update

* refactor how attention kernels from hub are used.

* up

* refactor according to Dhruv's ideas.

Co-authored-by: Dhruv Nair <dhruv@huggingface.co>

* empty

Co-authored-by: Dhruv Nair <dhruv@huggingface.co>

* empty

Co-authored-by: Dhruv Nair <dhruv@huggingface.co>

* empty

Co-authored-by: dn6 <dhruv@huggingface.co>

* up

---------

Co-authored-by: Dhruv Nair <dhruv@huggingface.co>
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

.github/workflows/nightly_tests.yml

@@ -73,6 +73,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #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 install pytest-reportlog
       - name: Environment
         run: |
@@ -84,7 +86,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -k "not Flax and not Onnx" \
+             -k "not Flax and not Onnx" \
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             --report-log=tests_pipeline_${{ matrix.module }}_cuda.log \
             tests/pipelines/${{ matrix.module }}
@@ -126,6 +128,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        #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 install pytest-reportlog
     - name: Environment
       run: python utils/print_env.py
@@ -190,6 +194,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
+        #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 
     - name: Environment
       run: |
         python utils/print_env.py
@@ -232,6 +238,8 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip install peft@git+https://github.com/huggingface/peft.git
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #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 install pytest-reportlog
       - name: Environment
         run: |
@@ -281,6 +289,8 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip install peft@git+https://github.com/huggingface/peft.git
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #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 
 
       - name: Environment
         run: |
@@ -358,6 +368,8 @@ jobs:
               uv pip install ${{ join(matrix.config.additional_deps, ' ') }}
           fi
           uv pip install pytest-reportlog
+          #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 
       - name: Environment
         run: |
           python utils/print_env.py
@@ -405,6 +417,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip install -U bitsandbytes optimum_quanto
+          #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 install pytest-reportlog
       - name: Environment
         run: |
@@ -531,7 +545,7 @@ jobs:
 #          HF_HOME: /System/Volumes/Data/mnt/cache
 #          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
 #        run: |
-#          ${CONDA_RUN} pytest -n 1 --make-reports=tests_torch_mps \
+#          ${CONDA_RUN} pytest -n 1  --make-reports=tests_torch_mps \
 #            --report-log=tests_torch_mps.log \
 #            tests/
 #      - name: Failure short reports
@@ -587,7 +601,7 @@ jobs:
 #          HF_HOME: /System/Volumes/Data/mnt/cache
 #          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
 #        run: |
-#          ${CONDA_RUN} pytest -n 1 --make-reports=tests_torch_mps \
+#          ${CONDA_RUN} pytest -n 1  --make-reports=tests_torch_mps \
 #            --report-log=tests_torch_mps.log \
 #            tests/
 #      - name: Failure short reports

.github/workflows/pr_modular_tests.yml

@@ -109,7 +109,8 @@ jobs:
     - name: Install dependencies
       run: |
         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 --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
 
     - name: Environment

.github/workflows/pr_tests.yml

@@ -115,7 +115,8 @@ jobs:
     - name: Install dependencies
       run: |
         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 --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
 
     - name: Environment
@@ -246,7 +247,8 @@ jobs:
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
-        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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |

.github/workflows/pr_tests_gpu.yml

@@ -1,4 +1,4 @@
-name: Fast GPU Tests on PR 
+name: Fast GPU Tests on PR
 
 on:
   pull_request:
@@ -71,7 +71,7 @@ jobs:
         if: ${{ failure() }}
         run: |
           echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
-  
+
   setup_torch_cuda_pipeline_matrix:
     needs: [check_code_quality, check_repository_consistency]
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
@@ -131,7 +131,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-          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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
       - name: Environment
         run: |
@@ -149,18 +150,18 @@ jobs:
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
-          if [ "${{ matrix.module }}" = "ip_adapters" ]; then 
+          if [ "${{ matrix.module }}" = "ip_adapters" ]; then
               pytest -n 1 --max-worker-restart=0 --dist=loadfile \
               -k "not Flax and not Onnx" \
               --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
               tests/pipelines/${{ matrix.module }}
-          else 
+          else
               pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
               pytest -n 1 --max-worker-restart=0 --dist=loadfile \
               -k "not Flax and not Onnx and $pattern" \
               --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
               tests/pipelines/${{ matrix.module }}
-          fi 
+          fi
 
       - name: Failure short reports
         if: ${{ failure() }}
@@ -201,7 +202,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -223,10 +225,10 @@ jobs:
         pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
         if [ -z "$pattern" ]; then
           pytest -n 1  --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx" tests/${{ matrix.module }} \
-          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+          --make-reports=tests_torch_cuda_${{ matrix.module }}
         else
           pytest -n 1  --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx and $pattern" tests/${{ matrix.module }} \
-          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+          --make-reports=tests_torch_cuda_${{ matrix.module }}
         fi
 
     - name: Failure short reports
@@ -262,7 +264,8 @@ jobs:
         nvidia-smi
     - name: Install dependencies
       run: |
-        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 --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 install -e ".[quality,training]"
 
     - name: Environment

.github/workflows/push_tests.yml

@@ -76,7 +76,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-          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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
       - name: Environment
         run: |
           python utils/print_env.py
@@ -128,7 +129,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -180,7 +182,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
-        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 --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
     - name: Environment
       run: |
         python utils/print_env.py

src/diffusers/models/attention_dispatch.py

@@ -16,6 +16,7 @@ import contextlib
 import functools
 import inspect
 import math
+from dataclasses import dataclass
 from enum import Enum
 from typing import TYPE_CHECKING, Any, Callable, Dict, List, Literal, Optional, Tuple, Union
 
@@ -42,7 +43,7 @@ from ..utils import (
     is_xformers_available,
     is_xformers_version,
 )
-from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS, DIFFUSERS_ENABLE_HUB_KERNELS
+from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS
 
 
 if TYPE_CHECKING:
@@ -82,24 +83,11 @@ else:
     flash_attn_3_func = None
     flash_attn_3_varlen_func = None
 
-
 if _CAN_USE_AITER_ATTN:
     from aiter import flash_attn_func as aiter_flash_attn_func
 else:
     aiter_flash_attn_func = None
 
-if DIFFUSERS_ENABLE_HUB_KERNELS:
-    if not is_kernels_available():
-        raise ImportError(
-            "To use FA3 kernel for your hardware from the Hub, the `kernels` library must be installed. Install with `pip install kernels`."
-        )
-    from ..utils.kernels_utils import _get_fa3_from_hub
-
-    flash_attn_interface_hub = _get_fa3_from_hub()
-    flash_attn_3_func_hub = flash_attn_interface_hub.flash_attn_func
-else:
-    flash_attn_3_func_hub = None
-
 if _CAN_USE_SAGE_ATTN:
     from sageattention import (
         sageattn,
@@ -261,6 +249,25 @@ class _AttentionBackendRegistry:
         return supports_context_parallel
 
 
+@dataclass
+class _HubKernelConfig:
+    """Configuration for downloading and using a hub-based attention kernel."""
+
+    repo_id: str
+    function_attr: str
+    revision: Optional[str] = None
+    kernel_fn: Optional[Callable] = None
+
+
+# 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", revision="fake-ops-return-probs"
+    )
+}
+
+
 @contextlib.contextmanager
 def attention_backend(backend: Union[str, AttentionBackendName] = AttentionBackendName.NATIVE):
     """
@@ -415,13 +422,9 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
 
     # TODO: add support Hub variant of FA3 varlen later
     elif backend in [AttentionBackendName._FLASH_3_HUB]:
-        if not DIFFUSERS_ENABLE_HUB_KERNELS:
-            raise RuntimeError(
-                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `DIFFUSERS_ENABLE_HUB_KERNELS` env var isn't set. Please set it like `export DIFFUSERS_ENABLE_HUB_KERNELS=yes`."
-            )
         if not is_kernels_available():
             raise RuntimeError(
-                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
+                f"Backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
             )
 
     elif backend == AttentionBackendName.AITER:
@@ -571,6 +574,29 @@ def _flex_attention_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
     return q_idx >= kv_idx
 
 
+# ===== Helpers for downloading kernels =====
+def _maybe_download_kernel_for_backend(backend: AttentionBackendName) -> None:
+    if backend not in _HUB_KERNELS_REGISTRY:
+        return
+    config = _HUB_KERNELS_REGISTRY[backend]
+
+    if config.kernel_fn is not None:
+        return
+
+    try:
+        from kernels import get_kernel
+
+        kernel_module = get_kernel(config.repo_id, revision=config.revision)
+        kernel_func = getattr(kernel_module, config.function_attr)
+
+        # Cache the downloaded kernel function in the config object
+        config.kernel_fn = kernel_func
+
+    except Exception as e:
+        logger.error(f"An error occurred while fetching kernel '{config.repo_id}' from the Hub: {e}")
+        raise
+
+
 # ===== torch op registrations =====
 # Registrations are required for fullgraph tracing compatibility
 # TODO: this is only required because the beta release FA3 does not have it. There is a PR adding
@@ -1418,7 +1444,8 @@ def _flash_attention_3_hub(
     return_attn_probs: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    out = flash_attn_3_func_hub(
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
+    out = func(
         q=query,
         k=key,
         v=value,

src/diffusers/models/modeling_utils.py

@@ -595,7 +595,11 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 attention as backend.
         """
         from .attention import AttentionModuleMixin
-        from .attention_dispatch import AttentionBackendName, _check_attention_backend_requirements
+        from .attention_dispatch import (
+            AttentionBackendName,
+            _check_attention_backend_requirements,
+            _maybe_download_kernel_for_backend,
+        )
 
         # TODO: the following will not be required when everything is refactored to AttentionModuleMixin
         from .attention_processor import Attention, MochiAttention
@@ -606,8 +610,10 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         available_backends = {x.value for x in AttentionBackendName.__members__.values()}
         if backend not in available_backends:
             raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
+
         backend = AttentionBackendName(backend)
         _check_attention_backend_requirements(backend)
+        _maybe_download_kernel_for_backend(backend)
 
         attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
         for module in self.modules():

src/diffusers/utils/constants.py

@@ -46,7 +46,6 @@ DIFFUSERS_ATTN_CHECKS = os.getenv("DIFFUSERS_ATTN_CHECKS", "0").upper() in ENV_V
 DEFAULT_HF_PARALLEL_LOADING_WORKERS = 8
 HF_ENABLE_PARALLEL_LOADING = os.environ.get("HF_ENABLE_PARALLEL_LOADING", "").upper() in ENV_VARS_TRUE_VALUES
 DIFFUSERS_DISABLE_REMOTE_CODE = os.getenv("DIFFUSERS_DISABLE_REMOTE_CODE", "false").upper() in ENV_VARS_TRUE_VALUES
-DIFFUSERS_ENABLE_HUB_KERNELS = os.environ.get("DIFFUSERS_ENABLE_HUB_KERNELS", "").upper() in ENV_VARS_TRUE_VALUES
 
 # Below should be `True` if the current version of `peft` and `transformers` are compatible with
 # PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are

src/diffusers/utils/kernels_utils.py

@@ -1,23 +0,0 @@
-from ..utils import get_logger
-from .import_utils import is_kernels_available
-
-
-logger = get_logger(__name__)
-
-
-_DEFAULT_HUB_ID_FA3 = "kernels-community/flash-attn3"
-
-
-def _get_fa3_from_hub():
-    if not is_kernels_available():
-        return None
-    else:
-        from kernels import get_kernel
-
-        try:
-            # TODO: temporary revision for now. Remove when merged upstream into `main`.
-            flash_attn_3_hub = get_kernel(_DEFAULT_HUB_ID_FA3, revision="fake-ops-return-probs")
-            return flash_attn_3_hub
-        except Exception as e:
-            logger.error(f"An error occurred while fetching kernel '{_DEFAULT_HUB_ID_FA3}' from the Hub: {e}")
-            raise

tests/others/test_attention_backends.py

@@ -7,7 +7,6 @@ To run this test suite:
 
 ```bash
 export RUN_ATTENTION_BACKEND_TESTS=yes
-export DIFFUSERS_ENABLE_HUB_KERNELS=yes
 
 pytest tests/others/test_attention_backends.py
 ```