update

.github/workflows/nightly_tests.yml

@@ -333,7 +333,7 @@ jobs:
             additional_deps: ["peft"]
           - backend: "gguf"
             test_location: "gguf"
-            additional_deps: ["peft"]
+            additional_deps: ["peft", "kernels"]
           - backend: "torchao"
             test_location: "torchao"
             additional_deps: []

docs/source/en/quantization/gguf.md

@@ -61,7 +61,7 @@ Optimized CUDA kernels can accelerate GGUF quantized model inference by approxim
 pip install -U kernels
 ```
 
-Once installed, GGUF inference automatically uses optimized kernels when available. Note that CUDA kernels may introduce minor numerical differences compared to the original GGUF implementation, potentially causing subtle visual variations in generated images. To disable CUDA kernel usage, set the environment variable `DIFFUSERS_GGUF_CUDA_KERNELS=false`.
+Once installed, set `DIFFUSERS_GGUF_CUDA_KERNELS=true`  to use optimized kernels when available. Note that CUDA kernels may introduce minor numerical differences compared to the original GGUF implementation, potentially causing subtle visual variations in generated images. To disable CUDA kernel usage, set the environment variable `DIFFUSERS_GGUF_CUDA_KERNELS=false`.
 
 ## Supported Quantization Types
 

src/diffusers/quantizers/gguf/utils.py

@@ -30,7 +30,7 @@ if is_accelerate_available():
 
 
 can_use_cuda_kernels = (
-    os.getenv("DIFFUSERS_GGUF_CUDA_KERNELS", "true").lower() in ["1", "true", "yes"]
+    os.getenv("DIFFUSERS_GGUF_CUDA_KERNELS", "false").lower() in ["1", "true", "yes"]
     and torch.cuda.is_available()
     and torch.cuda.get_device_capability()[0] >= 7
 )

src/diffusers/utils/testing_utils.py

@@ -35,6 +35,7 @@ from .import_utils import (
     is_compel_available,
     is_flax_available,
     is_gguf_available,
+    is_kernels_available,
     is_note_seq_available,
     is_onnx_available,
     is_opencv_available,
@@ -629,6 +630,18 @@ def require_torchao_version_greater_or_equal(torchao_version):
     return decorator
 
 
+def require_kernels_greater_or_equal(kernels_version):
+    def decorator(test_case):
+        correct_kernels_version = is_kernels_available() and version.parse(
+            version.parse(importlib.metadata.version("kernels")).base_version
+        ) >= version.parse(kernels_version)
+        return unittest.skipUnless(
+            correct_kernels_version, f"Test requires kernels with version greater than {kernels_version}."
+        )(test_case)
+
+    return decorator
+
+
 def deprecate_after_peft_backend(test_case):
     """
     Decorator marking a test that will be skipped after PEFT backend

tests/quantization/gguf/test_gguf.py

@@ -32,6 +32,7 @@ from diffusers.utils.testing_utils import (
     require_accelerator,
     require_big_accelerator,
     require_gguf_version_greater_or_equal,
+    require_kernels_version_greater_or_equal,
     require_peft_backend,
     torch_device,
 )
@@ -49,6 +50,7 @@ enable_full_determinism()
 @require_accelerate
 @require_accelerator
 @require_gguf_version_greater_or_equal("0.10.0")
+@require_kernels_version_greater_or_equal("0.9.0")
 class GGUFCudaKernelsTests(unittest.TestCase):
     def setUp(self):
         gc.collect()