (
+            batch_size * sequence_length
+        )
 
         # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
         expert_attention_mask = (
             attention_mask[None, :, :, None, None]
-            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .expand(
+                (num_hidden_layers, batch_size, sequence_length, top_k, num_experts)
+            )
             .reshape(-1, top_k, num_experts)
             .to(compute_device)
         )
 
         # Compute the percentage of tokens routed to each experts
-        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
-            expert_attention_mask, dim=0
-        )
+        tokens_per_expert = torch.sum(
+            expert_mask.float() * expert_attention_mask, dim=0
+        ) / torch.sum(expert_attention_mask, dim=0)
 
         # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
         router_per_expert_attention_mask = (
@@ -126,9 +132,9 @@ def load_balancing_loss_func(
         )
 
         # Compute the average probability of routing to these experts
-        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
-            router_per_expert_attention_mask, dim=0
-        )
+        router_prob_per_expert = torch.sum(
+            routing_weights * router_per_expert_attention_mask, dim=0
+        ) / torch.sum(router_per_expert_attention_mask, dim=0)
 
     overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
     return overall_loss * num_experts
@@ -147,7 +153,9 @@ def __init__(self, config: MixtralConfig):
         self.act_fn = ACT2FN[config.hidden_act]
 
     def forward(self, hidden_states):
-        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
+        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(
+            hidden_states
+        )
         current_hidden_states = self.w2(current_hidden_states)
         return current_hidden_states
 
@@ -174,7 +182,9 @@ def __init__(self, config):
         # gating
         self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
 
-        self.experts = nn.ModuleList([MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
+        self.experts = nn.ModuleList(
+            [MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)]
+        )
 
         # Jitter parameters
         self.jitter_noise = config.router_jitter_noise
@@ -183,24 +193,32 @@ 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 *= 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)
 
         routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
-        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.top_k, dim=-1
+        )
         routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
         # we cast back to the input dtype
         routing_weights = routing_weights.to(hidden_states.dtype)
 
         final_hidden_states = torch.zeros(
-            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device,
         )
 
         # One hot encode the selected experts to create an expert mask
         # this will be used to easily index which expert is going to be sollicitated
-        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+        expert_mask = torch.nn.functional.one_hot(
+            selected_experts, num_classes=self.num_experts
+        ).permute(2, 1, 0)
 
         # Loop over all available experts in the model and perform the computation on each expert
         for expert_idx in range(self.num_experts):
@@ -210,12 +228,18 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
             # the current expert. We need to make sure to multiply the output hidden
             # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
             current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
-            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+            current_hidden_states = (
+                expert_layer(current_state) * routing_weights[top_x, idx, None]
+            )
 
             # However `index_add_` only support torch tensors for indexing so we'll use
             # the `top_x` tensor here.
-            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
-        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+            final_hidden_states.index_add_(
+                0, top_x, current_hidden_states.to(hidden_states.dtype)
+            )
+        final_hidden_states = final_hidden_states.reshape(
+            batch_size, sequence_length, hidden_dim
+        )
         return final_hidden_states, router_logits
 
 
@@ -235,8 +259,12 @@ def __init__(self, config: MixtralConfig, layer_idx: int):
         self.self_attn = MixtralAttention(config, layer_idx)
 
         self.block_sparse_moe = MixtralSparseMoeBlock(config)
-        self.input_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.input_layernorm = MixtralRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = MixtralRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
 
     @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
     def forward(
@@ -300,7 +328,9 @@ def forward(
         **kwargs: Unpack[TransformersKwargs],
     ) -> MoeModelOutputWithPast:
         if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
 
         if use_cache and past_key_values is None:
             past_key_values = DynamicCache(config=self.config)
@@ -309,14 +339,22 @@ def forward(
             inputs_embeds = self.embed_tokens(input_ids)
 
         if cache_position is None:
-            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
             cache_position = torch.arange(
-                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
             )
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask
+        mask_function = (
+            create_causal_mask
+            if self.config.sliding_window is None
+            else create_sliding_window_causal_mask
+        )
         causal_mask = mask_function(
             config=self.config,
             input_embeds=inputs_embeds,
@@ -399,7 +437,9 @@ def forward(
         ```"""
 
         output_router_logits = (
-            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+            output_router_logits
+            if output_router_logits is not None
+            else self.config.output_router_logits
         )
 
         # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
@@ -417,7 +457,11 @@ def forward(
 
         hidden_states = outputs.last_hidden_state
         # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
         logits = self.lm_head(hidden_states[:, slice_indices, :])
 
         loss = None
@@ -433,7 +477,9 @@ def forward(
                 attention_mask,
             )
             if labels is not None:
-                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+                loss += self.router_aux_loss_coef * aux_loss.to(
+                    loss.device
+                )  # make sure to reside in the same device
 
         return MoeCausalLMOutputWithPast(
             loss=loss,

tests/models/mixtral/test_mixtral_torch_export.py

@@ -21,7 +21,7 @@
 
 from transformers import MixtralConfig
 from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
-from transformers.testing_utils import require_torch, torch_device
+from transformers.testing_utils import require_torch
 
 
 @require_torch
@@ -43,31 +43,30 @@ def test_moe_block_torch_export(self):
         # Create MoE block
         moe_block = MixtralSparseMoeBlock(self.config)
         moe_block.eval()
-        
+
         # Move to meta device for export testing
         moe_block = moe_block.to("meta")
-        
+
         # Create test input
         batch_size, seq_len = 2, 8
         hidden_states = torch.randn(
-            batch_size, seq_len, self.config.hidden_size, 
-            device="meta"
+            batch_size, seq_len, self.config.hidden_size, device="meta"
         )
-        
+
         # Test torch.export - should not raise GuardOnDataDependentSymNode error
         try:
             exported_program = te.export(
-                moe_block,
-                args=(hidden_states,),
-                kwargs={},
-                strict=False
+                moe_block, args=(hidden_states,), kwargs={}, strict=False
             )
             # If export succeeds, the test passes
             self.assertIsNotNone(exported_program)
         except Exception as e:
             # Check if it's the specific error we're trying to avoid
             error_msg = str(e)
-            if "GuardOnDataDependentSymNode" in error_msg or "nonzero" in error_msg.lower():
+            if (
+                "GuardOnDataDependentSymNode" in error_msg
+                or "nonzero" in error_msg.lower()
+            ):
                 self.fail(
                     f"torch.export failed with data-dependent operation error: {error_msg}\n"
                     "This suggests the .nonzero() fix is not working properly."
@@ -81,30 +80,29 @@ def test_moe_block_functionality(self):
         # Create MoE block
         moe_block = MixtralSparseMoeBlock(self.config)
         moe_block.eval()
-        
+
         # Create test input
         batch_size, seq_len = 2, 4
         hidden_states = torch.randn(batch_size, seq_len, self.config.hidden_size)
-        
+
         # Forward pass
         with torch.no_grad():
             output, router_logits = moe_block(hidden_states)
-        
+
         # Verify output shapes
         self.assertEqual(output.shape, hidden_states.shape)
         self.assertEqual(
-            router_logits.shape, 
-            (batch_size * seq_len, self.config.num_local_experts)
+            router_logits.shape, (batch_size * seq_len, self.config.num_local_experts)
         )
-        
+
         # Verify that outputs are not all zeros (computation happened)
         self.assertFalse(torch.allclose(output, torch.zeros_like(output)))
-        
+
         # Test with different input to ensure different outputs
         hidden_states2 = torch.randn(batch_size, seq_len, self.config.hidden_size)
         with torch.no_grad():
             output2, _ = moe_block(hidden_states2)
-        
+
         # Outputs should be different for different inputs
         self.assertFalse(torch.allclose(output, output2))
 
@@ -117,7 +115,7 @@ def test_moe_block_export_with_different_configs(self):
             (16, 2),
             (8, 4),
         ]
-        
+
         for num_experts, top_k in test_configs:
             with self.subTest(num_experts=num_experts, top_k=top_k):
                 config = MixtralConfig(
@@ -127,28 +125,27 @@ def test_moe_block_export_with_different_configs(self):
                     num_experts_per_tok=top_k,
                     router_jitter_noise=0.0,
                 )
-                
+
                 moe_block = MixtralSparseMoeBlock(config)
                 moe_block.eval()
                 moe_block = moe_block.to("meta")
-                
+
                 hidden_states = torch.randn(1, 4, config.hidden_size, device="meta")
-                
+
                 # Should export without errors
                 try:
                     exported_program = te.export(
-                        moe_block,
-                        args=(hidden_states,),
-                        kwargs={},
-                        strict=False
+                        moe_block, args=(hidden_states,), kwargs={}, strict=False
                     )
                     self.assertIsNotNone(exported_program)
                 except Exception as e:
                     if "GuardOnDataDependentSymNode" in str(e):
-                        self.fail(f"Export failed for config ({num_experts}, {top_k}): {e}")
+                        self.fail(
+                            f"Export failed for config ({num_experts}, {top_k}): {e}"
+                        )
                     else:
                         raise
 
 
 if __name__ == "__main__":
-    unittest.main()
+    unittest.main()