[gguf][torch.compile time] Convert to plain tensor earlier in dequantize_gguf_tensor (#13166)

[gguf] Convert to plain tensor earlier in dequantize_gguf_tensor

Once dequantize_gguf_tensor fetches the quant_type attributed from the
GGUFParamter tensor subclass, there is no further need of running the
actual dequantize operations on the Tensor subclass, we can just convert
to plain tensor right away.

This not only makes PyTorch eager faster, but reduces torch.compile
tracer compile time from 36 seconds to 10 seconds, because there is lot
less code to trace now.

.github/workflows/pr_modular_tests.yml

@@ -117,7 +117,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps

.github/workflows/pr_tests.yml

@@ -114,7 +114,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
@@ -191,7 +191,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers

.github/workflows/push_tests_mps.yml

@@ -41,7 +41,7 @@ jobs:
       shell: arch -arch arm64 bash {0}
       run: |
         ${CONDA_RUN} python -m pip install --upgrade pip uv
-        ${CONDA_RUN} python -m uv pip install -e ".[quality,test]"
+        ${CONDA_RUN} python -m uv pip install -e ".[quality]"
         ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio
         ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
         ${CONDA_RUN} python -m uv pip install transformers --upgrade

src/diffusers/models/attention_dispatch.py

@@ -38,7 +38,6 @@ from ..utils import (
     is_flash_attn_available,
     is_flash_attn_version,
     is_kernels_available,
-    is_kernels_version,
     is_sageattention_available,
     is_sageattention_version,
     is_torch_npu_available,
@@ -266,7 +265,6 @@ class _HubKernelConfig:
     repo_id: str
     function_attr: str
     revision: str | None = None
-    version: int | None = None
     kernel_fn: Callable | None = None
     wrapped_forward_attr: str | None = None
     wrapped_backward_attr: str | None = None
@@ -276,31 +274,27 @@ class _HubKernelConfig:
 
 # Registry for hub-based attention kernels
 _HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
+    # TODO: temporary revision for now. Remove when merged upstream into `main`.
     AttentionBackendName._FLASH_3_HUB: _HubKernelConfig(
-        repo_id="kernels-community/flash-attn3", function_attr="flash_attn_func", version=1
+        repo_id="kernels-community/flash-attn3", function_attr="flash_attn_func", revision="fake-ops-return-probs"
     ),
     AttentionBackendName._FLASH_3_VARLEN_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn3",
         function_attr="flash_attn_varlen_func",
-        version=1,
+        # revision="fake-ops-return-probs",
     ),
     AttentionBackendName.FLASH_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn2",
         function_attr="flash_attn_func",
-        version=1,
         revision=None,
         wrapped_forward_attr="flash_attn_interface._wrapped_flash_attn_forward",
         wrapped_backward_attr="flash_attn_interface._wrapped_flash_attn_backward",
     ),
     AttentionBackendName.FLASH_VARLEN_HUB: _HubKernelConfig(
-        repo_id="kernels-community/flash-attn2",
-        function_attr="flash_attn_varlen_func",
-        version=1,
+        repo_id="kernels-community/flash-attn2", function_attr="flash_attn_varlen_func", revision=None
     ),
     AttentionBackendName.SAGE_HUB: _HubKernelConfig(
-        repo_id="kernels-community/sage-attention",
-        function_attr="sageattn",
-        version=1,
+        repo_id="kernels-community/sage_attention", function_attr="sageattn", revision=None
     ),
 }
 
@@ -470,10 +464,6 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
             raise RuntimeError(
                 f"Backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
             )
-        if not is_kernels_version(">=", "0.12"):
-            raise RuntimeError(
-                f"Backend '{backend.value}' needs to be used with a `kernels` version of at least 0.12. Please update with `pip install -U kernels`."
-            )
 
     elif backend == AttentionBackendName.AITER:
         if not _CAN_USE_AITER_ATTN:

src/diffusers/quantizers/gguf/utils.py

@@ -516,6 +516,9 @@ def dequantize_gguf_tensor(tensor):
 
     block_size, type_size = GGML_QUANT_SIZES[quant_type]
 
+    # Conver to plain tensor to avoid unnecessary __torch_function__ overhead.
+    tensor = tensor.as_tensor()
+
     tensor = tensor.view(torch.uint8)
     shape = _quant_shape_from_byte_shape(tensor.shape, type_size, block_size)
 
@@ -525,7 +528,7 @@ def dequantize_gguf_tensor(tensor):
     dequant = dequant_fn(blocks, block_size, type_size)
     dequant = dequant.reshape(shape)
 
-    return dequant.as_tensor()
+    return dequant
 
 
 class GGUFParameter(torch.nn.Parameter):

src/diffusers/utils/__init__.py

@@ -86,7 +86,6 @@ from .import_utils import (
     is_inflect_available,
     is_invisible_watermark_available,
     is_kernels_available,
-    is_kernels_version,
     is_kornia_available,
     is_librosa_available,
     is_matplotlib_available,

src/diffusers/utils/import_utils.py

@@ -724,22 +724,6 @@ def is_transformers_version(operation: str, version: str):
     return compare_versions(parse(_transformers_version), operation, version)
 
 
-@cache
-def is_kernels_version(operation: str, version: str):
-    """
-    Compares the current Kernels version to a given reference with an operation.
-
-    Args:
-        operation (`str`):
-            A string representation of an operator, such as `">"` or `"<="`
-        version (`str`):
-            A version string
-    """
-    if not _kernels_available:
-        return False
-    return compare_versions(parse(_kernels_version), operation, version)
-
-
 @cache
 def is_hf_hub_version(operation: str, version: str):
     """

tests/models/testing_utils/lora.py

@@ -375,7 +375,7 @@ class LoraHotSwappingForModelTesterMixin:
             # additionally check if dynamic compilation works.
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output0_after = model(**inputs_dict)["sample"]
@@ -390,7 +390,7 @@ class LoraHotSwappingForModelTesterMixin:
         with torch.inference_mode():
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output1_after = model(**inputs_dict)["sample"]