From daa13cf3ade875cd0b88cc3f45555bda827aeaa2 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Mon, 16 Mar 2026 07:10:47 +0000
Subject: [PATCH] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 unsloth/kernels/fp8.py         | 25 ++++++++++++++++---------
 unsloth/models/_utils.py       |  5 +++++
 unsloth/models/loader_utils.py |  5 +----
 3 files changed, 22 insertions(+), 13 deletions(-)

diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index b7fb43adb1..a1d88444e6 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -628,9 +628,9 @@ def module_forward_patch(forward_function, scale_attr = "weight_scale"):
 
 def _compressed_linear_supports_unsloth_fp8(self):
     if (
-        CompressedLinear is None or
-        QuantizationStrategy is None or
-        QuantizationStatus is None
+        CompressedLinear is None
+        or QuantizationStrategy is None
+        or QuantizationStatus is None
     ):
         return False
 
@@ -639,10 +639,10 @@ def _compressed_linear_supports_unsloth_fp8(self):
     quantization_scheme = getattr(self, "quantization_scheme", None)
     quantization_args = getattr(quantization_scheme, "weights", None)
     if (
-        weight is None or
-        weight_scale is None or
-        quantization_args is None or
-        weight.dtype != torch.float8_e4m3fn
+        weight is None
+        or weight_scale is None
+        or quantization_args is None
+        or weight.dtype != torch.float8_e4m3fn
     ):
         return False
 
@@ -679,6 +679,7 @@ def _compressed_linear_forward_fallback(self, input):
         weight_data = self.compressor.decompress_module(self)
         param = nn.Parameter(weight_data, requires_grad = False)
         from compressed_tensors.utils import register_offload_parameter
+
         register_offload_parameter(self, "weight", param)
         self.quantization_status = QuantizationStatus.FROZEN
 
@@ -715,7 +716,11 @@ def _fp8_moe_lora_extractor(wrapper, weight_A, weight_B, scaling, num_experts):
 
     param_name = getattr(wrapper, "parameter_name", None)
 
-    if param_name == "down_proj" and intermediate_dim is not None and hidden_dim is not None:
+    if (
+        param_name == "down_proj"
+        and intermediate_dim is not None
+        and hidden_dim is not None
+    ):
         first_weight = weight_B.view(dim_B, num_experts, rank_per_expert)
         first_weight = first_weight.permute(1, 0, 2).contiguous()
         second_weight = weight_A.view(num_experts, rank_per_expert, dim_A)
@@ -764,7 +769,9 @@ def _patch_fp8_moe_experts():
     experts_interface["grouped_mm"] = forward_moe_backend
     experts_interface["batched_mm"] = forward_native_moe_loop
     if hasattr(finegrained_fp8, "FP8Experts"):
-        finegrained_fp8.FP8Experts._unsloth_lora_extractor_fn = staticmethod(_fp8_moe_lora_extractor)
+        finegrained_fp8.FP8Experts._unsloth_lora_extractor_fn = staticmethod(
+            _fp8_moe_lora_extractor
+        )
 
 
 # Patch the forward functions of the layers (for compiled models)
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 9eddab9f71..37d3e9a2bc 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -94,13 +94,18 @@ import functools
 import textwrap
 import logging
 import warnings, subprocess, inspect, psutil, os, math
+
 try:
     from transformers.utils import auto_docstring
 except:
+
     def auto_docstring(*args, **kwargs):
         def decorator(obj):
             return obj
+
         return decorator
+
+
 from unsloth_zoo.utils import Version, get_quant_type
 from importlib.metadata import version as importlib_version
 from ..device_type import (
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index 113506e73c..c5c417b0b2 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -454,10 +454,7 @@ def _get_fp8_mode_and_check_settings(
     assert load_in_fp8 is not False
     if load_in_fp8 is True:
         fp8_mode = "row"  # default
-        if (
-            not fast_inference
-            and os.environ.get("UNSLOTH_HAS_FBGEMM", "0") != "1"
-        ):
+        if not fast_inference and os.environ.get("UNSLOTH_HAS_FBGEMM", "0") != "1":
             fp8_mode = "block"
     else:
         fp8_mode = load_in_fp8