Avoid nvJitLink cache crashes during incremental LTO

cuda_core/cuda/core/_linker.pyi

@@ -153,7 +153,9 @@ class LinkerOptions:
     incremental : bool, optional
         Perform an incremental link. The result can be passed
         directly to a later :class:`Linker`. Requires nvJitLink 13.2 or newer
-        and is not supported by the driver linker backend.
+        and is not supported by the driver linker backend. Intermediate
+        caching is disabled automatically when combined with link-time
+        optimization.
         Default: False.
     ptx : bool, optional
         Emit PTX after linking instead of CUBIN; only supported with ``link_time_optimization=True``.
@@ -198,6 +200,8 @@ class LinkerOptions:
         Default: 1.
     no_cache : bool, optional
         Do not cache the intermediate steps of nvJitLink.
+        This is also enabled automatically for incremental links that use
+        link-time optimization.
         Default: False.
     numba_debug : bool, optional
         Non-functional. ``numba_debug`` is an NVVM/NVRTC *compiler* option;

cuda_core/cuda/core/_linker.pyx

@@ -271,7 +271,9 @@ class LinkerOptions:
     incremental : bool, optional
         Perform an incremental link. The result can be passed
         directly to a later :class:`Linker`. Requires nvJitLink 13.2 or newer
-        and is not supported by the driver linker backend.
+        and is not supported by the driver linker backend. Intermediate
+        caching is disabled automatically when combined with link-time
+        optimization.
         Default: False.
     ptx : bool, optional
         Emit PTX after linking instead of CUBIN; only supported with ``link_time_optimization=True``.
@@ -316,6 +318,8 @@ class LinkerOptions:
         Default: 1.
     no_cache : bool, optional
         Do not cache the intermediate steps of nvJitLink.
+        This is also enabled automatically for incremental links that use
+        link-time optimization.
         Default: False.
     numba_debug : bool, optional
         Non-functional. ``numba_debug`` is an NVVM/NVRTC *compiler* option;
@@ -431,7 +435,10 @@ class LinkerOptions:
             options.append(f"-split-compile={self.split_compile}")
         if self.split_compile_extended is not None:
             options.append(f"-split-compile-extended={self.split_compile_extended}")
-        if self.no_cache is True:
+        # Repeated incremental LTO links can crash in nvJitLink's intermediate
+        # cache (reproduced with 13.4.92 on Windows). Avoid reusing cached
+        # state for each incremental stage.
+        if self.no_cache is True or (self.incremental and self.link_time_optimization):
             options.append("-no-cache")
 
         if as_bytes:

cuda_core/tests/test_linker.py

@@ -218,6 +218,19 @@ def test_linker_options_incremental_as_bytes(value, expected_count):
     assert options.as_bytes().count(b"-r") == expected_count
 
 
+@pytest.mark.agent_authored(model="gpt-6")
+@pytest.mark.skipif(is_culink_backend, reason="as_bytes() only supported for nvjitlink backend")
+def test_linker_options_incremental_lto_disables_cache():
+    options = LinkerOptions(arch="sm_80", incremental=True, link_time_optimization=True)
+    assert b"-no-cache" in options.as_bytes()
+
+    options = LinkerOptions(arch="sm_80", incremental=True)
+    assert b"-no-cache" not in options.as_bytes()
+
+    options = LinkerOptions(arch="sm_80", link_time_optimization=True)
+    assert b"-no-cache" not in options.as_bytes()
+
+
 @pytest.mark.parametrize("backend", ("invalid", "driver"))
 def test_linker_options_as_bytes_invalid_backend(backend):
     """Test LinkerOptions.as_bytes() with invalid backend"""