Revert cleaning cache changes on Windows (#3217)

With unloaded DLL libraries, these changes are no longer necessary.
However, two tests that hold a reference to the compiled kernel need to
be adjusted - manually clear the cache (inside `JITFunction` object).

CI:
*
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12951798922
(passed)
*
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12955820922
(check status)

Blocked on #3251

Extra refs:
* python/cpython#87319

---------

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

python/test/regression/conftest.py

@@ -14,13 +14,9 @@ def device(request):
 
 @pytest.fixture
 def fresh_triton_cache():
-    try:
-        with tempfile.TemporaryDirectory() as tmpdir:
-            try:
-                os.environ["TRITON_CACHE_DIR"] = tmpdir
-                yield tmpdir
-            finally:
-                os.environ.pop("TRITON_CACHE_DIR", None)
-    except OSError:
-        # Ignore errors, such as PermissionError, on Windows
-        pass
+    with tempfile.TemporaryDirectory() as tmpdir:
+        try:
+            os.environ["TRITON_CACHE_DIR"] = tmpdir
+            yield tmpdir
+        finally:
+            os.environ.pop("TRITON_CACHE_DIR", None)

python/test/unit/conftest.py

@@ -17,16 +17,12 @@ def device(request):
 
 @pytest.fixture
 def fresh_triton_cache():
-    try:
-        with tempfile.TemporaryDirectory() as tmpdir:
-            try:
-                os.environ["TRITON_CACHE_DIR"] = tmpdir
-                yield tmpdir
-            finally:
-                os.environ.pop("TRITON_CACHE_DIR", None)
-    except OSError:
-        # Ignore errors, such as PermissionError, on Windows
-        pass
+    with tempfile.TemporaryDirectory() as tmpdir:
+        try:
+            os.environ["TRITON_CACHE_DIR"] = tmpdir
+            yield tmpdir
+        finally:
+            os.environ.pop("TRITON_CACHE_DIR", None)
 
 
 def pytest_configure(config):

python/test/unit/runtime/test_cache.py

@@ -164,6 +164,8 @@ def inc_counter(*args, **kwargs):
     for i in range(10):
         kernel[(1, )](x, 1, BLOCK=1024)
     assert counter == 1
+    device = getattr(torch, device).current_device()
+    kernel.device_caches[device][0].clear()
 
 
 @pytest.mark.parametrize('mode', ['enable', 'disable', 'disable_on_alignment'])
@@ -181,6 +183,10 @@ def inc_counter(*args, **kwargs):
     for i in [1, 2, 4, 8, 16, 32]:
         function[(1, )](x, i, BLOCK=512)
     assert counter == target
+    device = getattr(torch, device).current_device()
+    kernel.device_caches[device][0].clear()
+    kernel_nospec.device_caches[device][0].clear()
+    kernel_nospec_on_alignment.device_caches[device][0].clear()
 
 
 def test_annotation(device):
@@ -489,7 +495,7 @@ def cache_hook(*args, **kwargs):
     assert specialization_data is not None
 
     # clear the cache
-    shutil.rmtree(fresh_triton_cache, ignore_errors=True)
+    shutil.rmtree(fresh_triton_cache)
     kernel_add.device_caches[device][0].clear()
 
     # preload the kernel

python/test/unit/runtime/test_subproc.py

@@ -87,7 +87,7 @@ def test_compile_in_forked_subproc_with_forced_gc(fresh_triton_cache) -> None:
     compile_empty_kernel_with_gc()
 
     # stage 2.p
-    shutil.rmtree(fresh_triton_cache, ignore_errors=True)
+    shutil.rmtree(fresh_triton_cache)
     mp_ctx = multiprocessing.get_context(start_method)
     proc = mp_ctx.Process(target=compile_empty_kernel_with_gc)
 

python/triton/runtime/cache.py

@@ -131,15 +131,7 @@ def put(self, data, filename, binary=True) -> str:
             f.write(data)
         # Replace is guaranteed to be atomic on POSIX systems if it succeeds
         # so filepath cannot see a partial write
-        try:
-            os.replace(temp_path, filepath)
-        except PermissionError:
-            # Ignore PermissionError on Windows because it happens when another process already
-            # put a file into the cache and locked it by opening it.
-            if os.name == "nt":
-                os.remove(temp_path)
-            else:
-                raise
+        os.replace(temp_path, filepath)
         os.removedirs(temp_dir)
         return filepath
 

third_party/intel/backend/driver.py

@@ -172,7 +172,6 @@ class TritonLauncher:
 
     def __init__(self, cache_path: str):
         self.shared_library = ctypes.PyDLL(cache_path)
-        # breakpoint()
         self.shared_library.launch.restype = ctypes.py_object
         self.shared_library.launch.argtypes = (ctypes.py_object, )
 

third_party/intel/python/test/conftest.py

@@ -14,13 +14,9 @@ def device(request):
 
 @pytest.fixture
 def fresh_triton_cache():
-    try:
-        with tempfile.TemporaryDirectory() as tmpdir:
-            try:
-                os.environ["TRITON_CACHE_DIR"] = tmpdir
-                yield tmpdir
-            finally:
-                os.environ.pop("TRITON_CACHE_DIR", None)
-    except OSError:
-        # Ignore errors, such as PermissionError, on Windows
-        pass
+    with tempfile.TemporaryDirectory() as tmpdir:
+        try:
+            os.environ["TRITON_CACHE_DIR"] = tmpdir
+            yield tmpdir
+        finally:
+            os.environ.pop("TRITON_CACHE_DIR", None)