onnxruntime-directml 1.22.1.dev20250710002__cp313-cp313-win_amd64.whl → 1.24.1__cp313-cp313-win_amd64.whl

onnxruntime/ThirdPartyNotices.txt

@@ -5806,41 +5806,6 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 _____
 
-composable_kernel
-
-https://github.com/ROCmSoftwarePlatform/composable_kernel
-
-Copyright (c) 2018-    , Advanced Micro Devices, Inc. (Chao Liu, Jing Zhang)
-Copyright (c) 2019-    , Advanced Micro Devices, Inc. (Letao Qin, Qianfeng Zhang, Liang Huang, Shaojie Wang)
-Copyright (c) 2022-    , Advanced Micro Devices, Inc. (Anthony Chang, Chunyu Lai, Illia Silin, Adam Osewski, Poyen Chen, Jehandad Khan)
-Copyright (c) 2019-2021, Advanced Micro Devices, Inc. (Hanwen Chang)
-Copyright (c) 2019-2020, Advanced Micro Devices, Inc. (Tejash Shah)
-Copyright (c) 2020     , Advanced Micro Devices, Inc. (Xiaoyan Zhou)
-Copyright (c) 2021-2022, Advanced Micro Devices, Inc. (Jianfeng Yan)
-
-SPDX-License-Identifier: MIT
-Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-
-_____
-
 neural-speed
 
 https://github.com/intel/neural-speed

onnxruntime/__init__.py

@@ -8,7 +8,9 @@ For more information on ONNX Runtime, please see `aka.ms/onnxruntime <https://ak
 or the `Github project <https://github.com/microsoft/onnxruntime/>`_.
 """
 
-__version__ = "1.22.1"
+import contextlib
+
+__version__ = "1.24.1"
 __author__ = "Microsoft"
 
 # we need to do device version validation (for example to check Cuda version for an onnxruntime-training package).
@@ -30,9 +32,20 @@ try:
         NodeArg,  # noqa: F401
         OrtAllocatorType,  # noqa: F401
         OrtArenaCfg,  # noqa: F401
+        OrtCompileApiFlags,  # noqa: F401
+        OrtDeviceMemoryType,  # noqa: F401
+        OrtEpAssignedNode,  # noqa: F401
+        OrtEpAssignedSubgraph,  # noqa: F401
+        OrtEpDevice,  # noqa: F401
+        OrtExecutionProviderDevicePolicy,  # noqa: F401
+        OrtExternalInitializerInfo,  # noqa: F401
+        OrtHardwareDevice,  # noqa: F401
+        OrtHardwareDeviceType,  # noqa: F401
         OrtMemoryInfo,  # noqa: F401
+        OrtMemoryInfoDeviceType,  # noqa: F401
         OrtMemType,  # noqa: F401
         OrtSparseFormat,  # noqa: F401
+        OrtSyncStream,  # noqa: F401
         RunOptions,  # noqa: F401
         SessionIOBinding,  # noqa: F401
         SessionOptions,  # noqa: F401
@@ -44,11 +57,15 @@ try:
         get_available_providers,  # noqa: F401
         get_build_info,  # noqa: F401
         get_device,  # noqa: F401
+        get_ep_devices,  # noqa: F401
         get_version_string,  # noqa: F401
         has_collective_ops,  # noqa: F401
+        register_execution_provider_library,  # noqa: F401
         set_default_logger_severity,  # noqa: F401
         set_default_logger_verbosity,  # noqa: F401
+        set_global_thread_pool_sizes,  # noqa: F401
         set_seed,  # noqa: F401
+        unregister_execution_provider_library,  # noqa: F401
     )
 
     import_capi_exception = None
@@ -64,9 +81,11 @@ from onnxruntime.capi.onnxruntime_inference_collection import (
     AdapterFormat,  # noqa: F401
     InferenceSession,  # noqa: F401
     IOBinding,  # noqa: F401
+    ModelCompiler,  # noqa: F401
     OrtDevice,  # noqa: F401
     OrtValue,  # noqa: F401
     SparseTensor,  # noqa: F401
+    copy_tensors,  # noqa: F401
 )
 
 # TODO: thiagofc: Temporary experimental namespace for new PyTorch front-end
@@ -85,7 +104,7 @@ onnxruntime_validation.check_distro_info()
 
 
 def _get_package_version(package_name: str):
-    from importlib.metadata import PackageNotFoundError, version
+    from importlib.metadata import PackageNotFoundError, version  # noqa: PLC0415
 
     try:
         package_version = version(package_name)
@@ -95,7 +114,7 @@ def _get_package_version(package_name: str):
 
 
 def _get_package_root(package_name: str, directory_name: str | None = None):
-    from importlib.metadata import PackageNotFoundError, distribution
+    from importlib.metadata import PackageNotFoundError, distribution  # noqa: PLC0415
 
     root_directory_name = directory_name or package_name
     try:
@@ -118,14 +137,43 @@ def _get_package_root(package_name: str, directory_name: str | None = None):
     return None
 
 
+def _extract_cuda_major_version(version_str: str) -> str:
+    """Extract CUDA major version from version string (e.g., '12.1' -> '12').
+
+    Args:
+        version_str: CUDA version string to parse
+
+    Returns:
+        Major version as string, or "12" if parsing fails
+    """
+    return version_str.split(".")[0] if version_str else "12"
+
+
+def _get_cufft_version(cuda_major: str) -> str:
+    """Get cufft library version based on CUDA major version.
+
+    Args:
+        cuda_major: CUDA major version as string (e.g., "12", "13")
+
+    Returns:
+        cufft version as string
+    """
+    # cufft versions: CUDA 12.x -> 11, CUDA 13.x -> 12
+    return "12" if cuda_major == "13" else "11"
+
+
 def _get_nvidia_dll_paths(is_windows: bool, cuda: bool = True, cudnn: bool = True):
+    # Dynamically determine CUDA major version from build info
+    cuda_major_version = _extract_cuda_major_version(cuda_version)
+    cufft_version = _get_cufft_version(cuda_major_version)
+
     if is_windows:
         # Path is relative to site-packages directory.
         cuda_dll_paths = [
-            ("nvidia", "cublas", "bin", "cublasLt64_12.dll"),
-            ("nvidia", "cublas", "bin", "cublas64_12.dll"),
-            ("nvidia", "cufft", "bin", "cufft64_11.dll"),
-            ("nvidia", "cuda_runtime", "bin", "cudart64_12.dll"),
+            ("nvidia", "cublas", "bin", f"cublasLt64_{cuda_major_version}.dll"),
+            ("nvidia", "cublas", "bin", f"cublas64_{cuda_major_version}.dll"),
+            ("nvidia", "cufft", "bin", f"cufft64_{cufft_version}.dll"),
+            ("nvidia", "cuda_runtime", "bin", f"cudart64_{cuda_major_version}.dll"),
         ]
         cudnn_dll_paths = [
             ("nvidia", "cudnn", "bin", "cudnn_engines_runtime_compiled64_9.dll"),
@@ -139,12 +187,12 @@ def _get_nvidia_dll_paths(is_windows: bool, cuda: bool = True, cudnn: bool = Tru
     else:  # Linux
         # cublas64 depends on cublasLt64, so cublasLt64 should be loaded first.
         cuda_dll_paths = [
-            ("nvidia", "cublas", "lib", "libcublasLt.so.12"),
-            ("nvidia", "cublas", "lib", "libcublas.so.12"),
-            ("nvidia", "cuda_nvrtc", "lib", "libnvrtc.so.12"),
+            ("nvidia", "cublas", "lib", f"libcublasLt.so.{cuda_major_version}"),
+            ("nvidia", "cublas", "lib", f"libcublas.so.{cuda_major_version}"),
+            ("nvidia", "cuda_nvrtc", "lib", f"libnvrtc.so.{cuda_major_version}"),
             ("nvidia", "curand", "lib", "libcurand.so.10"),
-            ("nvidia", "cufft", "lib", "libcufft.so.11"),
-            ("nvidia", "cuda_runtime", "lib", "libcudart.so.12"),
+            ("nvidia", "cufft", "lib", f"libcufft.so.{cufft_version}"),
+            ("nvidia", "cuda_runtime", "lib", f"libcudart.so.{cuda_major_version}"),
         ]
 
         # Do not load cudnn sub DLLs (they will be dynamically loaded later) to be consistent with PyTorch in Linux.
@@ -157,10 +205,10 @@ def _get_nvidia_dll_paths(is_windows: bool, cuda: bool = True, cudnn: bool = Tru
 
 def print_debug_info():
     """Print information to help debugging."""
-    import importlib.util
-    import os
-    import platform
-    from importlib.metadata import distributions
+    import importlib.util  # noqa: PLC0415
+    import os  # noqa: PLC0415
+    import platform  # noqa: PLC0415
+    from importlib.metadata import distributions  # noqa: PLC0415
 
     print(f"{package_name} version: {__version__}")
     if cuda_version:
@@ -186,15 +234,17 @@ def print_debug_info():
 
     if cuda_version:
         # Print version of installed packages that is related to CUDA or cuDNN DLLs.
+        cuda_major = _extract_cuda_major_version(cuda_version)
+
         packages = [
             "torch",
-            "nvidia-cuda-runtime-cu12",
-            "nvidia-cudnn-cu12",
-            "nvidia-cublas-cu12",
-            "nvidia-cufft-cu12",
-            "nvidia-curand-cu12",
-            "nvidia-cuda-nvrtc-cu12",
-            "nvidia-nvjitlink-cu12",
+            f"nvidia-cuda-runtime-cu{cuda_major}",
+            f"nvidia-cudnn-cu{cuda_major}",
+            f"nvidia-cublas-cu{cuda_major}",
+            f"nvidia-cufft-cu{cuda_major}",
+            f"nvidia-curand-cu{cuda_major}",
+            f"nvidia-cuda-nvrtc-cu{cuda_major}",
+            f"nvidia-nvjitlink-cu{cuda_major}",
         ]
         for package in packages:
             directory_name = "nvidia" if package.startswith("nvidia-") else None
@@ -205,9 +255,9 @@ def print_debug_info():
                 print(f"{package} not installed")
 
     if platform.system() == "Windows":
-        print(f"\nEnvironment variable:\nPATH={os.environ['PATH']}")
+        print(f"\nEnvironment variable:\nPATH={os.environ.get('PATH', '(unset)')}")
     elif platform.system() == "Linux":
-        print(f"\nEnvironment variable:\nLD_LIBRARY_PATH={os.environ['LD_LIBRARY_PATH']}")
+        print(f"\nEnvironment variable:\nLD_LIBRARY_PATH={os.environ.get('LD_LIBRARY_PATH', '(unset)')}")
 
     if importlib.util.find_spec("psutil"):
 
@@ -217,7 +267,7 @@ def print_debug_info():
                 target_keywords = ["cufft", "cublas", "cudart", "nvrtc", "curand", "cudnn", *target_keywords]
             return any(keyword in path for keyword in target_keywords)
 
-        import psutil
+        import psutil  # noqa: PLC0415
 
         p = psutil.Process(os.getpid())
 
@@ -228,7 +278,7 @@ def print_debug_info():
 
         if cuda_version:
             if importlib.util.find_spec("cpuinfo") and importlib.util.find_spec("py3nvml"):
-                from .transformers.machine_info import get_device_info
+                from .transformers.machine_info import get_device_info  # noqa: PLC0415
 
                 print("\nDevice information:")
                 print(get_device_info())
@@ -239,7 +289,7 @@ def print_debug_info():
 
 
 def preload_dlls(cuda: bool = True, cudnn: bool = True, msvc: bool = True, directory=None):
-    """Preload CUDA 12.x and cuDNN 9.x DLLs in Windows or Linux, and MSVC runtime DLLs in Windows.
+    """Preload CUDA 12.x+ and cuDNN 9.x DLLs in Windows or Linux, and MSVC runtime DLLs in Windows.
 
        When the installed PyTorch is compatible (using same major version of CUDA and cuDNN),
        there is no need to call this function if `import torch` is done before `import onnxruntime`.
@@ -255,10 +305,10 @@ def preload_dlls(cuda: bool = True, cudnn: bool = True, msvc: bool = True, direc
            If directory is empty string (""), the search order: nvidia site packages, default DLL loading paths.
            If directory is a path, the search order: the directory, default DLL loading paths.
     """
-    import ctypes
-    import os
-    import platform
-    import sys
+    import ctypes  # noqa: PLC0415
+    import os  # noqa: PLC0415
+    import platform  # noqa: PLC0415
+    import sys  # noqa: PLC0415
 
     if platform.system() not in ["Windows", "Linux"]:
         return
@@ -274,30 +324,53 @@ def preload_dlls(cuda: bool = True, cudnn: bool = True, msvc: bool = True, direc
             print("Microsoft Visual C++ Redistributable is not installed, this may lead to the DLL load failure.")
             print("It can be downloaded at https://aka.ms/vs/17/release/vc_redist.x64.exe.")
 
-    if not (cuda_version and cuda_version.startswith("12.")) and (cuda or cudnn):
-        print(
-            f"\033[33mWARNING: {package_name} is not built with CUDA 12.x support. "
-            "Please install a version that supports CUDA 12.x, or call preload_dlls with cuda=False and cudnn=False.\033[0m"
-        )
-        return
-
-    if not (cuda_version and cuda_version.startswith("12.") and (cuda or cudnn)):
+    # Check if CUDA version is supported (12.x or 13.x+)
+    ort_cuda_major = None
+    if cuda_version:
+        try:
+            ort_cuda_major = int(cuda_version.split(".")[0])
+            if ort_cuda_major < 12 and (cuda or cudnn):
+                print(
+                    f"\033[33mWARNING: {package_name} is built with CUDA {cuda_version}, which is not supported for preloading. "
+                    f"CUDA 12.x or newer is required. Call preload_dlls with cuda=False and cudnn=False.\033[0m"
+                )
+                return
+        except ValueError:
+            print(
+                f"\033[33mWARNING: Unable to parse CUDA version '{cuda_version}'. "
+                "Skipping DLL preloading. Call preload_dlls with cuda=False and cudnn=False.\033[0m"
+            )
+            return
+    elif cuda or cudnn:
+        # No CUDA version info available but CUDA/cuDNN preloading requested
         return
 
     is_cuda_cudnn_imported_by_torch = False
 
     if is_windows:
         torch_version = _get_package_version("torch")
-        is_torch_for_cuda_12 = torch_version and "+cu12" in torch_version
+        # Check if torch CUDA version matches onnxruntime CUDA version
+        torch_cuda_major = None
+        if torch_version and "+cu" in torch_version:
+            with contextlib.suppress(ValueError):
+                # Extract CUDA version from torch (e.g., "2.0.0+cu121" -> 12)
+                cu_part = torch_version.split("+cu")[1]
+                torch_cuda_major = int(cu_part[:2])  # First 2 digits are major version
+
+        is_torch_cuda_compatible = (
+            torch_cuda_major == ort_cuda_major if (torch_cuda_major and ort_cuda_major) else False
+        )
+
         if "torch" in sys.modules:
-            is_cuda_cudnn_imported_by_torch = is_torch_for_cuda_12
-            if (torch_version and "+cu" in torch_version) and not is_torch_for_cuda_12:
+            is_cuda_cudnn_imported_by_torch = is_torch_cuda_compatible
+            if torch_cuda_major and ort_cuda_major and torch_cuda_major != ort_cuda_major:
                 print(
-                    f"\033[33mWARNING: The installed PyTorch {torch_version} does not support CUDA 12.x. "
-                    f"Please install PyTorch for CUDA 12.x to be compatible with {package_name}.\033[0m"
+                    f"\033[33mWARNING: The installed PyTorch {torch_version} uses CUDA {torch_cuda_major}.x, "
+                    f"but {package_name} is built with CUDA {ort_cuda_major}.x. "
+                    f"Please install PyTorch for CUDA {ort_cuda_major}.x to be compatible.\033[0m"
                 )
 
-        if is_torch_for_cuda_12 and directory is None:
+        if is_torch_cuda_compatible and directory is None:
             torch_root = _get_package_root("torch", "torch")
             if torch_root:
                 directory = os.path.join(torch_root, "lib")

onnxruntime/capi/build_and_package_info.py

@@ -1,2 +1,2 @@
 package_name = 'onnxruntime-directml'
-__version__ = '1.22.1.dev20250710002'
+__version__ = '1.24.1'