From 387cfd268b1289d2ad8fde618d1fdc89691ed429 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 10:47:50 +0000
Subject: [PATCH 01/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_precision.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 38a3434490..9b8dba28e3 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -121,7 +121,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:

From 91780a250ef73b3a5f88d4e2f0c0343a0828dd5f Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:26:10 +0000
Subject: [PATCH 02/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_precision.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:

From 7410f8fa1ce0fdda7890a25ddb61d4e02d1d6c5b Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:28:55 +0000
Subject: [PATCH 03/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_speed.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index b774370519..9184f29f88 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -240,6 +240,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).

From 394f7985df0f7b0e7b42bd7e249bf95df0647dce Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:30:37 +0000
Subject: [PATCH 04/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 6e1341983f..7cb1074425 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -257,6 +257,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail

From ab27ea85203ce9d1a968ac753262e903c0131896 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:31:14 +0000
Subject: [PATCH 05/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 3 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index b774370519..9184f29f88 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -240,6 +240,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 6e1341983f..7cb1074425 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -257,6 +257,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail

From 1ab9db02ffbcb07a71b85aa077faee97137612f0 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:32:10 +0000
Subject: [PATCH 06/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index b774370519..9184f29f88 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -240,6 +240,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 6e1341983f..7cb1074425 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -257,6 +257,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index eb497a46d2..dc34e60cef 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -445,6 +445,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From c1b4ed1233526c1e581361f83168f9b3d4c3bd63 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:32:41 +0000
Subject: [PATCH 07/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index b774370519..9184f29f88 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -240,6 +240,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 1afdd4200d..cb6873a661 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -284,6 +284,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index eb497a46d2..dc34e60cef 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -445,6 +445,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From a8e87ac77ba30e19cdb5f6af072d9bdb5160218c Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:33:16 +0000
Subject: [PATCH 08/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 62bab8dcba..208cb8b513 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -249,6 +249,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 1afdd4200d..cb6873a661 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -284,6 +284,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index baaa8696f9..d03c48b477 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -445,6 +445,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From ffe5b697732e1018a1708c28847d6c3ebae74935 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:33:47 +0000
Subject: [PATCH 09/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 62bab8dcba..208cb8b513 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -249,6 +249,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 94be813a94..9e2e56a2a5 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -313,6 +313,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index baaa8696f9..d03c48b477 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -445,6 +445,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From db081d67deaa8c67a41dc15b6c022d16806a4d53 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:34:18 +0000
Subject: [PATCH 10/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 62bab8dcba..208cb8b513 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -249,6 +249,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 94be813a94..9e2e56a2a5 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -313,6 +313,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index baaa8696f9..d03c48b477 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -445,6 +445,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From 44beb54df5ea0fc79cc7faf9dbd68518a84317c9 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:36:04 +0000
Subject: [PATCH 11/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 62bab8dcba..208cb8b513 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -249,6 +249,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 94be813a94..9e2e56a2a5 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -313,6 +313,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index c37ca3ef14..d0b6997d07 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -465,6 +465,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From 5728670f6e441d0cf82522ecba36696e8b34909e Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 11:57:18 +0000
Subject: [PATCH 12/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion_attention.py         | 1 -
 studio/backend/core/inference/diffusion_precision.py         | 5 ++++-
 studio/backend/core/inference/diffusion_speed.py             | 2 ++
 studio/backend/core/inference/diffusion_transformer_quant.py | 1 +
 studio/backend/tests/test_diffusion_routes.py                | 2 ++
 5 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 662fff74dd..2738691e58 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -282,6 +282,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 94be813a94..9e2e56a2a5 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -313,6 +313,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index b00af9cfd6..b7fbb3f50c 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -487,6 +487,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots

From a927f4dc1e2343c02b34e96a902d5080505adc5e Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 13:58:10 +0000
Subject: [PATCH 13/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/diffusion.py    |  4 +++-
 .../core/inference/diffusion_attention.py     |  1 -
 .../core/inference/diffusion_precision.py     |  5 ++++-
 .../backend/core/inference/diffusion_speed.py |  2 ++
 .../inference/diffusion_transformer_quant.py  |  1 +
 .../backend/tests/test_diffusion_backend.py   |  4 +---
 studio/backend/tests/test_diffusion_routes.py |  2 ++
 studio/backend/tests/test_diffusion_sdxl.py   | 22 +++++++++----------
 8 files changed, 24 insertions(+), 17 deletions(-)

diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index 816e3a0e01..eaaa93dede 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -1538,7 +1538,9 @@ class DiffusionBackend:
                         mask_pil = mask_pil.resize(init_pil.size, _PILImage.NEAREST)
                 if init_pil is not None:
                     # Keep the VAE encode dtype consistent with the input image.
-                    self._align_vae_dtype(pipe, getattr(state.family, "denoiser_attr", "transformer"))
+                    self._align_vae_dtype(
+                        pipe, getattr(state.family, "denoiser_attr", "transformer")
+                    )
 
                 # Pipelines vary in which kwargs they accept (img2img derives size from the
                 # input image and may reject width/height; a distilled pipe may take no
diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 04f812a13e..e652b0068c 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -225,7 +225,6 @@ def _reset_global_backend_to_native(logger: Any) -> None:
             AttentionBackendName,
             _AttentionBackendRegistry,
         )
-
         _AttentionBackendRegistry.set_active_backend(AttentionBackendName.NATIVE)
     except Exception:  # noqa: BLE001 — best-effort; leave the global as-is on any change
         pass
diff --git a/studio/backend/core/inference/diffusion_precision.py b/studio/backend/core/inference/diffusion_precision.py
index 68036aae8d..030b6da7f2 100644
--- a/studio/backend/core/inference/diffusion_precision.py
+++ b/studio/backend/core/inference/diffusion_precision.py
@@ -120,7 +120,10 @@ def _cast_fp8(encoder: Any, target: Any) -> None:
     # gets cast to fp8 and, sharing one tensor, drags the embedding to fp8 with it. The
     # embedding then emits fp8 activations that crash the first RMSNorm. Skip the tied
     # projection so the shared tensor stays dense (lm_head is unused for prompt encoding).
-    get_out, get_in = getattr(encoder, "get_output_embeddings", None), getattr(encoder, "get_input_embeddings", None)
+    get_out, get_in = (
+        getattr(encoder, "get_output_embeddings", None),
+        getattr(encoder, "get_input_embeddings", None),
+    )
     out_emb = get_out() if callable(get_out) else None
     in_emb = get_in() if callable(get_in) else None
     if out_emb is not None and in_emb is not None and out_emb.weight is in_emb.weight:
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 662fff74dd..2738691e58 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -282,6 +282,8 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
     except Exception as exc:  # noqa: BLE001 — optimisation only
         _warn(logger, "cudnn_benchmark", exc)
         return False
+
+
 # The TF32 flag values from before the first max load flipped them, so a later
 # non-max load / unload can put the process back exactly as it found it (rather than
 # forcing a hardcoded default that might clobber another component's choice).
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index 94be813a94..9e2e56a2a5 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -313,6 +313,7 @@ def _make_quant_config(scheme: str, fast_accum: Optional[bool] = None) -> Any:
             return Float8DynamicActivationFloat8WeightConfig()
     if scheme == TQ_NVFP4:
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
+
         # Select the CUTLASS FP4 path, not the default Triton kernel: torchao defaults
         # use_triton_kernel=True, which needs MSLK installed. On a Blackwell box with the
         # CUTLASS FP4 extension but no MSLK, the default would make the smoke probe fail
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py
index 0b0e1e11cb..b05d6a3940 100644
--- a/studio/backend/tests/test_diffusion_backend.py
+++ b/studio/backend/tests/test_diffusion_backend.py
@@ -889,9 +889,7 @@ def test_load_sdxl_single_file_uses_pipeline_from_single_file(fake_runtime, tmp_
     assert status["loaded"] is True
     assert status["family"] == "sdxl"
     # The whole-pipeline single-file path was taken with the base repo as config.
-    assert _FakePipeline.last_single_file["path"] == str(
-        (tmp_path / "sdxl.safetensors").resolve()
-    )
+    assert _FakePipeline.last_single_file["path"] == str((tmp_path / "sdxl.safetensors").resolve())
     assert _FakePipeline.last_single_file["config"] == "stabilityai/stable-diffusion-xl-base-1.0"
     # The transformer-only single-file build was NOT taken.
     assert _FakeTransformer.last == {}
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index b00af9cfd6..b7fbb3f50c 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -487,6 +487,8 @@ def test_invalid_attention_backend_returns_422(client):
         json = {"model_path": "x/z-image", "gguf_filename": "q.gguf", "attention_backend": "bogus"},
     )
     assert resp.status_code == 422
+
+
 def test_prequant_path_doc_describes_allowlist_not_toggle():
     # The field help must match the code: UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH is a
     # directory allowlist, not a =1 toggle (diffusion_prequant._allowed_prequant_roots
diff --git a/studio/backend/tests/test_diffusion_sdxl.py b/studio/backend/tests/test_diffusion_sdxl.py
index a0b3044952..41ee904b94 100644
--- a/studio/backend/tests/test_diffusion_sdxl.py
+++ b/studio/backend/tests/test_diffusion_sdxl.py
@@ -46,7 +46,7 @@ def test_sdxl_detection_by_repo_and_override():
     assert detect_family("stabilityai/sdxl-turbo").name == "sdxl"
     assert detect_family("some-org/My-Cool-SDXL-Merge").name == "sdxl"
     assert detect_family("some-org/stable-diffusion-xl-anime").name == "sdxl"
-    assert detect_family("x", override="sdxl").name == "sdxl"
+    assert detect_family("x", override = "sdxl").name == "sdxl"
     # A GGUF DiT family must NOT be swallowed by the SDXL match.
     assert detect_family("unsloth/FLUX.1-schnell-GGUF").name == "flux.1"
 
@@ -91,9 +91,9 @@ class _FakeVae:
         self.moved_to = None
 
     def parameters(self):
-        yield types.SimpleNamespace(dtype=self._dtype)
+        yield types.SimpleNamespace(dtype = self._dtype)
 
-    def to(self, dtype=None):
+    def to(self, dtype = None):
         self.moved_to = dtype
         self._dtype = dtype
 
@@ -101,29 +101,29 @@ class _FakeVae:
 def test_align_vae_dtype_uses_unet_denoiser():
     # For SDXL the denoiser lives at pipe.unet; _align_vae_dtype must read it (a pipe
     # with only .unet and no .transformer) and cast the VAE to the U-Net's dtype.
-    vae = _FakeVae(dtype="float32")
-    pipe = types.SimpleNamespace(unet=types.SimpleNamespace(dtype="bfloat16"), vae=vae)
+    vae = _FakeVae(dtype = "float32")
+    pipe = types.SimpleNamespace(unet = types.SimpleNamespace(dtype = "bfloat16"), vae = vae)
     DiffusionBackend._align_vae_dtype(pipe, "unet")
     assert vae.moved_to == "bfloat16"
 
 
 def test_align_vae_dtype_transformer_default_unchanged():
     # DiT default: reads pipe.transformer; a pipe with no transformer is a safe no-op.
-    vae = _FakeVae(dtype="float32")
-    pipe = types.SimpleNamespace(transformer=types.SimpleNamespace(dtype="bfloat16"), vae=vae)
+    vae = _FakeVae(dtype = "float32")
+    pipe = types.SimpleNamespace(transformer = types.SimpleNamespace(dtype = "bfloat16"), vae = vae)
     DiffusionBackend._align_vae_dtype(pipe)
     assert vae.moved_to == "bfloat16"
     # No denoiser attribute -> no-op (does not raise, does not move the VAE).
-    vae2 = _FakeVae(dtype="float32")
-    DiffusionBackend._align_vae_dtype(types.SimpleNamespace(vae=vae2), "unet")
+    vae2 = _FakeVae(dtype = "float32")
+    DiffusionBackend._align_vae_dtype(types.SimpleNamespace(vae = vae2), "unet")
     assert vae2.moved_to is None
 
 
 def test_sdxl_lora_supported_on_diffusers():
     # SDXL is bf16/bnb-4bit on diffusers -> LoRA is allowed (unlike GGUF-via-diffusers).
     assert diffusion_lora.supports_lora(
-        engine="diffusers", family="sdxl", model_kind="pipeline", transformer_quant=None
+        engine = "diffusers", family = "sdxl", model_kind = "pipeline", transformer_quant = None
     )
     assert diffusion_lora.supports_lora(
-        engine="diffusers", family="sdxl", model_kind="single_file", transformer_quant=None
+        engine = "diffusers", family = "sdxl", model_kind = "single_file", transformer_quant = None
     )

From 0f1c4a35ba293986a163fc2944b6d366092bd4f0 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 1 Jul 2026 14:48:15 +0000
Subject: [PATCH 14/14] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 .../training/diffusion_training_service.py    | 65 +++++++++--------
 studio/backend/models/training.py             | 32 +++++----
 studio/backend/routes/training.py             | 11 +--
 .../backend/tests/test_diffusion_training.py  | 69 +++++++++++++------
 4 files changed, 107 insertions(+), 70 deletions(-)

diff --git a/studio/backend/core/training/diffusion_training_service.py b/studio/backend/core/training/diffusion_training_service.py
index 166d2a9193..bdefcfe25b 100644
--- a/studio/backend/core/training/diffusion_training_service.py
+++ b/studio/backend/core/training/diffusion_training_service.py
@@ -34,8 +34,7 @@ _TERMINAL = ("complete", "error")
 def _default_target(*, event_queue: Any, stop_queue: Any, config: dict) -> None:
     # Imported lazily so this module (and the route layer) stays torch-free at import.
     from .diffusion_lora_trainer import run_diffusion_training_process
-
-    run_diffusion_training_process(event_queue=event_queue, stop_queue=stop_queue, config=config)
+    run_diffusion_training_process(event_queue = event_queue, stop_queue = stop_queue, config = config)
 
 
 def _idle_state() -> dict[str, Any]:
@@ -94,20 +93,23 @@ class DiffusionTrainingService:
             if self._proc is not None and self._proc.is_alive():
                 raise RuntimeError("A diffusion training job is already running.")
             if self._pump is not None and self._pump.is_alive():
-                self._pump.join(timeout=5.0)
+                self._pump.join(timeout = 5.0)
 
             job_id = uuid.uuid4().hex
             event_queue = self._ctx.Queue()
             self._stop_queue = self._ctx.Queue()
             self._proc = self._ctx.Process(
-                target=self._target,
-                kwargs={"event_queue": event_queue, "stop_queue": self._stop_queue, "config": config},
-                daemon=True,
+                target = self._target,
+                kwargs = {
+                    "event_queue": event_queue,
+                    "stop_queue": self._stop_queue,
+                    "config": config,
+                },
+                daemon = True,
             )
             self._proc.start()
             try:
                 from utils.process_lifetime import adopt_pid
-
                 adopt_pid(self._proc.pid)  # bind to parent lifetime (no zombie on exit)
             except Exception:  # noqa: BLE001 -- lifetime binding is best-effort
                 pass
@@ -115,11 +117,15 @@ class DiffusionTrainingService:
             now = time.time()
             self._state = _idle_state()
             self._state.update(
-                active=True, job_id=job_id, status="running",
-                message="Starting diffusion LoRA training...", started_at=now, updated_at=now,
+                active = True,
+                job_id = job_id,
+                status = "running",
+                message = "Starting diffusion LoRA training...",
+                started_at = now,
+                updated_at = now,
             )
             self._pump = threading.Thread(
-                target=self._pump_loop, args=(event_queue, self._proc), daemon=True
+                target = self._pump_loop, args = (event_queue, self._proc), daemon = True
             )
             self._pump.start()
             return job_id
@@ -149,7 +155,7 @@ class DiffusionTrainingService:
     def _pump_loop(self, event_queue: Any, proc: Any) -> None:
         while True:
             try:
-                ev = event_queue.get(timeout=1.0)
+                ev = event_queue.get(timeout = 1.0)
             except Exception:  # noqa: BLE001 -- Empty (timeout) or a closed queue
                 if not proc.is_alive():
                     # Drain anything buffered, then decide if it exited cleanly.
@@ -163,9 +169,10 @@ class DiffusionTrainingService:
                     with self._lock:
                         if self._state.get("status") not in ("completed", "stopped", "error"):
                             self._state.update(
-                                active=False, status="error",
-                                message="Training process exited unexpectedly.",
-                                updated_at=time.time(),
+                                active = False,
+                                status = "error",
+                                message = "Training process exited unexpectedly.",
+                                updated_at = time.time(),
                             )
                     _ = drained
                     return
@@ -182,33 +189,33 @@ class DiffusionTrainingService:
             s = self._state
             s["updated_at"] = time.time()
             if etype == "model_load_started":
-                s.update(in_model_load=True, status="running", message="Loading base model...")
+                s.update(in_model_load = True, status = "running", message = "Loading base model...")
                 if ev.get("num_images") is not None:
                     s["num_images"] = ev.get("num_images")
             elif etype == "model_load_completed":
-                s.update(in_model_load=False, message="Training...")
+                s.update(in_model_load = False, message = "Training...")
             elif etype == "progress":
                 s.update(
-                    status="running",
-                    step=ev.get("step", s["step"]),
-                    total_steps=ev.get("total_steps", s["total_steps"]),
-                    loss=ev.get("loss", s["loss"]),
-                    avg_loss=ev.get("avg_loss", s["avg_loss"]),
-                    learning_rate=ev.get("learning_rate", s["learning_rate"]),
-                    message="Training...",
+                    status = "running",
+                    step = ev.get("step", s["step"]),
+                    total_steps = ev.get("total_steps", s["total_steps"]),
+                    loss = ev.get("loss", s["loss"]),
+                    avg_loss = ev.get("avg_loss", s["avg_loss"]),
+                    learning_rate = ev.get("learning_rate", s["learning_rate"]),
+                    message = "Training...",
                 )
             elif etype == "complete":
                 s.update(
-                    active=False,
-                    status="stopped" if ev.get("stopped") else "completed",
-                    output_dir=ev.get("output_dir"),
-                    lora_path=ev.get("lora_path"),
-                    message="Stopped (partial adapter saved)."
+                    active = False,
+                    status = "stopped" if ev.get("stopped") else "completed",
+                    output_dir = ev.get("output_dir"),
+                    lora_path = ev.get("lora_path"),
+                    message = "Stopped (partial adapter saved)."
                     if ev.get("stopped")
                     else "Training complete.",
                 )
             elif etype == "error":
-                s.update(active=False, status="error", message=str(ev.get("message", "error")))
+                s.update(active = False, status = "error", message = str(ev.get("message", "error")))
 
 
 _service: Optional[DiffusionTrainingService] = None
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index 85f0949810..9a7774d8be 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -675,28 +675,30 @@ class DiffusionTrainingStartRequest(BaseModel):
     rest carry the trainer's defaults.
     """
 
-    model_config = ConfigDict(protected_namespaces=())
+    model_config = ConfigDict(protected_namespaces = ())
 
-    base_model: str = Field(..., description="HF repo id or local path to an SDXL pipeline")
-    data_dir: str = Field(..., description="Folder of training images (+ captions)")
-    output_dir: str = Field(..., description="Directory to write the LoRA .safetensors into")
+    base_model: str = Field(..., description = "HF repo id or local path to an SDXL pipeline")
+    data_dir: str = Field(..., description = "Folder of training images (+ captions)")
+    output_dir: str = Field(..., description = "Directory to write the LoRA .safetensors into")
     instance_prompt: Optional[str] = Field(
-        None, description="Dreambooth caption applied to images without their own caption"
+        None, description = "Dreambooth caption applied to images without their own caption"
     )
-    resolution: int = Field(1024, ge=64, le=2048, description="Square training resolution (multiple of 8)")
-    train_steps: int = Field(500, ge=1, le=100000)
-    learning_rate: float = Field(1e-4, gt=0)
-    train_batch_size: int = Field(1, ge=1, le=64)
-    gradient_accumulation_steps: int = Field(1, ge=1, le=256)
-    lora_rank: int = Field(16, ge=1, le=320)
-    lora_alpha: Optional[int] = Field(None, ge=1, le=640, description="Defaults to lora_rank")
-    lora_dropout: float = Field(0.0, ge=0.0, le=1.0)
+    resolution: int = Field(
+        1024, ge = 64, le = 2048, description = "Square training resolution (multiple of 8)"
+    )
+    train_steps: int = Field(500, ge = 1, le = 100000)
+    learning_rate: float = Field(1e-4, gt = 0)
+    train_batch_size: int = Field(1, ge = 1, le = 64)
+    gradient_accumulation_steps: int = Field(1, ge = 1, le = 256)
+    lora_rank: int = Field(16, ge = 1, le = 320)
+    lora_alpha: Optional[int] = Field(None, ge = 1, le = 640, description = "Defaults to lora_rank")
+    lora_dropout: float = Field(0.0, ge = 0.0, le = 1.0)
     seed: int = Field(42)
     mixed_precision: Literal["bf16", "fp16", "no"] = Field("bf16")
-    snr_gamma: Optional[float] = Field(5.0, description="Min-SNR loss weighting; null disables")
+    snr_gamma: Optional[float] = Field(5.0, description = "Min-SNR loss weighting; null disables")
     gradient_checkpointing: bool = Field(True)
     lr_scheduler: str = Field("constant")
-    lr_warmup_steps: int = Field(0, ge=0)
+    lr_warmup_steps: int = Field(0, ge = 0)
     center_crop: bool = Field(False)
     random_flip: bool = Field(True)
     caption_column: str = Field("text")
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index b2ea6dfaa4..c96189bef2 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -1031,8 +1031,7 @@ async def stream_training_progress(
 
 @router.post("/diffusion/start", response_model = DiffusionTrainingStartResponse)
 async def start_diffusion_training(
-    body: DiffusionTrainingStartRequest,
-    current_subject: str = Depends(get_current_subject),
+    body: DiffusionTrainingStartRequest, current_subject: str = Depends(get_current_subject)
 ):
     """Start an SDXL LoRA training job from an image + caption dataset."""
     from core.training.diffusion_training_service import get_diffusion_training_service
@@ -1047,8 +1046,11 @@ async def start_diffusion_training(
         raise HTTPException(status_code = 409, detail = str(e))
     except Exception as e:
         raise log_and_http_error(
-            e, 500, "Failed to start diffusion training",
-            event = "diffusion_training.start_failed", log = logger,
+            e,
+            500,
+            "Failed to start diffusion training",
+            event = "diffusion_training.start_failed",
+            log = logger,
         )
     return DiffusionTrainingStartResponse(job_id = job_id, status = "running")
 
@@ -1066,5 +1068,4 @@ async def stop_diffusion_training(current_subject: str = Depends(get_current_sub
 async def diffusion_training_status(current_subject: str = Depends(get_current_subject)):
     """Poll the current diffusion training job's status/progress (JSON)."""
     from core.training.diffusion_training_service import get_diffusion_training_service
-
     return DiffusionTrainingStatusResponse(**get_diffusion_training_service().status())
diff --git a/studio/backend/tests/test_diffusion_training.py b/studio/backend/tests/test_diffusion_training.py
index ae95e6ce7e..ffc231673c 100644
--- a/studio/backend/tests/test_diffusion_training.py
+++ b/studio/backend/tests/test_diffusion_training.py
@@ -32,8 +32,8 @@ class _FakeQueue:
     def put(self, x):
         self._q.put(x)
 
-    def get(self, timeout=None):
-        return self._q.get(timeout=timeout)  # raises queue.Empty on timeout
+    def get(self, timeout = None):
+        return self._q.get(timeout = timeout)  # raises queue.Empty on timeout
 
     def get_nowait(self):
         return self._q.get_nowait()
@@ -50,7 +50,7 @@ class _FakeProc:
         self.pid = 4321
 
     def start(self):
-        self._thread = threading.Thread(target=self._target, kwargs=self._kwargs, daemon=True)
+        self._thread = threading.Thread(target = self._target, kwargs = self._kwargs, daemon = True)
         self._thread.start()
 
     def is_alive(self):
@@ -69,10 +69,24 @@ def _happy_target(*, event_queue, stop_queue, config):
     event_queue.put({"type": "model_load_started", "num_images": 3})
     event_queue.put({"type": "model_load_completed"})
     event_queue.put(
-        {"type": "progress", "step": 1, "total_steps": 2, "loss": 0.5, "avg_loss": 0.5, "learning_rate": 1e-4}
+        {
+            "type": "progress",
+            "step": 1,
+            "total_steps": 2,
+            "loss": 0.5,
+            "avg_loss": 0.5,
+            "learning_rate": 1e-4,
+        }
     )
     event_queue.put(
-        {"type": "progress", "step": 2, "total_steps": 2, "loss": 0.4, "avg_loss": 0.45, "learning_rate": 1e-4}
+        {
+            "type": "progress",
+            "step": 2,
+            "total_steps": 2,
+            "loss": 0.4,
+            "avg_loss": 0.45,
+            "learning_rate": 1e-4,
+        }
     )
     event_queue.put(
         {
@@ -86,8 +100,10 @@ def _happy_target(*, event_queue, stop_queue, config):
 
 def _stoppable_target(*, event_queue, stop_queue, config):
     event_queue.put({"type": "model_load_completed"})
-    stop_queue.get(timeout=5.0)  # block until stop() signals
-    event_queue.put({"type": "complete", "output_dir": config["output_dir"], "lora_path": "x", "stopped": True})
+    stop_queue.get(timeout = 5.0)  # block until stop() signals
+    event_queue.put(
+        {"type": "complete", "output_dir": config["output_dir"], "lora_path": "x", "stopped": True}
+    )
 
 
 def _crashing_target(*, event_queue, stop_queue, config):
@@ -98,7 +114,11 @@ def _crashing_target(*, event_queue, stop_queue, config):
 _CFG = {"base_model": "b", "data_dir": "d", "output_dir": "/tmp/out", "train_steps": 2}
 
 
-def _wait_status(svc, *terminal, timeout=3.0):
+def _wait_status(
+    svc,
+    *terminal,
+    timeout = 3.0,
+):
     end = time.time() + timeout
     while time.time() < end:
         st = svc.status()
@@ -109,7 +129,7 @@ def _wait_status(svc, *terminal, timeout=3.0):
 
 
 def test_service_happy_path():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_happy_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _happy_target)
     job_id = svc.start(dict(_CFG))
     assert job_id
     st = _wait_status(svc, "completed")
@@ -122,7 +142,7 @@ def test_service_happy_path():
 
 
 def test_service_rejects_bad_config_before_spawn():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_happy_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _happy_target)
     with pytest.raises(ValueError):
         svc.start({**_CFG, "train_steps": 0})
     # Nothing was spawned; still idle.
@@ -130,7 +150,7 @@ def test_service_rejects_bad_config_before_spawn():
 
 
 def test_service_rejects_second_concurrent_job():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_stoppable_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _stoppable_target)
     svc.start(dict(_CFG))
     _wait_status(svc, "running")
     with pytest.raises(RuntimeError):
@@ -140,7 +160,7 @@ def test_service_rejects_second_concurrent_job():
 
 
 def test_service_stop_marks_stopped():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_stoppable_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _stoppable_target)
     svc.start(dict(_CFG))
     _wait_status(svc, "running")
     assert svc.stop() is True
@@ -152,7 +172,7 @@ def test_service_stop_marks_stopped():
 
 
 def test_service_crash_without_terminal_event_is_error():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_crashing_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _crashing_target)
     svc.start(dict(_CFG))
     st = _wait_status(svc, "error")
     assert st["status"] == "error"
@@ -160,7 +180,7 @@ def test_service_crash_without_terminal_event_is_error():
 
 
 def test_apply_event_transitions():
-    svc = DiffusionTrainingService(ctx=_FakeCtx(), target=_happy_target)
+    svc = DiffusionTrainingService(ctx = _FakeCtx(), target = _happy_target)
     svc._apply_event({"type": "model_load_started", "num_images": 5})
     assert svc.status()["in_model_load"] is True and svc.status()["num_images"] == 5
     svc._apply_event({"type": "model_load_completed"})
@@ -212,25 +232,32 @@ def client(monkeypatch):
         "core.training.diffusion_training_service.get_diffusion_training_service", lambda: fake
     )
     app = FastAPI()
-    app.include_router(training_router, prefix="/api/train")
+    app.include_router(training_router, prefix = "/api/train")
     app.dependency_overrides[get_current_subject] = lambda: "test-user"
     c = TestClient(app)
     c._fake = fake  # type: ignore[attr-defined]
     return c
 
 
-_BODY = {"base_model": "stabilityai/sdxl-turbo", "data_dir": "/data", "output_dir": "/out", "train_steps": 10}
+_BODY = {
+    "base_model": "stabilityai/sdxl-turbo",
+    "data_dir": "/data",
+    "output_dir": "/out",
+    "train_steps": 10,
+}
 
 
 def test_route_start_ok(client):
-    r = client.post("/api/train/diffusion/start", json=_BODY)
+    r = client.post("/api/train/diffusion/start", json = _BODY)
     assert r.status_code == 200, r.text
     assert r.json() == {"job_id": "job-123", "status": "running"}
     assert client._fake.started_with["base_model"] == "stabilityai/sdxl-turbo"
 
 
 def test_route_start_missing_required_is_422(client):
-    r = client.post("/api/train/diffusion/start", json={"base_model": "x"})  # no data_dir/output_dir
+    r = client.post(
+        "/api/train/diffusion/start", json = {"base_model": "x"}
+    )  # no data_dir/output_dir
     assert r.status_code == 422
 
 
@@ -239,7 +266,7 @@ def test_route_start_bad_config_maps_to_400(client, monkeypatch):
         raise ValueError("resolution must be a multiple of 8")
 
     client._fake.start = _raise  # type: ignore[assignment]
-    r = client.post("/api/train/diffusion/start", json=_BODY)
+    r = client.post("/api/train/diffusion/start", json = _BODY)
     assert r.status_code == 400
     assert "multiple of 8" in r.json()["detail"]
 
@@ -249,12 +276,12 @@ def test_route_start_conflict_maps_to_409(client):
         raise RuntimeError("A diffusion training job is already running.")
 
     client._fake.start = _raise  # type: ignore[assignment]
-    r = client.post("/api/train/diffusion/start", json=_BODY)
+    r = client.post("/api/train/diffusion/start", json = _BODY)
     assert r.status_code == 409
 
 
 def test_route_status_and_stop(client):
-    client.post("/api/train/diffusion/start", json=_BODY)
+    client.post("/api/train/diffusion/start", json = _BODY)
     s = client.get("/api/train/diffusion/status")
     assert s.status_code == 200 and s.json()["status"] == "running"
     st = client.post("/api/train/diffusion/stop")