triton-windows 3.2.0.post13__cp311-cp311-win_amd64.whl → 3.2.0.post16__cp311-cp311-win_amd64.whl

triton/backends/nvidia/compiler.py

@@ -358,11 +358,10 @@ class CUDABackend(BaseBackend):
     @staticmethod
     def make_cubin(src, metadata, opt, capability):
         ptxas, _ = _path_to_binary("ptxas")
-        # On Windows, we need to set delete=False, close the temp file before reading it, and manually remove it
         with tempfile.NamedTemporaryFile(delete=False, mode='w', suffix='.ptx') as fsrc, \
             tempfile.NamedTemporaryFile(delete=False, mode='r', suffix='.log') as flog:
             fsrc.write(src)
-            fsrc.close()
+            fsrc.flush()
             fbin = fsrc.name + '.o'
 
             line_info = [] if os.environ.get('TRITON_DISABLE_LINE_INFO') else ['-lineinfo']
@@ -373,15 +372,10 @@ class CUDABackend(BaseBackend):
                 ptxas, *line_info, *fmad, '-v', *opt_level, f'--gpu-name=sm_{capability}{suffix}', fsrc.name, '-o', fbin
             ]
             try:
-                subprocess.run(ptxas_cmd, check=True, close_fds=False, stdout=flog, stderr=flog)
-                try_remove(fsrc.name)
-                flog.close()
-                try_remove(flog.name)
+                subprocess.run(ptxas_cmd, check=True, close_fds=True, stdout=flog, stderr=flog)
             except subprocess.CalledProcessError as e:
-                flog.close()
                 with open(flog.name) as log_file:
                     log = log_file.read()
-                try_remove(flog.name)
 
                 if e.returncode == 255:
                     error = 'Internal Triton PTX codegen error'
@@ -394,9 +388,13 @@ class CUDABackend(BaseBackend):
                                    f'`ptxas` stderr:\n{log}\n'
                                    f'Repro command: {" ".join(ptxas_cmd)}\n')
 
-            with open(fbin, 'rb') as f:
-                cubin = f.read()
-            try_remove(fbin)
+        with open(fbin, 'rb') as f:
+            cubin = f.read()
+        try_remove(fbin)
+
+        # It's better to remove the temp files outside the context managers
+        try_remove(fsrc.name)
+        try_remove(flog.name)
         return cubin
 
     def add_stages(self, stages, options):

triton/backends/nvidia/driver.c

@@ -9,6 +9,7 @@
 
 #include <stdbool.h>
 #define PY_SSIZE_T_CLEAN
+#define Py_LIMITED_API 0x03090000
 #include <Python.h>
 // #include <stdatomic.h>
 

triton/backends/nvidia/driver.py

@@ -179,6 +179,8 @@ def make_launcher(constants, signature, ids):
     src = f"""
 #include \"cuda.h\"
 #include <stdbool.h>
+#define PY_SSIZE_T_CLEAN
+#define Py_LIMITED_API 0x03090000
 #include <Python.h>
 
 #ifndef _WIN32

{triton_windows-3.2.0.post13.dist-info → triton_windows-3.2.0.post16.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: triton-windows
-Version: 3.2.0.post13
+Version: 3.2.0.post16
 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-3.2.0.post13.dist-info → triton_windows-3.2.0.post16.dist-info}/RECORD

@@ -3,7 +3,7 @@ triton/_internal_testing.py,sha256=4pzyc_36u05khUveZ1TWL3MQ-7EVMJ1C2c1dRd8BMsw,4
 triton/errors.py,sha256=8WfnuRKLG578mgY6cBA3ECruVMf9ULEKFNgRcJ6IhWM,89
 triton/testing.py,sha256=fX3pn9bjC3Z-z5qzSKW56C_2WF8h3mHLy5RJqpZ-HsA,19382
 triton/windows_utils.py,sha256=aQMItmuZNXaki8zSB7HMvhy1RsXlmz1GcO-lORm1IIk,10852
-triton/_C/libtriton.pyd,sha256=kaDz-YoCvQQPWFekXkNoDVI_-9dGZMbGerPS8MneZG0,87245312
+triton/_C/libtriton.pyd,sha256=RtgL6yjGrRjSbfEtGtq3tV6yB9_72p5jnGLUxonJX3U,87245312
 triton/backends/__init__.py,sha256=opAo_vgEMt3tLO_bYFrYGksnIu0qohbmyuu_s3-rNAs,1595
 triton/backends/compiler.py,sha256=JZiiEbB9Wws3tjU6KXrydKtlOQI7Suk-mTYPlafa0Qk,11388
 triton/backends/driver.py,sha256=QX_6P1Go9ajdlHZi4Hv3nCtdHyDA6o8_lM3NMnlH1mk,1386
@@ -105,9 +105,9 @@ triton/backends/amd/include/roctracer/ext/prof_protocol.h,sha256=6FAcvVD-dNM7uul
 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=n5uOEkwELHK5cVs5N9f3jLpgie-1dul3tCHSCl_PHeA,16581
-triton/backends/nvidia/driver.c,sha256=sg-C-zQEW2B9BkTQJ3xKfnRf1rOcC8_eHoQCTuZr_Ow,18872
-triton/backends/nvidia/driver.py,sha256=pfB_lZgnN5RlZOod-N9ZVh1eIHu4lHeF3K2xldU_QZQ,17534
+triton/backends/nvidia/compiler.py,sha256=7-WtsBYdNn_RIHj6-IhkkDwh9fJvqkfjarj3Q757Z3s,16420
+triton/backends/nvidia/driver.c,sha256=8m1JB-t52GN1LoAFNlpLmM9DHuBxNIHL2VHFLTyJp8g,18906
+triton/backends/nvidia/driver.py,sha256=daDEQB2ExjTZ2F0-k-iv5DZBRSvKlVnt-IuvsDZtZHs,17593
 triton/backends/nvidia/bin/ptxas.exe,sha256=2m-Z7N6rSw_zUngI0Zza4eMzhqglEEQQ6sMkev8cAr8,19580416
 triton/backends/nvidia/include/cuda.h,sha256=6SN2FpBhqz9-wdKnWN4_R3q18-auO67gdcBp-cEiSpE,1072817
 triton/backends/nvidia/lib/libdevice.10.bc,sha256=XC-uN8huaMOjhgWpX1EtfRLV89uYYxC-R_VzBKpype4,473728
@@ -238,7 +238,7 @@ triton/tools/compile.py,sha256=b3yNnVgoBk8WzOs87JrZPDIyasdSgAslOWmxse1J6yM,6761
 triton/tools/disasm.py,sha256=BBO4bALdLcWgWDLhQdYHLlTx3oo8g_d8maeE_Uu-FmU,5088
 triton/tools/experimental_descriptor.py,sha256=0Wqy96Cc6YLh9o0eTknW-Lfvha6lfRSfe8bswkcPHMs,1260
 triton/tools/link.py,sha256=u7qtfZRLriZkAMEGNvj8YF-k1cthmLL7BwHYqBgT63E,11871
-triton_windows-3.2.0.post13.dist-info/METADATA,sha256=uAvTirqnN40FaRwsUR3KrEBNUtDiNqa8fUfcMTpV6Mc,1514
-triton_windows-3.2.0.post13.dist-info/WHEEL,sha256=KJXh1V8Kwbv4sLWZhe-NnXz7pUnJH10x-MhKeMsQYvA,101
-triton_windows-3.2.0.post13.dist-info/top_level.txt,sha256=iIyUoyO6Ld8jQHWY9plUyco-emBXKTqVmuAXJAGvmEk,211
-triton_windows-3.2.0.post13.dist-info/RECORD,,
+triton_windows-3.2.0.post16.dist-info/METADATA,sha256=zxegA-uRJZoybHNk_X6ScqsXXzaeU5-4iv9mRfshjS0,1514
+triton_windows-3.2.0.post16.dist-info/WHEEL,sha256=kynURWPI8D6MJZUv3wOGtm1TQb0_WLqkF127RFCfG9M,101
+triton_windows-3.2.0.post16.dist-info/top_level.txt,sha256=iIyUoyO6Ld8jQHWY9plUyco-emBXKTqVmuAXJAGvmEk,211
+triton_windows-3.2.0.post16.dist-info/RECORD,,

{triton_windows-3.2.0.post13.dist-info → triton_windows-3.2.0.post16.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (76.0.0)
+Generator: setuptools (77.0.1)
 Root-Is-Purelib: false
 Tag: cp311-cp311-win_amd64