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 01/11] [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 02/11] [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 03/11] [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 04/11] [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 05/11] [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 06/11] [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 07/11] [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 08/11] [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 09/11] [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 10/11] [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 11/11] [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