diff --git a/studio/backend/core/training/diffusion_train_common.py b/studio/backend/core/training/diffusion_train_common.py
index 9d1dfc3475..dbbcda32d7 100644
--- a/studio/backend/core/training/diffusion_train_common.py
+++ b/studio/backend/core/training/diffusion_train_common.py
@@ -264,7 +264,6 @@ def bf16_unsupported_reason(resolved_family: str) -> Optional[str]:
         return None
     try:
         import torch
-
         if torch.cuda.is_available() and not torch.cuda.is_bf16_supported():
             return (
                 "This trainer requires a bfloat16-capable GPU (Ampere or newer); this CUDA "
diff --git a/studio/backend/tests/test_diffusion_base_precision.py b/studio/backend/tests/test_diffusion_base_precision.py
index 88f83ce307..c6cf0ec33c 100644
--- a/studio/backend/tests/test_diffusion_base_precision.py
+++ b/studio/backend/tests/test_diffusion_base_precision.py
@@ -81,7 +81,9 @@ def test_base_precision_denies_fp8_for_corrupted_family():
     # The deny is fp8-specific: int8 (per-token, unaffected) and the other dense modes stay
     # allowed for the same Qwen base.
     for mode in ("nf4", "bf16", "int8", "auto"):
-        norm = _cfg(base_model = _QWEN_DENSE, base_precision = mode, mixed_precision = "bf16").normalized()
+        norm = _cfg(
+            base_model = _QWEN_DENSE, base_precision = mode, mixed_precision = "bf16"
+        ).normalized()
         assert norm.resolved_family == "qwen-image"
         assert norm.base_precision == mode