triton-windows 3.2.0.post19__cp312-cp312-win_amd64.whl → 3.2.0.post21__cp312-cp312-win_amd64.whl

triton/backends/amd/driver.py

@@ -362,11 +362,14 @@ 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) {{
@@ -410,6 +413,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
 
@@ -423,6 +427,7 @@ 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

@@ -122,7 +122,7 @@ class CUDAOptions:
     cluster_dims: tuple = (1, 1, 1)
     ptx_version: int = None
     enable_fp_fusion: bool = True
-    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")
@@ -164,8 +164,6 @@ class CUDABackend(BaseBackend):
         args = {k: opts[k] for k in CUDAOptions.__dataclass_fields__.keys() if k in opts}
         if "supported_fp8_dtypes" not in args:
             supported_fp8_dtypes = set(CUDAOptions.supported_fp8_dtypes)
-            if self.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

@@ -314,11 +314,14 @@ 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) {{
@@ -330,7 +333,7 @@ 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");
@@ -427,6 +430,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
   // raise exception asap
@@ -445,7 +449,7 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     Py_DECREF(args);
     if (!ret)
       return NULL;
-
+    Py_DECREF(ret);
   }}
 
   // return None

triton/runtime/autotuner.py

@@ -182,9 +182,9 @@ 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

@@ -94,17 +94,17 @@ 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: