[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

@@ -118,7 +118,10 @@ def te_quant_supported(target: Any, mode: str) -> bool:
 # 4-bit is a steeper quality cost), never an auto pick, consistent with the DiT ladder.
 _TE_AUTO_LADDER: tuple[tuple[tuple[int, int], tuple[str, ...]], ...] = (
     ((8, 9), (TE_QUANT_FP8_DYNAMIC, TE_QUANT_INT8, TE_QUANT_FP8)),  # Ada sm_89 / Hopper / Blackwell
-    ((8, 0), (TE_QUANT_INT8, TE_QUANT_FP8)),  # Ampere sm_80/86: no fp8 GEMM -> int8 or layerwise fp8
+    (
+        (8, 0),
+        (TE_QUANT_INT8, TE_QUANT_FP8),
+    ),  # Ampere sm_80/86: no fp8 GEMM -> int8 or layerwise fp8
 )
 
 # Text encoders whose activation ranges break a scheme at the MODEL level (measured hidden-state
@@ -145,7 +148,6 @@ def _te_scheme_probe(scheme: str, device: str) -> bool:
         return True
     try:
         from .diffusion_transformer_quant import _smoke_probe
-
         return _smoke_probe(tq, device)
     except Exception:
         return False
@@ -179,7 +181,9 @@ def select_te_quant_scheme(
             # Consumer GDDR parts run int8 full-rate but halve fp8 FP32-accumulate: prefer int8.
             ordered = (
                 (TE_QUANT_INT8,) + tuple(s for s in schemes if s != TE_QUANT_INT8)
-                if TE_QUANT_INT8 in schemes and schemes[0] != TE_QUANT_INT8 and _is_consumer_gpu(device)
+                if TE_QUANT_INT8 in schemes
+                and schemes[0] != TE_QUANT_INT8
+                and _is_consumer_gpu(device)
                 else schemes
             )
             for scheme in ordered:

studio/backend/models/inference.py

@@ -1745,7 +1745,9 @@ class DiffusionLoadRequest(BaseModel):
         "default (also regional torch.compile where eligible), "
         "max (also TF32 + fused QKV).",
     )
-    text_encoder_quant: Optional[Literal["auto", "none", "off", "fp8", "fp8_dynamic", "int8", "nvfp4"]] = Field(
+    text_encoder_quant: Optional[
+        Literal["auto", "none", "off", "fp8", "fp8_dynamic", "int8", "nvfp4"]
+    ] = Field(
         None,
         description = "Quantise the companion text encoder(s). auto (the default when unset) picks "
         "the fastest accurate scheme for this GPU + model family (fp8_dynamic on fp8-GEMM silicon, "
@@ -2348,7 +2350,9 @@ class VideoLoadRequest(BaseModel):
             "backend's transformer_quant field.",
         )
     )
-    text_encoder_quant: Optional[Literal["auto", "none", "off", "fp8", "fp8_dynamic", "int8", "nvfp4"]] = Field(
+    text_encoder_quant: Optional[
+        Literal["auto", "none", "off", "fp8", "fp8_dynamic", "int8", "nvfp4"]
+    ] = Field(
         None,
         description = "Quantise the dense companion text encoder (Gemma3 / UMT5 / Qwen2.5-VL), "
         "which loads bf16 from the base repo regardless of how the DiT was sourced and is often "

studio/backend/tests/test_diffusion_precision.py

@@ -404,7 +404,13 @@ def test_nvfp4_filter_keeps_vision_tower_dense(monkeypatch):
 # ── auto ladder (select_te_quant_scheme) ────────────────────────────────────────
 
 
-def _stub_tq_select(monkeypatch, *, cc, consumer = False, smoke = True):
+def _stub_tq_select(
+    monkeypatch,
+    *,
+    cc,
+    consumer = False,
+    smoke = True,
+):
     """Stub the transformer module's shared helpers that select_te_quant_scheme imports:
     capability, GPU class, and the kernel smoke probe (bool or a (tq, dev) predicate)."""
     dtq = types.ModuleType("core.inference.diffusion_transformer_quant")