diff --git a/scripts/verify_comment_only_diff.py b/scripts/verify_comment_only_diff.py
index 068d3244df..90eafb7f8f 100644
--- a/scripts/verify_comment_only_diff.py
+++ b/scripts/verify_comment_only_diff.py
@@ -35,6 +35,7 @@ Example:
     git diff --name-only origin/main..HEAD \\
       | xargs python scripts/verify_comment_only_diff.py --base origin/main
 """
+
 from __future__ import annotations
 
 import argparse
@@ -49,7 +50,9 @@ import yaml
 
 def _git_show(rev: str, path: str) -> str:
     return subprocess.check_output(
-        ["git", "show", f"{rev}:{path}"], text = True, stderr = subprocess.DEVNULL,
+        ["git", "show", f"{rev}:{path}"],
+        text = True,
+        stderr = subprocess.DEVNULL,
     )
 
 
@@ -145,8 +148,7 @@ def _walk_yaml_diff(b: Any, a: Any, prefix: str = "") -> None:
     elif isinstance(b, list):
         if len(b) != len(a):
             print(
-                f"     list len at {prefix or '/'}: "
-                f"{len(b)} -> {len(a)}",
+                f"     list len at {prefix or '/'}: " f"{len(b)} -> {len(a)}",
             )
         for i, (bi, ai) in enumerate(zip(b, a)):
             _walk_yaml_diff(bi, ai, f"{prefix}[{i}]")
diff --git a/tests/test_gemma4_moe_4bit_swap.py b/tests/test_gemma4_moe_4bit_swap.py
index 2093067624..f12b592dc4 100644
--- a/tests/test_gemma4_moe_4bit_swap.py
+++ b/tests/test_gemma4_moe_4bit_swap.py
@@ -6,6 +6,7 @@ that exercise the swap helper's shape contract, idempotence, and gating
 behaviour. The full repro (resident VRAM 46 GB -> 14.27 GB, cosine sim 0.994
 vs BF16) is documented in the PR description.
 """
+
 import importlib
 import os
 
@@ -54,12 +55,14 @@ def _stub_gemma4_module():
         requires_grad = False,
     )
     from transformers.activations import ACT2FN
+
     module.act_fn = ACT2FN[_StubConfig.hidden_activation]
     return module
 
 
 def test_is_enabled_reads_env_var():
     from unsloth.models import gemma4_moe_4bit
+
     old = os.environ.pop("UNSLOTH_GEMMA4_MOE_4BIT", None)
     try:
         assert gemma4_moe_4bit.is_gemma4_moe_4bit_enabled() is False
@@ -78,6 +81,7 @@ def test_swap_skips_models_without_gemma4_experts():
     from unsloth.models.gemma4_moe_4bit import (
         swap_gemma4_experts_to_per_expert_linear4bit,
     )
+
     model = nn.Sequential(nn.Linear(8, 8), nn.Linear(8, 8))
     assert swap_gemma4_experts_to_per_expert_linear4bit(model) == 0
 
@@ -111,6 +115,7 @@ def test_swap_idempotent_on_stub_module_without_cuda():
     from unsloth.models.gemma4_moe_4bit import (
         swap_gemma4_experts_to_per_expert_linear4bit,
     )
+
     if not torch.cuda.is_available():
         # CPU-only: bnb's Linear4bit init would fail. Validate the model-walk
         # path on an empty Sequential to confirm the helper is side-effect-free.
diff --git a/unsloth/models/gemma4_moe_4bit.py b/unsloth/models/gemma4_moe_4bit.py
index d8453beb98..0863fa477c 100644
--- a/unsloth/models/gemma4_moe_4bit.py
+++ b/unsloth/models/gemma4_moe_4bit.py
@@ -32,6 +32,7 @@ Gated on UNSLOTH_GEMMA4_MOE_4BIT (default off) and on load_in_4bit=True.
 Default off until the matching per-expert LoRA path lands; opt in via
 the env var if you want the VRAM win without QLoRA training.
 """
+
 from __future__ import annotations
 
 import os
@@ -67,7 +68,8 @@ def _per_expert_forward(
     final_hidden_states = torch.zeros_like(hidden_states)
     with torch.no_grad():
         expert_mask = torch.nn.functional.one_hot(
-            top_k_index, num_classes = self.num_experts,
+            top_k_index,
+            num_classes = self.num_experts,
         )
         expert_mask = expert_mask.permute(2, 1, 0)
         expert_hit = torch.greater(expert_mask.sum(dim = (-1, -2)), 0).nonzero()
@@ -81,9 +83,7 @@ def _per_expert_forward(
         gate_up = self.gate_up_proj_4bit[expert_idx](current_state)
         gate, up = gate_up.chunk(2, dim = -1)
         current_hidden_states = self.act_fn(gate) * up
-        current_hidden_states = self.down_proj_4bit[expert_idx](
-            current_hidden_states
-        )
+        current_hidden_states = self.down_proj_4bit[expert_idx](current_hidden_states)
         current_hidden_states = (
             current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
         )
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index c459a153fe..68cfb8a3fe 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1060,6 +1060,7 @@ class FastBaseModel:
                         is_gemma4_moe_4bit_enabled,
                         swap_gemma4_experts_to_per_expert_linear4bit,
                     )
+
                     if is_gemma4_moe_4bit_enabled():
                         _swapped = swap_gemma4_experts_to_per_expert_linear4bit(
                             model,