triton-windows 3.4.0.post20__cp313-cp313-win_amd64.whl → 3.5.0.post21__cp313-cp313-win_amd64.whl

triton/tools/compile.py

@@ -3,12 +3,29 @@ import hashlib
 import importlib.util
 import sys
 from argparse import ArgumentParser
+from dataclasses import dataclass
 from pathlib import Path
 from typing import List
 
 import triton
 import triton.backends
-from triton.backends.nvidia.driver import ty_to_cpp
+
+
+@dataclass
+class CompileArgs:
+    '''
+    A class to contain arguments from command-line parser.
+    '''
+    path: str = ''
+    kernel_name: str = ''
+    signature: str = ''
+    grid: str = ''
+    target: str | None = None
+    num_warps: int = 1
+    num_stages: int = 3
+    out_name: str | None = None
+    out_path: Path | None = None
+
 
 desc = """
 Triton ahead-of-time compiler:
@@ -36,14 +53,18 @@ NOTE: when resolving the scope of /path/to/kernel.py, the file will be executed
 used to run this `compile.py` script
 """
 
-if __name__ == "__main__":
 
+def main():
     # command-line arguments
     parser = ArgumentParser(description=desc)
     parser.add_argument("path",
                         help="Path to Python source containing desired kernel in its scope. File will be executed.")
     parser.add_argument("--kernel-name", "-n", type=str, default="", help="Name of the kernel to compile",
                         required=True)
+    parser.add_argument(
+        "--target", "-t", type=str, default=None,
+        help="The target to compile towards, in format of '<backend>:<arch>:<warp-size>'; "
+        "e.g., 'cuda:80:32', 'hip:gfx942:64'. Default to None, which means using current machine's GPU target")
     parser.add_argument("--num-warps", "-w", type=int, default=1, help="Number of warps to launch the kernel")
     parser.add_argument("--num-stages", "-ns", type=int, default=3,
                         help="Number of stages (meta-parameter of the kernel)")
@@ -51,8 +72,12 @@ if __name__ == "__main__":
     parser.add_argument("--out-path", "-o", type=Path, default=None, help="Out filename")
     parser.add_argument("--signature", "-s", type=str, help="Signature of the kernel", required=True)
     parser.add_argument("--grid", "-g", type=str, help="Launch grid of the kernel", required=True)
-    args = parser.parse_args()
+    cli_args = parser.parse_args()
+    args = CompileArgs(**vars(cli_args))  # A sanity check to ensure class CompileArgs is updated as well.
+    compile_kernel(args)
 
+
+def compile_kernel(args: CompileArgs):
     out_name = args.out_name if args.out_name else args.kernel_name
     out_path = args.out_path if args.out_path else Path(out_name)
 
@@ -108,10 +133,18 @@ if __name__ == "__main__":
         assert h in [1, 16], f"Only 1 and 16 are valid hints, got {h}"
     attrs = {k: [["tt.divisibility", 16]] for k, v in hints.items() if v == 16}
     src = triton.compiler.ASTSource(fn=kernel, constexprs=constants, signature=signature, attrs=attrs)
-    opts = {"num_warps": args.num_warps, "num_stages": args.num_stages}
-    ccinfo = triton.compile(src, options=opts)
-    if ccinfo.metadata.global_scratch_size > 0:
+
+    target = triton.backends.compiler.GPUTarget(*args.target.split(":")) \
+        if args.target else triton.runtime.driver.active.get_current_target()
+    backend = triton.compiler.make_backend(target)
+    kwargs = {"num_warps": args.num_warps, "num_stages": args.num_stages}
+    options = backend.parse_options(kwargs)
+    ccinfo = triton.compile(src, target=target, options=options.__dict__)
+
+    if getattr(ccinfo.metadata, "global_scratch_size", 0) > 0:
         raise RuntimeError("AOT compiling kernels with global scratch requirements is not yet implemented")
+    if ccinfo.metadata.profile_scratch_size > 0:
+        raise RuntimeError("AOT compiling kernels with profile scratch requirements is not yet implemented")
 
     arg_names = []
     arg_types = []
@@ -136,8 +169,12 @@ if __name__ == "__main__":
         if hints.get((i, ), None) == 16:
             suffix += 'd'
     func_name = '_'.join([out_name, sig_hash, suffix])
-    asm = ccinfo.asm["cubin"]  # store binary data once
+    asm = ccinfo.asm[backend.binary_ext]  # store binary data once
+
     hex_ = str(binascii.hexlify(asm))[2:-1]
+
+    ty_to_cpp = triton.runtime.driver.active.map_python_to_cpp_type
+
     params = {
         "kernel_name": func_name,
         "triton_kernel_name": args.kernel_name,
@@ -145,18 +182,29 @@ if __name__ == "__main__":
         "bin_data": ", ".join([f"0x{x}{y}" for x, y in zip(hex_[::2], hex_[1::2])]),
         "signature": ", ".join([f"{ty_to_cpp(ty)} {name}" for name, ty in zip(arg_names_not_1, arg_types_not_1)]),
         "full_signature": ", ".join([f"{ty_to_cpp(ty)} {name}" for name, ty in zip(arg_names, arg_types)]),
-        "arg_pointers": ", ".join([f"&{arg}" for arg in arg_names_not_1] + ["&global_scratch"]),
-        "num_args": len(arg_names_not_1) + 1,
+        "arg_pointers": ", ".join([f"&{arg}" for arg in arg_names_not_1] + ["&global_scratch"] + ["&profile_scratch"]),
+        "num_args": len(arg_names_not_1) + 2,  # +2 for global and profile scratch
         "kernel_docstring": doc_string,
         "shared": ccinfo.metadata.shared,
         "num_warps": args.num_warps,
-        "algo_info": '_'.join([const_sig, meta_sig]),
+        "algo_info": "_".join([const_sig, meta_sig]),
         "gridX": grid[0],
         "gridY": grid[1],
         "gridZ": grid[2],
         "_placeholder": "",
     }
-    for ext in ['h', 'c']:
-        template_path = Path(__file__).parent / "extra" / "cuda" / f"compile.{ext}"
-        with out_path.with_suffix(f".{sig_hash}_{suffix}.{ext}").open("w") as fp:
-            fp.write(Path(template_path).read_text().format(**params))
+    output_files = []
+    backend_name = target.backend
+    template_dir = Path(__file__).parent / "extra" / backend_name
+    for template_path in template_dir.glob('compile.*'):
+        ext = template_path.suffix
+        output_file = out_path.with_suffix(f".{sig_hash}_{suffix}{ext}")
+        with output_file.open("w") as fp:
+            fp.write(template_path.read_text().format(**params))
+        output_files.append(output_file)
+
+    return func_name, output_files
+
+
+if __name__ == "__main__":
+    main()

triton/tools/extra/cuda/compile.c

@@ -61,6 +61,7 @@ CUresult {kernel_name}(CUstream stream, {signature}) {{
     unsigned int gY = {gridY};
     unsigned int gZ = {gridZ};
     CUdeviceptr global_scratch = 0;
+    CUdeviceptr profile_scratch = 0;
     void *args[{num_args}] = {{ {arg_pointers} }};
     // TODO: shared memory
     if(gX * gY * gZ > 0)

triton/tools/extra/hip/compile.cpp

@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+/* clang-format off */
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <hip/hip_runtime.h>
+
+// helpers to check for hip errors
+#define HIP_CHECK(ans) {{\
+    gpuAssert((ans), __FILE__, __LINE__);\
+  }}\
+
+static inline void gpuAssert(hipError_t code, const char *file, int line) {{
+  if (code != hipSuccess) {{
+    const char *prefix = "Triton Error [HIP]: ";
+    const char *str;
+    hipDrvGetErrorString(code, &str);
+    char err[1024] = {{0}};
+    strcat(err, prefix);
+    strcat(err, str);
+    printf("%s\\n", err);
+    exit(code);
+  }}
+}}
+
+// globals
+#define HSACO_NAME {kernel_name}_hsaco
+hipModule_t {kernel_name}_mod = nullptr;
+hipFunction_t {kernel_name}_func = nullptr;
+unsigned char HSACO_NAME[{bin_size}] = {{ {bin_data} }};
+
+
+void unload_{kernel_name}(void) {{
+    HIP_CHECK(hipModuleUnload({kernel_name}_mod));
+}}
+
+
+void load_{kernel_name}() {{
+    int dev = 0;
+    void *bin = (void *)&HSACO_NAME;
+    int shared = {shared};
+    HIP_CHECK(hipModuleLoadData(&{kernel_name}_mod, bin));
+    HIP_CHECK(hipModuleGetFunction(&{kernel_name}_func, {kernel_name}_mod, "{triton_kernel_name}"));
+}}
+
+/*
+{kernel_docstring}
+*/
+hipError_t {kernel_name}(hipStream_t stream, {signature}) {{
+    if ({kernel_name}_func == nullptr)
+       load_{kernel_name}();
+    unsigned int gX = {gridX};
+    unsigned int gY = {gridY};
+    unsigned int gZ = {gridZ};
+    hipDeviceptr_t global_scratch = 0;
+    hipDeviceptr_t profile_scratch = 0;
+    void *args[{num_args}] = {{ {arg_pointers} }};
+    // TODO: shared memory
+    if(gX * gY * gZ > 0)
+      return hipModuleLaunchKernel({kernel_name}_func, gX, gY, gZ, {num_warps} * warpSize, 1, 1, {shared}, stream, args, nullptr);
+    else
+      return hipErrorInvalidValue;
+}}

triton/tools/extra/hip/compile.h

@@ -0,0 +1,13 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <inttypes.h>
+#include <stdint.h>
+#include <stdio.h>
+
+void unload_{kernel_name}(void);
+void load_{kernel_name}(void);
+hipError_t{_placeholder} {kernel_name}(hipStream_t stream, {signature});

triton/tools/ragged_tma.py

@@ -0,0 +1,92 @@
+import triton
+import triton.language as tl
+from triton.tools.tensor_descriptor import TensorDescriptor
+
+# fmt: off
+
+
+def create_ragged_descriptor(T, block_shape, ragged_dim=0):
+    """
+    Given a 2- or 3-dimensional tensor T, this creates a 'ragged descriptor'
+    which behaves like a concatenation (along the first axis) of subarrays
+    of potentially unequal size.
+
+    The load_ragged and store_ragged device functions can be used to read
+    and write from subarrays T[batch_offset : batch_offset + batch_size]
+    with hardware bounds-checking preventing any sort of leakage outside
+    the subarray.
+    """
+
+    block_shape = list(block_shape)
+    tensor_shape = list(T.shape)
+    rank = len(tensor_shape)
+
+    if ragged_dim < 0:
+        ragged_dim += rank
+
+    assert 0 <= ragged_dim < rank - 1, "last dimension cannot be ragged"
+    assert rank <= 3, "read-write ragged descriptors must have at most 3 dimensions"
+
+    assert len(block_shape) == rank, "block shape must have same length as tensor shape"
+
+    max_int = 0x7fff0000
+    billion = 0x40000000  # == 2**30
+
+    assert tensor_shape[ragged_dim] <= billion, "number of rows may not exceed 2**30"
+    tensor_shape[ragged_dim] = billion
+    ragged_stride = T.stride(ragged_dim)
+
+    # we prepend an extra two dimensions and rely on the fact that pointers
+    # have 64-bit wraparound semantics:
+    tma_stride = [2**34 - ragged_stride, ragged_stride] + [T.stride(i) for i in range(rank)]
+    tma_shape  = [max_int, max_int] + tensor_shape
+    box_shape  = [1, 1] + block_shape
+
+    return TensorDescriptor(T, tma_shape, tma_stride, box_shape)
+
+
+@triton.jit
+def to_ragged_indices(batch_offset, batch_size, row):
+    """
+    Helper function for load_ragged and store_ragged.
+    """
+
+    billion = 0x40000000  # == 2**30
+    x = billion - batch_size + row
+    y = batch_offset + batch_size
+
+    return billion, y, x
+
+
+@triton.jit
+def load_ragged(TMA, batch_offset, batch_size, coords, ragged_dim: tl.constexpr = 0):
+    """
+    Read from a subarray T[batch_offset : batch_offset + batch_size] with
+    hardware bounds-checking, where reading outside the subarray gives zeros.
+
+    Coords should be an appropriately-sized list of integers, just like in
+    TMA.load().
+    """
+
+    tl.static_assert(len(TMA.shape) == len(coords) + 2, "TMA must be a read-write ragged descriptor")
+
+    c0, c1, c2 = to_ragged_indices(batch_offset, batch_size, coords[ragged_dim])
+    data = TMA.load([c0, c1] + coords[:ragged_dim] + [c2] + coords[ragged_dim + 1:])
+    data = tl.reshape(data, data.shape[2:])
+    return data
+
+
+@triton.jit
+def store_ragged(TMA, batch_offset, batch_size, coords, data, ragged_dim: tl.constexpr = 0):
+    """
+    Write to a subarray T[batch_offset : batch_offset + batch_size] with
+    hardware bounds-checking, where writes outside the subarray are masked
+    correctly.
+
+    Coords should be an appropriately-sized list of integers, just like in
+    TMA.store().
+    """
+
+    c0, c1, c2 = to_ragged_indices(batch_offset, batch_size, coords[ragged_dim])
+    data = tl.reshape(data, [1, 1] + data.shape)
+    TMA.store([c0, c1] + coords[:ragged_dim] + [c2] + coords[ragged_dim + 1:], data)

triton/tools/tensor_descriptor.py

@@ -9,6 +9,7 @@ class TensorDescriptor:
     shape: List[int]
     strides: List[int]
     block_shape: List[int]
+    padding: str = "zero"
 
     def __post_init__(self):
         rank = len(self.shape)
@@ -17,20 +18,17 @@ class TensorDescriptor:
         assert rank > 0, "rank must not be zero"
         assert rank <= 5, "rank cannot be more than 5"
         ty = type(self.base)
-        type_name = f"{ty.__module__}.{ty.__name__}"
-        if type_name not in ("torch.FakeTensor", "torch.FunctionalTensor"):
+        if ty.__name__ not in ("FakeTensor", "FunctionalTensor"):
             assert self.base.data_ptr() % 16 == 0, "base must be 16-byte aligned"
         validate_block_shape(self.block_shape)
         elem_bytes = self.base.dtype.itemsize
         for stride in self.strides[:-1]:
             assert (stride * elem_bytes) % 16 == 0, "strides must be 16-byte aligned"
         assert self.strides[-1] == 1, "Last dimension must be contiguous"
+        assert self.padding == "zero" or self.padding == "nan", "Illegal value for padding"
+        if self.padding == "nan":
+            assert self.base.dtype.is_floating_point, "Padding option `nan` is only supported for floating point tensors"
 
     @staticmethod
-    def from_tensor(tensor: Any, block_shape: List[int]):
-        return TensorDescriptor(
-            tensor,
-            tensor.shape,
-            tensor.stride(),
-            block_shape,
-        )
+    def from_tensor(tensor: Any, block_shape: List[int], padding="zero"):
+        return TensorDescriptor(tensor, tensor.shape, tensor.stride(), block_shape, padding)

triton/windows_utils.py

@@ -54,14 +54,11 @@ def max_version(
 
 
 def check_msvc(msvc_base_path: Path, version: str) -> bool:
-    return all(
-        x.exists()
-        for x in [
-            msvc_base_path / version / "bin" / "Hostx64" / "x64" / "cl.exe",
-            msvc_base_path / version / "include" / "vcruntime.h",
-            msvc_base_path / version / "lib" / "x64" / "vcruntime.lib",
-        ]
-    )
+    return all(x.exists() for x in [
+        msvc_base_path / version / "bin" / "Hostx64" / "x64" / "cl.exe",
+        msvc_base_path / version / "include" / "vcruntime.h",
+        msvc_base_path / version / "lib" / "x64" / "vcruntime.lib",
+    ])
 
 
 def find_msvc_env() -> tuple[Optional[Path], Optional[str]]:
@@ -72,20 +69,16 @@ def find_msvc_env() -> tuple[Optional[Path], Optional[str]]:
 
     version = os.getenv("VCToolsVersion")
     if not check_msvc(msvc_base_path, version):
-        warnings.warn(
-            f"Environment variables VCINSTALLDIR = {os.getenv('VCINSTALLDIR')}, "
-            f"VCToolsVersion = {os.getenv('VCToolsVersion')} are set, "
-            "but this MSVC installation is incomplete."
-        )
+        warnings.warn(f"Environment variables VCINSTALLDIR = {os.getenv('VCINSTALLDIR')}, "
+                      f"VCToolsVersion = {os.getenv('VCToolsVersion')} are set, "
+                      "but this MSVC installation is incomplete.")
         return None, None
 
     return msvc_base_path, version
 
 
 def find_msvc_vswhere() -> tuple[Optional[Path], Optional[str]]:
-    vswhere_path = find_in_program_files(
-        r"Microsoft Visual Studio\Installer\vswhere.exe"
-    )
+    vswhere_path = find_in_program_files(r"Microsoft Visual Studio\Installer\vswhere.exe")
     if vswhere_path is None:
         return None, None
 
@@ -111,9 +104,7 @@ def find_msvc_vswhere() -> tuple[Optional[Path], Optional[str]]:
     if not msvc_base_path.exists():
         return None, None
 
-    version = max_version(
-        os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
-    )
+    version = max_version(os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path))
     if version is None:
         return None, None
 
@@ -132,9 +123,7 @@ def find_msvc_envpath() -> tuple[Optional[Path], Optional[str]]:
         if not msvc_base_path.exists():
             continue
 
-        version = max_version(
-            os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
-        )
+        version = max_version(os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path))
         if version is None:
             continue
 
@@ -153,9 +142,7 @@ def find_msvc_hardcoded() -> tuple[Optional[Path], Optional[str]]:
     paths = sorted(paths)[::-1]
     for msvc_base_path in paths:
         msvc_base_path = Path(msvc_base_path)
-        version = max_version(
-            os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
-        )
+        version = max_version(os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path))
         if version is None:
             continue
         return msvc_base_path, version
@@ -188,13 +175,10 @@ def find_msvc(env_only: bool) -> tuple[Optional[str], list[str], list[str]]:
 
 
 def check_winsdk(winsdk_base_path: Path, version: str) -> bool:
-    return all(
-        x.exists()
-        for x in [
-            winsdk_base_path / "Include" / version / "ucrt" / "stdlib.h",
-            winsdk_base_path / "Lib" / version / "ucrt" / "x64" / "ucrt.lib",
-        ]
-    )
+    return all(x.exists() for x in [
+        winsdk_base_path / "Include" / version / "ucrt" / "stdlib.h",
+        winsdk_base_path / "Lib" / version / "ucrt" / "x64" / "ucrt.lib",
+    ])
 
 
 def find_winsdk_env() -> tuple[Optional[Path], Optional[str]]:
@@ -205,18 +189,16 @@ def find_winsdk_env() -> tuple[Optional[Path], Optional[str]]:
 
     version = os.getenv("WindowsSDKVersion")
     if version is None:
-        warnings.warn(
-            f"Environment variable WindowsSdkDir = {os.getenv('WindowsSdkDir')}, "
-            "but WindowsSDKVersion is not set."
-        )
+        version = os.getenv("WindowsSDKVer")
+    if version is None:
+        warnings.warn(f"Environment variable WindowsSdkDir = {winsdk_base_path}, "
+                      "but WindowsSDKVersion (or WindowsSDKVer) is not set.")
         return None, None
     version = version.rstrip("\\")
     if not check_winsdk(winsdk_base_path, version):
-        warnings.warn(
-            f"Environment variables WindowsSdkDir = {os.getenv('WindowsSdkDir')}, "
-            f"WindowsSDKVersion = {os.getenv('WindowsSDKVersion')} are set, "
-            "but this Windows SDK installation is incomplete."
-        )
+        warnings.warn(f"Environment variables WindowsSdkDir = {winsdk_base_path}, "
+                      f"WindowsSDKVersion (or WindowsSDKVer) = {version} are set, "
+                      "but this Windows SDK installation is incomplete.")
         return None, None
 
     return winsdk_base_path, version
@@ -225,9 +207,7 @@ def find_winsdk_env() -> tuple[Optional[Path], Optional[str]]:
 def find_winsdk_registry() -> tuple[Optional[Path], Optional[str]]:
     try:
         reg = winreg.ConnectRegistry(None, winreg.HKEY_LOCAL_MACHINE)
-        key = winreg.OpenKeyEx(
-            reg, r"SOFTWARE\WOW6432Node\Microsoft\Microsoft SDKs\Windows\v10.0"
-        )
+        key = winreg.OpenKeyEx(reg, r"SOFTWARE\WOW6432Node\Microsoft\Microsoft SDKs\Windows\v10.0")
         folder = winreg.QueryValueEx(key, "InstallationFolder")[0]
         winreg.CloseKey(key)
     except OSError:
@@ -294,9 +274,7 @@ def find_winsdk(env_only: bool) -> tuple[list[str], list[str]]:
 
 
 @functools.lru_cache
-def find_msvc_winsdk(
-    env_only: bool = False,
-) -> tuple[Optional[str], list[str], list[str]]:
+def find_msvc_winsdk(env_only: bool = False, ) -> tuple[Optional[str], list[str], list[str]]:
     msvc_bin_path, msvc_inc_dirs, msvc_lib_dirs = find_msvc(env_only)
     winsdk_inc_dirs, winsdk_lib_dirs = find_winsdk(env_only)
     return (
@@ -312,9 +290,9 @@ def find_python() -> list[str]:
     if sysconfig.get_config_var("Py_GIL_DISABLED"):
         version += "t"
     for python_base_path in [
-        sys.exec_prefix,
-        sys.base_exec_prefix,
-        os.path.dirname(sys.executable),
+            sys.exec_prefix,
+            sys.base_exec_prefix,
+            os.path.dirname(sys.executable),
     ]:
         python_lib_dir = Path(python_base_path) / "libs"
         if (python_lib_dir / f"python{version}.lib").exists():
@@ -326,14 +304,11 @@ def find_python() -> list[str]:
 
 def check_and_find_cuda(base_path: Path) -> tuple[Optional[str], list[str], list[str]]:
     # pip
-    if all(
-        x.exists()
-        for x in [
+    if all(x.exists() for x in [
             base_path / "cuda_nvcc" / "bin" / "ptxas.exe",
             base_path / "cuda_runtime" / "include" / "cuda.h",
             base_path / "cuda_runtime" / "lib" / "x64" / "cuda.lib",
-        ]
-    ):
+    ]):
         return (
             str(base_path / "cuda_nvcc" / "bin"),
             [str(base_path / "cuda_runtime" / "include")],
@@ -341,14 +316,11 @@ def check_and_find_cuda(base_path: Path) -> tuple[Optional[str], list[str], list
         )
 
     # conda
-    if all(
-        x.exists()
-        for x in [
+    if all(x.exists() for x in [
             base_path / "bin" / "ptxas.exe",
             base_path / "include" / "cuda.h",
             base_path / "lib" / "cuda.lib",
-        ]
-    ):
+    ]):
         return (
             str(base_path / "bin"),
             [str(base_path / "include")],
@@ -356,14 +328,11 @@ def check_and_find_cuda(base_path: Path) -> tuple[Optional[str], list[str], list
         )
 
     # bundled or system-wide
-    if all(
-        x.exists()
-        for x in [
+    if all(x.exists() for x in [
             base_path / "bin" / "ptxas.exe",
             base_path / "include" / "cuda.h",
             base_path / "lib" / "x64" / "cuda.lib",
-        ]
-    ):
+    ]):
         return (
             str(base_path / "bin"),
             [str(base_path / "include")],
@@ -380,9 +349,7 @@ def find_cuda_env() -> tuple[Optional[str], list[str], list[str]]:
             continue
 
         cuda_base_path = Path(cuda_base_path)
-        cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = check_and_find_cuda(
-            cuda_base_path
-        )
+        cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = check_and_find_cuda(cuda_base_path)
         if cuda_bin_path:
             return cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs
 
@@ -390,9 +357,7 @@ def find_cuda_env() -> tuple[Optional[str], list[str], list[str]]:
 
 
 def find_cuda_bundled() -> tuple[Optional[str], list[str], list[str]]:
-    cuda_base_path = (
-        Path(sysconfig.get_paths()["platlib"]) / "triton" / "backends" / "nvidia"
-    )
+    cuda_base_path = (Path(sysconfig.get_paths()["platlib"]) / "triton" / "backends" / "nvidia")
     return check_and_find_cuda(cuda_base_path)
 
 
@@ -416,9 +381,7 @@ def find_cuda_hardcoded() -> tuple[Optional[str], list[str], list[str]]:
     paths = sorted(paths)[::-1]
     for cuda_base_path in paths:
         cuda_base_path = Path(cuda_base_path)
-        cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = check_and_find_cuda(
-            cuda_base_path
-        )
+        cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = check_and_find_cuda(cuda_base_path)
         if cuda_bin_path:
             return cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs
 
@@ -428,11 +391,11 @@ def find_cuda_hardcoded() -> tuple[Optional[str], list[str], list[str]]:
 @functools.lru_cache
 def find_cuda() -> tuple[Optional[str], list[str], list[str]]:
     for f in [
-        find_cuda_env,
-        find_cuda_bundled,
-        find_cuda_pip,
-        find_cuda_conda,
-        find_cuda_hardcoded,
+            find_cuda_env,
+            find_cuda_bundled,
+            find_cuda_pip,
+            find_cuda_conda,
+            find_cuda_hardcoded,
     ]:
         cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = f()
         if cuda_bin_path:

{triton_windows-3.4.0.post20.dist-info → triton_windows-3.5.0.post21.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: triton-windows
-Version: 3.4.0.post20
+Version: 3.5.0.post21
 Summary: A language and compiler for custom Deep Learning operations
 Home-page: https://github.com/woct0rdho/triton-windows
 Author: Philippe Tillet, Dian Wu
@@ -10,14 +10,13 @@ Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Developers
 Classifier: Topic :: Software Development :: Build Tools
 Classifier: License :: OSI Approved :: MIT License
-Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
-Requires-Python: >=3.9,<3.14
+Classifier: Programming Language :: Python :: 3.14
+Requires-Python: >=3.10,<3.15
 License-File: LICENSE
-Requires-Dist: setuptools>=40.8.0
 Requires-Dist: importlib-metadata; python_version < "3.10"
 Provides-Extra: build
 Requires-Dist: cmake<4.0,>=3.20; extra == "build"