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

triton/backends/nvidia/compiler.py

@@ -121,7 +121,7 @@ class CUDAOptions:
     enable_fp_fusion: bool = True
     launch_cooperative_grid: bool = False
     launch_pdl: bool = False
-    supported_fp8_dtypes: Tuple[str] = ("fp8e5", "fp8e4b15")
+    supported_fp8_dtypes: Tuple[str] = ("fp8e4nv", "fp8e5", "fp8e4b15")
     deprecated_fp8_dot_operand_dtypes: Tuple[str] = ()
     default_dot_input_precision: str = "tf32"
     allowed_dot_input_precisions: Tuple[str] = ("tf32", "tf32x3", "ieee")
@@ -177,8 +177,6 @@ class CUDABackend(BaseBackend):
 
         if "supported_fp8_dtypes" not in args:
             supported_fp8_dtypes = set(CUDAOptions.supported_fp8_dtypes)
-            if capability >= 89:
-                supported_fp8_dtypes.add("fp8e4nv")
             args["supported_fp8_dtypes"] = tuple(sorted(supported_fp8_dtypes))
 
         if "deprecated_fp8_dot_operand_dtypes" not in args:
@@ -433,7 +431,8 @@ class CUDABackend(BaseBackend):
             try:
                 # close_fds=True on Windows and False on Linux, see https://github.com/triton-lang/triton/pull/4357
                 # On Windows, both stdout and stderr need to be redirected to flog
-                subprocess.run(ptxas_cmd, check=True, close_fds=True if os.name == 'nt' else False, stdout=flog, stderr=flog)
+                subprocess.run(ptxas_cmd, check=True, close_fds=True if os.name == 'nt' else False, stdout=flog,
+                               stderr=flog)
             except subprocess.CalledProcessError as e:
                 with open(flog.name) as log_file:
                     log = log_file.read()

triton/runtime/build.py

@@ -116,7 +116,7 @@ def _build(name: str, src: str, srcdir: str, library_dirs: list[str], include_di
     cc_cmd = _cc_cmd(cc, src, so, include_dirs, library_dirs, libraries)
 
     try:
-        ret = subprocess.check_call(cc_cmd)
+        subprocess.check_call(cc_cmd)
     except Exception as e:
         print("Failed to compile. cc_cmd:", cc_cmd)
         raise e

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-3.4.0.post21.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: triton-windows
-Version: 3.4.0.post20
+Version: 3.4.0.post21
 Summary: A language and compiler for custom Deep Learning operations
 Home-page: https://github.com/woct0rdho/triton-windows
 Author: Philippe Tillet, Dian Wu

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

@@ -5,8 +5,8 @@ triton/_utils.py,sha256=XTYb3qDDaVmbhmXbm6ChMTYTW9jeE538jZwJE_eliQg,3539
 triton/errors.py,sha256=8WfnuRKLG578mgY6cBA3ECruVMf9ULEKFNgRcJ6IhWM,89
 triton/knobs.py,sha256=VOdRM_J0TejBYP2H7QZZzLwcRJq-Eppm9hBxDt1pCgA,14916
 triton/testing.py,sha256=vbEQRNrOnnzRQvVVSaiZrUo8AC0XPV40GJxfvkKYLh0,20276
-triton/windows_utils.py,sha256=7uwsR-yQZQzzvHajkukryRTJlOotyRA5mlNlJOFVGsE,13130
-triton/_C/libtriton.pyd,sha256=feY3eFQa0O4dU5046FmNzjDDJeYUoZfWp82hHDU1PCo,89961984
+triton/windows_utils.py,sha256=JMi6mjOApzh2-cw3Wl_nl6ji7JkwexYI7xgo2Et3ihU,12903
+triton/_C/libtriton.pyd,sha256=_TR8zPM-fGXdMumvlF2VH2hQHqeJ79qanE6SqlVgjN4,89965056
 triton/backends/__init__.py,sha256=X7290kf96Fk9QnfLScsX4UDG3zPyH_-31E4A7pVOijM,1612
 triton/backends/compiler.py,sha256=MY2_cQG26p68z8VwRv2Nlj_h2DfEhwBbN-30caMgep0,2840
 triton/backends/driver.py,sha256=AN60upJlPgia0JwvZ8vIVgLITNPuI0fdz8zMIIHPpF4,1450
@@ -18,7 +18,7 @@ triton/backends/amd/lib/asanrtl.bc,sha256=1xv2RlU3WvbdsghHlmhwiHewGM2B5dKts5bERM
 triton/backends/amd/lib/ockl.bc,sha256=wQKCzkKukIHbu0lyjKUYlhndc7S27xto6L54J0Bn-C0,246124
 triton/backends/amd/lib/ocml.bc,sha256=UPNTXW0gCXUNB-c6orSYwb-mz9_mjUc7zny_vfFza44,205964
 triton/backends/nvidia/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-triton/backends/nvidia/compiler.py,sha256=9ebNiX61uhApSojj5zf51HCh750FjXk2nzYlVQDJudk,19495
+triton/backends/nvidia/compiler.py,sha256=7Yf79DdocAgYaWLR-MSTAv_QAJjVAZkhbr1-ovsnfnQ,19452
 triton/backends/nvidia/driver.c,sha256=rH8RDtMMv_UHr7qiLnrSdNg3xojOQe_fF1zW67LFjaE,17882
 triton/backends/nvidia/driver.py,sha256=22YT4HhTPI1aJ-mjw31Lvw1mcffj17pGJry3m9W6jio,26391
 triton/backends/nvidia/bin/ptxas.exe,sha256=f28E0l5aerLAfBXk7yagfOwIEE6_6_NkMx-vqPPEQ9Y,24753152
@@ -66,7 +66,7 @@ triton/language/extra/hip/libdevice.py,sha256=Rf-AmBzcO6ORVzSxSuLXOy0lpoZTsnRAuT
 triton/runtime/__init__.py,sha256=mKL5cqIBDUw2WO80NRCh4s1G8KYaqgM59TTAbTkPPjQ,621
 triton/runtime/_allocation.py,sha256=zaW4B7I7c-2rkVuN7IZaUB6IQSI1t4FvnTPZH-r7DTk,798
 triton/runtime/autotuner.py,sha256=cfWBuLpL6-eBv-J2tFIbL0gE3ZGHOFSJ0e3n0GfrzLw,20244
-triton/runtime/build.py,sha256=r44tiyJFWp5Xoxb2WUyw0nqxOA2rtyWN7O8krW9gZKE,6209
+triton/runtime/build.py,sha256=hiFHKRV-fwDoT5lMbWpzA3hLp6LrE6ccw7zTo2AvHo0,6203
 triton/runtime/cache.py,sha256=uMV-CwCaS9cthIzKoLlTHXjhw_RoaIUaVP7zmgsdsIo,9689
 triton/runtime/driver.py,sha256=seGhU4efCFPVN0KVzd4gmZ1x5s0I_sFyM5NC8brXWF8,1798
 triton/runtime/errors.py,sha256=CwfJXciwel_-K3BfQfKUpLPDWrSyTnGsfJkqJojrdfQ,1052
@@ -178,9 +178,9 @@ triton/tools/mxfp.py,sha256=YQdpBrGkOVNOtnLeRjMCeVFHWkSwUubGeWsItIjO8TU,11737
 triton/tools/tensor_descriptor.py,sha256=mt4iVVRcNg0gjoytb6iCP4l5vt-H2V3MGeAQfJcStJo,1289
 triton/tools/extra/cuda/compile.c,sha256=TdIENsqk6wrvv1C4Mk-sq9keXe3SJuMQcf0UpxmjNZk,2153
 triton/tools/extra/cuda/compile.h,sha256=n9QKIFZTL4RSsiXtAxBP9XGSnxjyaevQQ9bBpwDsvAg,332
-triton_windows-3.4.0.post20.dist-info/licenses/LICENSE,sha256=kmQPuXIi_Qppj_KM4MN4LBcmI_jWxgm1V2NqgPKPuUY,1132
-triton_windows-3.4.0.post20.dist-info/METADATA,sha256=XVQISF1LWaRTh_mbQoa1OXk_1uhmigdfEfWJ2whPIPc,1794
-triton_windows-3.4.0.post20.dist-info/WHEEL,sha256=qV0EIPljj1XC_vuSatRWjn02nZIz3N1t8jsZz7HBr2U,101
-triton_windows-3.4.0.post20.dist-info/entry_points.txt,sha256=cztF9ZYXxoMhibI_OttiKCl1EBP2LQaV8naJ-BcuES4,76
-triton_windows-3.4.0.post20.dist-info/top_level.txt,sha256=WBiIZyv6n9Y7MIh-HPHSv2w1RDk7EFL__7ZgQRrmHYs,7
-triton_windows-3.4.0.post20.dist-info/RECORD,,
+triton_windows-3.4.0.post21.dist-info/licenses/LICENSE,sha256=kmQPuXIi_Qppj_KM4MN4LBcmI_jWxgm1V2NqgPKPuUY,1132
+triton_windows-3.4.0.post21.dist-info/METADATA,sha256=IHOZbw7LyV2c8dx0dvpgjzw8LmdA99209iNKB1UC01o,1794
+triton_windows-3.4.0.post21.dist-info/WHEEL,sha256=qV0EIPljj1XC_vuSatRWjn02nZIz3N1t8jsZz7HBr2U,101
+triton_windows-3.4.0.post21.dist-info/entry_points.txt,sha256=cztF9ZYXxoMhibI_OttiKCl1EBP2LQaV8naJ-BcuES4,76
+triton_windows-3.4.0.post21.dist-info/top_level.txt,sha256=WBiIZyv6n9Y7MIh-HPHSv2w1RDk7EFL__7ZgQRrmHYs,7
+triton_windows-3.4.0.post21.dist-info/RECORD,,