triton-windows 3.3.1.post19__cp313-cp313-win_amd64.whl → 3.3.1.post21__cp313-cp313-win_amd64.whl

triton/backends/amd/driver.py

@@ -383,11 +383,14 @@ static inline DevicePtrInfo getPointer(PyObject *obj, int idx) {{
     if (!PyLong_Check(ret)) {{
       PyErr_SetString(PyExc_TypeError, "data_ptr method of Pointer object must return 64-bit int");
       ptr_info.valid = false;
+      Py_DECREF(ret);
       return ptr_info;
     }}
     ptr_info.dev_ptr = (hipDeviceptr_t)PyLong_AsUnsignedLongLong(ret);
-    if(!ptr_info.dev_ptr)
+    if(!ptr_info.dev_ptr) {{
+      Py_DECREF(ret);
       return ptr_info;
+    }}
     uint64_t dev_ptr;
     hipError_t status = hipSymbolTable.hipPointerGetAttribute(&dev_ptr, HIP_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr);
     if (status == hipErrorInvalidValue) {{
@@ -433,6 +436,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
 
@@ -446,6 +450,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
   if(PyErr_Occurred()) {{

triton/backends/nvidia/compiler.py

@@ -152,7 +152,7 @@ class CUDAOptions:
     ptx_version: int = None
     enable_fp_fusion: bool = True
     launch_cooperative_grid: bool = False
-    supported_fp8_dtypes: Tuple[str] = ("fp8e5", "fp8e4b15")
+    supported_fp8_dtypes: Tuple[str] = ("fp8e4nv", "fp8e5", "fp8e4b15")
     deprecated_fp8_dtypes: Tuple[str] = ()
     default_dot_input_precision: str = "tf32"
     allowed_dot_input_precisions: Tuple[str] = ("tf32", "tf32x3", "ieee")
@@ -203,8 +203,6 @@ 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_dtypes" not in args:

triton/backends/nvidia/driver.py

@@ -372,11 +372,14 @@ static inline DevicePtrInfo getPointer(PyObject *obj, int idx) {{
     if (!PyLong_Check(ret)) {{
       PyErr_SetString(PyExc_TypeError, "data_ptr method of Pointer object must return 64-bit int");
       ptr_info.valid = false;
+      Py_DECREF(ret);
       return ptr_info;
     }}
     ptr_info.dev_ptr = PyLong_AsUnsignedLongLong(ret);
-    if(!ptr_info.dev_ptr)
+    if(!ptr_info.dev_ptr) {{
+      Py_DECREF(ret);
       return ptr_info;
+    }}
     uint64_t dev_ptr;
     int status = cuPointerGetAttribute(&dev_ptr, CU_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr);
     if (status == CUDA_ERROR_INVALID_VALUE) {{
@@ -388,7 +391,7 @@ static inline DevicePtrInfo getPointer(PyObject *obj, int idx) {{
         ptr_info.valid = false;
     }}
     ptr_info.dev_ptr = dev_ptr;
-    Py_DECREF(ret);  // Thanks ChatGPT!
+    Py_DECREF(ret);
     return ptr_info;
   }}
   PyErr_SetString(PyExc_TypeError, "Pointer argument must be either uint64 or have data_ptr method");
@@ -488,6 +491,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
   CUdeviceptr global_scratch = 0;
@@ -515,7 +519,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
-
+    Py_DECREF(ret);
   }}
 
   Py_RETURN_NONE;

triton/runtime/autotuner.py

@@ -188,9 +188,9 @@ class Autotuner(KernelInterface):
                 # prune configs
                 used_cached_result = False
                 pruned_configs = self.prune_configs(kwargs)
-                bench_start = time.time()
+                bench_start = time.perf_counter()
                 timings = {config: self._bench(*args, config=config, **kwargs) for config in pruned_configs}
-                bench_end = time.time()
+                bench_end = time.perf_counter()
                 self.bench_time = bench_end - bench_start
                 self.cache[key] = builtins.min(timings, key=timings.get)
                 full_nargs = {**self.nargs, **kwargs, **self.cache[key].all_kwargs()}

triton/runtime/build.py

@@ -80,17 +80,17 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries):
         scheme = 'posix_prefix'
     py_include_dir = sysconfig.get_paths(scheme=scheme)["include"]
     custom_backend_dirs = set(os.getenv(var) for var in ('TRITON_CUDACRT_PATH', 'TRITON_CUDART_PATH'))
+    # Don't append in place
     include_dirs = include_dirs + [srcdir, py_include_dir, *custom_backend_dirs]
     if os.name == "nt":
-        library_dirs += find_python()
+        library_dirs = library_dirs + find_python()
     # Link against Python stable ABI
-    # libraries is modified in place
     if "python3" not in libraries:
-        libraries += ["python3"]
+        libraries = libraries + ["python3"]
     if is_msvc(cc):
         _, msvc_winsdk_inc_dirs, msvc_winsdk_lib_dirs = find_msvc_winsdk()
-        include_dirs += msvc_winsdk_inc_dirs
-        library_dirs += msvc_winsdk_lib_dirs
+        include_dirs = include_dirs + msvc_winsdk_inc_dirs
+        library_dirs = library_dirs + msvc_winsdk_lib_dirs
     cc_cmd = _cc_cmd(cc, src, so, include_dirs, library_dirs, libraries)
 
     try:

triton/windows_utils.py

@@ -204,12 +204,19 @@ def find_winsdk_env() -> tuple[Optional[Path], Optional[str]]:
     winsdk_base_path = Path(winsdk_base_path)
 
     version = os.getenv("WindowsSDKVersion")
-    if version:
-        version = version.rstrip("\\")
+    if version is None:
+        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, "
+            f"Environment variables WindowsSdkDir = {winsdk_base_path}, "
+            f"WindowsSDKVersion (or WindowsSDKVer) = {version} are set, "
             "but this Windows SDK installation is incomplete."
         )
         return None, None

{triton_windows-3.3.1.post19.dist-info → triton_windows-3.3.1.post21.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: triton-windows
-Version: 3.3.1.post19
+Version: 3.3.1.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-3.3.1.post19.dist-info → triton_windows-3.3.1.post21.dist-info}/RECORD

@@ -3,118 +3,21 @@ triton/_internal_testing.py,sha256=OBY28huiEWItqGgiukgZzHLLaSbS8yj9kdhn_u562Yg,5
 triton/_utils.py,sha256=5RiCLwW14w0Q3mdZ-9yz-VO5KiSexNj9xeDt4gaNsvE,1014
 triton/errors.py,sha256=8WfnuRKLG578mgY6cBA3ECruVMf9ULEKFNgRcJ6IhWM,89
 triton/testing.py,sha256=ivFf1Fq9frmfVahaVUp0bgJxmvVZNACZfj3Sai6zfAs,20048
-triton/windows_utils.py,sha256=YUl-1QbLINQRaAAMNYPjLiTFZlVsCNi9mTFck5aemwk,12778
-triton/_C/libtriton.pyd,sha256=-4eS6fQVllONYof2Sq9gIj1DYPV2NJ7UhDcjvbG7e2Y,87287808
+triton/windows_utils.py,sha256=KOxhtcaTpNbs-UNtZvBJ4lCahJeMgbXUzWdv2wW3z98,13036
+triton/_C/libtriton.pyd,sha256=-9nmO4enVytCA6tMWQGbKo3UbdnkjJ0xZJhJXClm8-o,87410688
 triton/backends/__init__.py,sha256=opAo_vgEMt3tLO_bYFrYGksnIu0qohbmyuu_s3-rNAs,1595
 triton/backends/compiler.py,sha256=ymaG0kpveAuESbQ9QZ0RyXjr0Aq4el_G5XGYogJ2gNA,3588
 triton/backends/driver.py,sha256=AN60upJlPgia0JwvZ8vIVgLITNPuI0fdz8zMIIHPpF4,1450
 triton/backends/amd/compiler.py,sha256=27jurEV7tH5J6BHtOXrdPJyCMYIPHo5G4Op_O66gv4E,19135
 triton/backends/amd/driver.c,sha256=obiiiPndny5NyhUcJ8iyrVHrXU1ruLpLGd_LgaKQEbU,8459
-triton/backends/amd/driver.py,sha256=p8FcTiAq-829p2gRQZ5sPE1_d1SQQ_sOhb5WxI6rXME,20379
-triton/backends/amd/include/hip/channel_descriptor.h,sha256=gTYe7SzIg-m3ThOQY2vr5Rh6-uWvUP_d37v8F4T2Q14,1773
-triton/backends/amd/include/hip/device_functions.h,sha256=vkybrdk6wyZP-T1I5PRjtfcMqGYXDeBpB5jhYj358GU,1589
-triton/backends/amd/include/hip/driver_types.h,sha256=m1HI80HC80qkTeco2Jd07woL_jTy48lz9JiDCV_8zsg,18985
-triton/backends/amd/include/hip/hip_bf16.h,sha256=lLw6K5ltb6AqSuINYTq8flxxsDkBP8Y2zbqmUjBcG9c,1571
-triton/backends/amd/include/hip/hip_bfloat16.h,sha256=Nqoy9VjfjglVx2_NJcp8hyT1sJUukXRWj8XMlidv1yA,1755
-triton/backends/amd/include/hip/hip_common.h,sha256=q5aPhG3DHW0iUJ7ayS5lfM_ZnZQNbMmLmfdHlOwbPdA,3450
-triton/backends/amd/include/hip/hip_complex.h,sha256=TmdzQP5oVPfhBVARJYcR5eyv9HInmKMFuFoQ_1ECk_I,1594
-triton/backends/amd/include/hip/hip_cooperative_groups.h,sha256=gMLvaYQ3b-f1vcoMtEwtkN0hO5__zNfP5p5oBKmv_SE,1878
-triton/backends/amd/include/hip/hip_deprecated.h,sha256=gFLuCuKn7R_xCfum_i_Q-vi3Lg8NWHKphKZKze8DwEo,6340
-triton/backends/amd/include/hip/hip_ext.h,sha256=mlzOesF-X62g9AnWdA4MP99GRu_VtwMbffOJFloLtRc,8609
-triton/backends/amd/include/hip/hip_fp16.h,sha256=vKJh-zgDWUW7NyXxtv2ho6aVLXX8BIPfzCigEQ5d6I4,1523
-triton/backends/amd/include/hip/hip_fp8.h,sha256=C4qn0im7Uhvp226VmL_QbF2b45Hhss2eokdSbqWkBcs,1433
-triton/backends/amd/include/hip/hip_gl_interop.h,sha256=-GwkSFMBneM8akFE7pqlhi0k-Ft2uz5674wGoiaU43Q,1438
-triton/backends/amd/include/hip/hip_hcc.h,sha256=RYrArDlnTEP89xKbzIpW17_bsBY5moCitq00PL-4oWI,1307
-triton/backends/amd/include/hip/hip_math_constants.h,sha256=8bSfve5E7cDuvNAUkFUeQwSLg3iJJHuqhuD4FmHNxEM,1588
-triton/backends/amd/include/hip/hip_profile.h,sha256=sjsNuduu5Jd6s7sJndZvZLlE0RZ0wN1rTVwv5nR7If0,1304
-triton/backends/amd/include/hip/hip_runtime.h,sha256=uy90l8Nep6xNUzeGcHMoDv84BT3hMpieTV-5ijkpL5A,3058
-triton/backends/amd/include/hip/hip_runtime_api.h,sha256=H9vPaMQ7VBkEtkxI5RyJ9UOXCmKzbPF70pcAsYNbP_A,402475
-triton/backends/amd/include/hip/hip_texture_types.h,sha256=AhkvjG4cDjf_ZFLg5SsSTfBnXG614PBK1XVPa7irZbk,1237
-triton/backends/amd/include/hip/hip_vector_types.h,sha256=6FcBMBkP3ZN1Enalpa9hV0VopxdBJvbUCuaxISgzbTY,1630
-triton/backends/amd/include/hip/hip_version.h,sha256=3MuxlZXlLLrhLllRr-1yrIZyv7l_zknuZrT8bDDtm3g,407
-triton/backends/amd/include/hip/hiprtc.h,sha256=dBxesFd3NCvG-NQ99s61i4fUQXfoHNmYhdF6CtE9QHQ,15904
-triton/backends/amd/include/hip/library_types.h,sha256=tPOJTQedPH5qC9meawLgKpnbFrQC2WKlfo6s0rhKoZc,2370
-triton/backends/amd/include/hip/math_functions.h,sha256=frzdJ4veBG8n9ALO4EmRrdOiDguR6FP6ygLnvOnVVSM,1815
-triton/backends/amd/include/hip/surface_types.h,sha256=uQHjITphDM7k4pnuEoDEupMUxBobzvhJpSy0unpegh4,1959
-triton/backends/amd/include/hip/texture_types.h,sha256=CtmdykZfDikhnrVfdJk3w2VK5X3Af_6rEKzU-VgLu24,6687
-triton/backends/amd/include/hip/amd_detail/amd_channel_descriptor.h,sha256=_2myGIdBTE0plFbGKOSx8HUqGZd0UBHo-YvKe2xkpbU,11708
-triton/backends/amd/include/hip/amd_detail/amd_device_functions.h,sha256=1TwUYh4RJhoiz850vLNicO1fFxFHR-vC-16ttDGDGGw,31367
-triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h,sha256=KfiNN5FjV6mqSAOam6VJDu9tI7l8uaRUrQ5EMiP5Ymw,50830
-triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h,sha256=AWsz68vSfi9PnCsIIMOQkCStRLbEat7f6om9_J2Ahj4,61352
-triton/backends/amd/include/hip/amd_detail/amd_hip_bfloat16.h,sha256=cFJlQEELGau_9geACeuiiFHyuAWCD6-VuSqcTnqajX0,9484
-triton/backends/amd/include/hip/amd_detail/amd_hip_common.h,sha256=dzkuIzuklqTRaNJjKLqfFEm6Fh4tK_FkTjYHFsZkmCI,1370
-triton/backends/amd/include/hip/amd_detail/amd_hip_complex.h,sha256=SEygl8X_MCXDVXxNIBm5Ds0eWwa-ojVXUUW48SIgsX8,5855
-triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h,sha256=t8rqzpZS--LGLI1D-pNBNCp7TCKGS0H136ps6m6VOlI,31861
-triton/backends/amd/include/hip/amd_detail/amd_hip_fp16.h,sha256=86Nw97iaiC4QV5xBv8d3Bwc4FioMh5DQuCHj3sh_Yrw,57854
-triton/backends/amd/include/hip/amd_detail/amd_hip_fp8.h,sha256=wc1l471D-BBroLMoV1uEsPXTp-zQjSpq-jF8bqeNZX8,53765
-triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h,sha256=djlpeoEOqtX0gVVRgqqakQZkwNzLudnK_ixbShizU7M,3861
-triton/backends/amd/include/hip/amd_detail/amd_hip_math_constants.h,sha256=u1fIaf-AiWF70ZA1zxVkUIbRqoJLu5lrfYbgt_usySk,5890
-triton/backends/amd/include/hip/amd_detail/amd_hip_runtime.h,sha256=ZvDsQ0AiZnJ178NuAsA7AuHrySXbN3aFs5Z9m2tsIDg,13954
-triton/backends/amd/include/hip/amd_detail/amd_hip_runtime_pt_api.h,sha256=fc4mtHBkWmiSRh8m-dxIxvu9zsweLTwEgohkntYcgJw,9997
-triton/backends/amd/include/hip/amd_detail/amd_hip_unsafe_atomics.h,sha256=w9nJ1S32GRl_ejDiGacteM6Zf84iovIifAzWX8Bze0Q,24202
-triton/backends/amd/include/hip/amd_detail/amd_hip_vector_types.h,sha256=qPdmRJnzlgtjVshkafoHxdHoMLkoYS9U-ZD-TjLznr0,57088
-triton/backends/amd/include/hip/amd_detail/amd_math_functions.h,sha256=46wiaEMStCczEsHtccgHlATfw_0O5j6Z8rlFkC7bmUA,3171
-triton/backends/amd/include/hip/amd_detail/amd_surface_functions.h,sha256=rsQuylNqmNhLb7PZjBz7WbruD_6YIXtOptY2BNJDxVU,11062
-triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h,sha256=-uJy_hBwSxRA9gzGp4UEZ3co8D_UHVNoMr_Rvx19qik,20042
-triton/backends/amd/include/hip/amd_detail/amd_warp_sync_functions.h,sha256=2B25tjpJ_KJGijRBaEy_a2_4HLt_TXQ6eJ57zRNoPOw,11706
-triton/backends/amd/include/hip/amd_detail/concepts.hpp,sha256=7EOkpr2w2-jclUQ115yxtFCkBWJ7btUzhBOe-mR0N0M,1252
-triton/backends/amd/include/hip/amd_detail/device_library_decls.h,sha256=4clSpgf898UVjfZFVnDkcYi75A27crPsuFtLcs1s4KU,7457
-triton/backends/amd/include/hip/amd_detail/functional_grid_launch.hpp,sha256=u7hRB9kQXX575a5C7cV3gKow55DSBUCwO0dTjIswlag,8129
-triton/backends/amd/include/hip/amd_detail/grid_launch.h,sha256=tNS7CQw9gy-z930CElH3n6c5iMvpsQ_WFZK024mNzEo,1830
-triton/backends/amd/include/hip/amd_detail/grid_launch.hpp,sha256=EuAlM3olyrArebqwW5eSxo4gfjvWCGOAGAuLLmFttgw,1370
-triton/backends/amd/include/hip/amd_detail/grid_launch_GGL.hpp,sha256=KpQAuyy1Dyt45WcPaR_x-Ex-onPGEHA01DBbla7TT-k,1219
-triton/backends/amd/include/hip/amd_detail/helpers.hpp,sha256=hi2pW1mXQnbIwvmwWt_nG6A38sqLOd-QP5S9sETTs60,5707
-triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp,sha256=UMsCaqurHFSRNiHpwjaA8DExAldhh4y0UP_KOxLa25Y,101025
-triton/backends/amd/include/hip/amd_detail/hip_assert.h,sha256=fNsG23KISuY-k5JFoX-5hZ7qGQScisXuHcdEwYlXOqw,3978
-triton/backends/amd/include/hip/amd_detail/hip_cooperative_groups_helper.h,sha256=tQ_XIvGKhvrj1h7gY-IVLmKvIPhsQa0YsBflxdhUHP8,7957
-triton/backends/amd/include/hip/amd_detail/hip_fp16_gcc.h,sha256=BtFsKmTptN4TOHocEicfNbBl2JCdZWKm_bd5mc5OzYY,6660
-triton/backends/amd/include/hip/amd_detail/hip_fp16_math_fwd.h,sha256=63tKWMPdW56qWlH_HbCaF_isVXufm514ol_SxL4YjTQ,5134
-triton/backends/amd/include/hip/amd_detail/hip_ldg.h,sha256=KAEZb9H4z4DDrkaloMOeWzahiDfI2V6c68vWT3jb5fU,3652
-triton/backends/amd/include/hip/amd_detail/hip_prof_str.h,sha256=Xb0M9ztYHRqRLjKFBumbAR2rjbUnoZulweCZKTU8ejU,648748
-triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h,sha256=eZUI5cscYKb9xC9IqnaY53Jgy1dQjkERwQaw034Bt5g,2758
-triton/backends/amd/include/hip/amd_detail/host_defines.h,sha256=ZyCvaPsp4vcFiEEIt2BhapfYVvWoVI_Q2P9rZHlFmCY,7197
-triton/backends/amd/include/hip/amd_detail/hsa_helpers.hpp,sha256=Os-sJQOFI_0Abh8Ql05s0Rtfruk4NsSMfg7BtugxMgg,3232
-triton/backends/amd/include/hip/amd_detail/macro_based_grid_launch.hpp,sha256=6ocsArNa9_R6D6XCuNy8Zq23KG-j2uYsjqNCtnMrJws,67925
-triton/backends/amd/include/hip/amd_detail/math_fwd.h,sha256=nup5YhceJnngoLJCESI8qX08dNpbZci0i78WKu-wfdI,17000
-triton/backends/amd/include/hip/amd_detail/ockl_image.h,sha256=LzRPGMb515_iIAIIcbb2uQB-bTvT4xOjY51VdARD7lc,10538
-triton/backends/amd/include/hip/amd_detail/program_state.hpp,sha256=8QE9OmB8OKTy7rBr3EYEizJI2s-_1tgXpgU7zCA2Ky0,3154
-triton/backends/amd/include/hip/amd_detail/texture_fetch_functions.h,sha256=Ex1lF2gBWJxtC3yP9pXRSFywMp3gbEmyl0Sw8iL91yM,17787
-triton/backends/amd/include/hip/amd_detail/texture_indirect_functions.h,sha256=KkW5o5gMpoVMTRwzfXHA7-kZ9ynI8OaIw6jJ1EB1s98,18447
-triton/backends/amd/include/hsa/Brig.h,sha256=5H-btCHq40qgjjpwVAoRWf3E0ccf-J6UCPEcKx_hGKw,32705
-triton/backends/amd/include/hsa/amd_hsa_common.h,sha256=q_zN0eq-dwR7FnQ84PcpV3yZyvjHsouIAjJgKltGoX8,3912
-triton/backends/amd/include/hsa/amd_hsa_elf.h,sha256=lBT57uuPT6ra01HQtM1O3RhVKW3n41AjESDSi6KVN7w,17306
-triton/backends/amd/include/hsa/amd_hsa_kernel_code.h,sha256=C55F8a480QsW16-iwN9TIT3cKnGh6GoeoEaEv3aVh4g,12659
-triton/backends/amd/include/hsa/amd_hsa_queue.h,sha256=ZJ-k5wY30heLmQnGB0VUz36XCiVHRmspg5FRNMGIk_U,4766
-triton/backends/amd/include/hsa/amd_hsa_signal.h,sha256=FDegZnWQC04GtnqHjXOBsB-AoVSaqdhNY6Mwbua5FGA,2947
-triton/backends/amd/include/hsa/hsa.h,sha256=KyOkPu9bhpsaDCJFqlnis9oPDAdN6l2ujQSepL86jjs,191193
-triton/backends/amd/include/hsa/hsa_amd_tool.h,sha256=pyZSyIVl-UA5AOhte78jvn4V3hCd0dxJAIv7KeADsPs,2843
-triton/backends/amd/include/hsa/hsa_api_trace.h,sha256=GRBMFBWlv2dYJIT8BAWa1y_q-veLIgEVa1uIm-eM-Yw,29602
-triton/backends/amd/include/hsa/hsa_api_trace_version.h,sha256=yU2aATDJyhnCZF-FmgeF2wnp-EvPLE7n5zSuw8CNAqc,3248
-triton/backends/amd/include/hsa/hsa_ext_amd.h,sha256=6XS5at7KdbW8-dzNVZwJ_1bM5OWiojdbN668pz9srbc,117980
-triton/backends/amd/include/hsa/hsa_ext_finalize.h,sha256=sv0AZbDM-B1wIdQ3cHTMlpUtNacQN2PkOgX90IZol_o,20227
-triton/backends/amd/include/hsa/hsa_ext_image.h,sha256=t5YJm_aw9EePCeFL1hoIfQ8ubIjBte-ptfReq6Ts-8Y,54232
-triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h,sha256=IR_vo3kSqcTs-gnRdpXW3JE1RHQ71FBi4S81G8GtpZ4,19823
-triton/backends/amd/include/hsa/hsa_ven_amd_loader.h,sha256=c6cxPAzAox7u6IbFzEkQZfCuRl-Kr39WhY2_w23X1R4,26146
-triton/backends/amd/include/hsa/hsa_ven_amd_pc_sampling.h,sha256=1jQvi96s94-GjOGktmvpqjVRsTGANugOsV6XmgMfivk,18767
-triton/backends/amd/include/roctracer/hip_ostream_ops.h,sha256=Si6RFHPn0QtMpSBboKFyfkrcOMYhTiom5oUnDEqRnP0,188184
-triton/backends/amd/include/roctracer/hsa_ostream_ops.h,sha256=jSHN5Ltd_EJtecE59xJS4wL_CuLuNkbO3QMqDgMpunk,69906
-triton/backends/amd/include/roctracer/hsa_prof_str.h,sha256=f0pO8T0JBnG_stZ_Bgawx-5_B2BNS7s9ogvEldA0UFo,124187
-triton/backends/amd/include/roctracer/roctracer.h,sha256=B8sHz2DMNprP7EqNWIGwVLY1KQMpxmhfVy4UoR8dzzY,23849
-triton/backends/amd/include/roctracer/roctracer_ext.h,sha256=vLaZ8peAxSy0cwrdEalKnUApkKspfa04iw1Mr_Zcio0,2940
-triton/backends/amd/include/roctracer/roctracer_hcc.h,sha256=NlF3R8JQ9oX9lGpm0b2n-EWJ0r3y9sP9wbwnoucaCuY,1303
-triton/backends/amd/include/roctracer/roctracer_hip.h,sha256=RCzYuNw1vLR7xK4rb06TtM9TU546UYKHJ83IMHmZEm8,1432
-triton/backends/amd/include/roctracer/roctracer_hsa.h,sha256=M8APM64XNAWSslxQisM-pcmKoUQaUdTMaKvSACyt0Ag,4108
-triton/backends/amd/include/roctracer/roctracer_plugin.h,sha256=8GGE1zDbdPCVJtbmwOCYq7X0mwFjfWRtzDYKLD4cKys,4786
-triton/backends/amd/include/roctracer/roctracer_roctx.h,sha256=gBjBk5vb0l3PbBSQ7V9iFtaM_RzkIDJEW1A_PXBihBM,2014
-triton/backends/amd/include/roctracer/roctx.h,sha256=RhJXUXRhSJ5LRE_1gm7E6-bjEMrfcFBLDLuf3UxAIh8,6717
-triton/backends/amd/include/roctracer/ext/prof_protocol.h,sha256=6FAcvVD-dNM7uulFs2B-aTxw5xOAWGy6evdD4yUaebA,3849
+triton/backends/amd/driver.py,sha256=y7w0jkoQ-jcpWRNDc9i6ZO-gQUOV8detjmAonqBr-zo,20473
 triton/backends/amd/lib/asanrtl.bc,sha256=1xv2RlU3WvbdsghHlmhwiHewGM2B5dKts5bERM6S89o,24508
 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=4BnLUapsoh8lsQPYlm-_ZCyKvU3DgJY3NgDMo8leV9c,19996
+triton/backends/nvidia/compiler.py,sha256=IKcoPBln5j40P3heHjpPVczZ4XL2PpPkRGiAHTtoGM8,19922
 triton/backends/nvidia/driver.c,sha256=HqtiJDPRxVav9pHl_swBja28RasjHZ7oFOpWKMT627c,18912
-triton/backends/nvidia/driver.py,sha256=kMMVl4SMbJc6obI2Brxoj7RNJt6YBGKYSB56dPeUY6w,21037
+triton/backends/nvidia/driver.py,sha256=9BOCsdHXFgGcKiUMicPu988ay0Ixo68Sa5SAyYtr-Eo,21110
 triton/backends/nvidia/bin/ptxas.exe,sha256=iCva9hIYg-Q2NybchwaQJFkwDzNiliFOTDdZrHPLb6A,24732160
 triton/backends/nvidia/include/cuda.h,sha256=Fn44OjeRImxegJ39apYUspseEfTWNGwpqSGUOnHj5WY,1183268
 triton/backends/nvidia/lib/libdevice.10.bc,sha256=XC-uN8huaMOjhgWpX1EtfRLV89uYYxC-R_VzBKpype4,473728
@@ -141,8 +44,8 @@ triton/language/extra/hip/__init__.py,sha256=ieSER4LeX9_0horChGUUVwpuKAprkuka8uG
 triton/language/extra/hip/libdevice.py,sha256=EVraUfeXzQmN3F5Lleg2mohVcbFWOWlLaAH1nkbqtV4,16841
 triton/runtime/__init__.py,sha256=mKL5cqIBDUw2WO80NRCh4s1G8KYaqgM59TTAbTkPPjQ,621
 triton/runtime/_allocation.py,sha256=zaW4B7I7c-2rkVuN7IZaUB6IQSI1t4FvnTPZH-r7DTk,798
-triton/runtime/autotuner.py,sha256=0ku0wjPo8xOvom6P4uEVZHsgPkxAFOqP1LjUVue0HLM,17854
-triton/runtime/build.py,sha256=6akPf7jgSRDWqcDGvJXxxiO0MzH6503YrM7xmr24l-Y,3825
+triton/runtime/autotuner.py,sha256=ZkdEh1hFXOA9OJvaMgBzVWTNwwYcnNYrniEN1phYd24,17870
+triton/runtime/build.py,sha256=Kp-mqYUZV5YsiOlqA8f7cAaT5dtcb71l5Xg3-xvK4Cg,3869
 triton/runtime/cache.py,sha256=uoU1UH3HPxkxT0r-69HIZgLBls3T-TDxJedRULX9lbM,10583
 triton/runtime/driver.py,sha256=VZ-883Xri71R72lHB6usIpLo3gGLbZJkAlLP3ewWSpc,1509
 triton/runtime/errors.py,sha256=CwfJXciwel_-K3BfQfKUpLPDWrSyTnGsfJkqJojrdfQ,1052
@@ -254,7 +157,7 @@ triton/tools/link.py,sha256=u7qtfZRLriZkAMEGNvj8YF-k1cthmLL7BwHYqBgT63E,11871
 triton/tools/mxfp.py,sha256=YQdpBrGkOVNOtnLeRjMCeVFHWkSwUubGeWsItIjO8TU,11737
 triton/tools/extra/cuda/compile.c,sha256=TdIENsqk6wrvv1C4Mk-sq9keXe3SJuMQcf0UpxmjNZk,2153
 triton/tools/extra/cuda/compile.h,sha256=n9QKIFZTL4RSsiXtAxBP9XGSnxjyaevQQ9bBpwDsvAg,332
-triton_windows-3.3.1.post19.dist-info/METADATA,sha256=LkTgBQiKqh82quqnNxm1RdVG2qYoJ1uvC74OaClT6-0,1627
-triton_windows-3.3.1.post19.dist-info/WHEEL,sha256=qV0EIPljj1XC_vuSatRWjn02nZIz3N1t8jsZz7HBr2U,101
-triton_windows-3.3.1.post19.dist-info/top_level.txt,sha256=KhMzHYsArnZ3IkjAQ-xLnx1n_FjvEpJNelg2xPiDl-U,254
-triton_windows-3.3.1.post19.dist-info/RECORD,,
+triton_windows-3.3.1.post21.dist-info/METADATA,sha256=kau-qMUZQ7G7g7qy5MvcD4KdIf8u0RngjQyogPyQaGY,1627
+triton_windows-3.3.1.post21.dist-info/WHEEL,sha256=qV0EIPljj1XC_vuSatRWjn02nZIz3N1t8jsZz7HBr2U,101
+triton_windows-3.3.1.post21.dist-info/top_level.txt,sha256=KhMzHYsArnZ3IkjAQ-xLnx1n_FjvEpJNelg2xPiDl-U,254
+triton_windows-3.3.1.post21.dist-info/RECORD,,