Merge branch 'main' into rluo8→fix/ipc-peer-access-skip

Author: rwgk

cuda_bindings/cuda/bindings/_bindings/cydriver.pyx.in

@@ -519,7 +519,7 @@ cdef int _cuPythonInit() except -1 nogil:
     cdef char libPath[260]
 
     with gil, __symbol_lock:
-        usePTDS = os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=0)
+        usePTDS = bool(int(os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=0)))
 
         # Load library
         libPath[0] = 0

cuda_bindings/cuda/bindings/_bindings/cyruntime.pyx.in

@@ -15,7 +15,7 @@ cdef int _cudaPythonInit() except -1 nogil:
         global __usePTDS
 
         with gil:
-            __usePTDS = os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=False)
+            __usePTDS = bool(int(os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=0)))
         __cudaPythonInit = True
         return __usePTDS
 

cuda_bindings/docs/source/release/13.2.1-notes.rst

@@ -0,0 +1,14 @@
+.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+.. SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+.. module:: cuda.bindings
+
+``cuda-bindings`` 13.2.1 Release notes
+======================================
+
+Bugfixes
+--------
+
+* Per-thread default stream mode would be used whenever the
+  ``CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM`` environment variable was set,
+  even if it was set to ``0``.

cuda_bindings/tests/nvml/test_device.py

@@ -4,6 +4,7 @@
 
 from functools import cache
 
+import numpy as np
 import pytest
 
 from cuda.bindings import nvml
@@ -78,7 +79,7 @@ def test_get_nv_link_supported_bw_modes(all_devices):
         assert not hasattr(modes, "total_bw_modes")
 
         for mode in modes.bw_modes:
-            assert isinstance(mode, int)
+            assert isinstance(mode, np.uint8)
 
 
 def test_device_get_pdi(all_devices):

cuda_core/cuda/core/_utils/enum_explanations_helpers.py

@@ -35,9 +35,11 @@
 def _binding_version() -> tuple[int, int, int]:
     """Return the installed ``cuda-bindings`` version, or a conservative old value."""
     try:
-        parts = importlib.metadata.version("cuda-bindings").split(".")[:3]
+        version = importlib.metadata.version("cuda-bindings")
     except importlib.metadata.PackageNotFoundError:
         return (0, 0, 0)  # For very old versions of cuda-python
+
+    parts = version.partition("+")[0].split(".")[:3]
     return tuple(int(v) for v in parts)
 
 

cuda_core/cuda/core/utils/_program_cache/_file_stream.py

@@ -422,11 +422,17 @@ def _path_for_key(self, key: object) -> Path:
         k = _as_key_bytes(key)
         # Hash the key to a fixed-length identifier so arbitrary-length user
         # keys never exceed per-component filename limits (typically 255 on
-        # ext4 / NTFS). With a 256-bit blake2b digest, the cache relies on
-        # cryptographic collision resistance for key uniqueness -- two
-        # distinct keys hashing to the same path is astronomically unlikely
-        # (~2^-128 with the 32-byte digest in use here).
-        digest = hashlib.blake2b(k, digest_size=32).hexdigest()
+        # ext4 / NTFS).
+        #
+        # FIPS: must use a FIPS-approved hash algorithm. FIPS-enforcing
+        # systems can disable non-approved hashlib algorithms (for example
+        # blake2b) at the OpenSSL level. See #2043.
+        #
+        # With a 256-bit SHA-256 digest, the cache relies on collision
+        # resistance for key uniqueness -- two distinct keys hashing to the
+        # same path is astronomically unlikely (~2^128 practical collision
+        # work).
+        digest = hashlib.sha256(k, usedforsecurity=False).hexdigest()
         return self._entries / digest[:2] / digest[2:]
 
     # -- mapping API ---------------------------------------------------------

cuda_core/cuda/core/utils/_program_cache/_keys.py

@@ -35,7 +35,7 @@
 )
 
 # Bump when the key schema changes in a way that invalidates existing caches.
-_KEY_SCHEMA_VERSION = 1
+_KEY_SCHEMA_VERSION = 2
 
 _VALID_CODE_TYPES = frozenset({"c++", "ptx", "nvvm"})
 _VALID_TARGET_TYPES = frozenset({"ptx", "cubin", "ltoir"})
@@ -768,7 +768,10 @@ def make_program_cache_key(
     option_bytes = backend.option_fingerprint(options, target_type)
     name_tags = backend.encode_name_expressions(name_expressions)
 
-    hasher = hashlib.blake2b(digest_size=32)
+    # IMPORTANT: Must use a FIPS-approved hash algorithm (SHA-2 family).
+    # FIPS-enforcing systems can disable non-approved hashlib algorithms
+    # (for example blake2b) at the OpenSSL level. See #2043.
+    hasher = hashlib.sha256(usedforsecurity=False)
 
     def _update(label: str, payload: bytes) -> None:
         hasher.update(label.encode("ascii"))

cuda_core/tests/test_memory.py

@@ -1474,11 +1474,11 @@ def test_pinned_mr_numa_id_default_no_ipc(init_cuda):
     device = Device()
     skip_if_pinned_memory_unsupported(device)
 
-    mr = PinnedMemoryResource(PinnedMemoryResourceOptions())
+    mr = create_pinned_memory_resource_or_xfail(PinnedMemoryResourceOptions(), xfail_device=device)
     assert mr.numa_id == -1
     mr.close()
 
-    mr = PinnedMemoryResource(PinnedMemoryResourceOptions(ipc_enabled=False))
+    mr = create_pinned_memory_resource_or_xfail(PinnedMemoryResourceOptions(ipc_enabled=False), xfail_device=device)
     assert mr.numa_id == -1
     mr.close()
 
@@ -1497,7 +1497,9 @@ def test_pinned_mr_numa_id_default_with_ipc(init_cuda):
     if expected_numa_id < 0:
         pytest.skip("System does not support NUMA")
 
-    mr = PinnedMemoryResource(PinnedMemoryResourceOptions(ipc_enabled=True, max_size=POOL_SIZE))
+    mr = create_pinned_memory_resource_or_xfail(
+        PinnedMemoryResourceOptions(ipc_enabled=True, max_size=POOL_SIZE), xfail_device=device
+    )
     assert mr.numa_id == expected_numa_id
     mr.close()
 
@@ -1511,7 +1513,7 @@ def test_pinned_mr_numa_id_explicit(init_cuda):
     if host_numa_id < 0:
         pytest.skip("System does not support NUMA")
 
-    mr = PinnedMemoryResource(PinnedMemoryResourceOptions(numa_id=host_numa_id))
+    mr = create_pinned_memory_resource_or_xfail(PinnedMemoryResourceOptions(numa_id=host_numa_id), xfail_device=device)
     assert mr.numa_id == host_numa_id
     mr.close()
 
@@ -1520,7 +1522,10 @@ def test_pinned_mr_numa_id_explicit(init_cuda):
     if not supports_ipc_mempool(device):
         pytest.skip("Driver rejects IPC-enabled mempool creation on this platform")
 
-    mr = PinnedMemoryResource(PinnedMemoryResourceOptions(ipc_enabled=True, numa_id=host_numa_id, max_size=POOL_SIZE))
+    mr = create_pinned_memory_resource_or_xfail(
+        PinnedMemoryResourceOptions(ipc_enabled=True, numa_id=host_numa_id, max_size=POOL_SIZE),
+        xfail_device=device,
+    )
     assert mr.numa_id == host_numa_id
     mr.close()
 

cuda_core/tests/test_program_cache.py

@@ -1773,7 +1773,7 @@ def test_filestream_cache_size_cap_counts_tmp_files(tmp_path):
 
 def test_filestream_cache_handles_long_keys(tmp_path):
     """Arbitrary-length keys must not overflow per-component filename limits.
-    The filename is a fixed-length 256-bit blake2b digest; key uniqueness
+    The filename is a fixed-length 256-bit digest; key uniqueness
     relies on the digest's collision resistance."""
     from cuda.core.utils import FileStreamProgramCache
 

scripts/run_tests.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 set -euo pipefail
@@ -133,7 +133,7 @@ status_from_rc() {
 run_pytest() {
   # Run pytest safely under set -e and return its exit code
   set +e
-  python -m pytest "${PYTEST_FLAGS[@]}" "$@"
+  CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM=0 python -m pytest "${PYTEST_FLAGS[@]}" "$@"
   local rc=$?
   set -e
   return ${rc}