tests: fix parametrize patterns rejected by pytest 9.1.0 (#2212)

Two latent test-code bugs that older pytest tolerated but pytest 9.1.0
flags as collection errors, breaking every Test job on main since the
pytest 9.1.0 release:

* cuda_core/tests/test_utils.py:151 had a stray trailing comma in the
  `parametrize` name string (`"in_arr,"`). pytest 9 now splits names on
  comma and counts them, mismatching against the multi-element value
  tuples. Drop the comma.

* cuda_bindings/tests/test_nvfatbin.py had two tests using
  `@pytest.mark.parametrize("arch", ["sm_80"], indirect=True)` to
  override the fixture-level `arch` parametrization. pytest 9 now
  rejects this combination as "duplicate parametrization of 'arch'".
  Extract the CUBIN-building logic into a `_build_cubin(arch)` helper,
  drop the indirect override on the two tests, and call the helper
  inline with the hardcoded `"sm_80"` they need. Preserves intent (the
  override existed because target arch "75" must not match the CUBIN's
  arch).

Both fixes are pytest-version-agnostic; verified collecting cleanly
under pytest 9.1.0, 9.0.2, and 8.4.2 with minimal reproductions of
each pattern.

Author: leofang

cuda_bindings/tests/test_nvfatbin.py

@@ -121,8 +121,7 @@ def nvcc_smoke(tmpdir) -> str:
     return nvcc
 
 
-@pytest.fixture
-def CUBIN(arch):
+def _build_cubin(arch):
     def CHECK_NVRTC(err):
         if err != nvrtc.nvrtcResult.NVRTC_SUCCESS:
             raise RuntimeError(repr(err))
@@ -141,6 +140,11 @@ def CHECK_NVRTC(err):
     return cubin
 
 
+@pytest.fixture
+def CUBIN(arch):
+    return _build_cubin(arch)
+
+
 # create a valid LTOIR input for testing
 @pytest.fixture
 def LTOIR(arch):
@@ -259,11 +263,11 @@ def test_nvfatbin_add_ptx(PTX, arch):
     nvfatbin.destroy(handle)
 
 
-@pytest.mark.parametrize("arch", ["sm_80"], indirect=True)
-def test_nvfatbin_add_cubin_ELF_SIZE_MISMATCH(CUBIN, arch):
+def test_nvfatbin_add_cubin_ELF_SIZE_MISMATCH():
+    cubin = _build_cubin("sm_80")
     handle = nvfatbin.create([], 0)
     with pytest.raises(nvfatbin.nvFatbinError, match="ERROR_ELF_ARCH_MISMATCH"):
-        nvfatbin.add_cubin(handle, CUBIN, len(CUBIN), "75", "inc")
+        nvfatbin.add_cubin(handle, cubin, len(cubin), "75", "inc")
 
     nvfatbin.destroy(handle)
 
@@ -280,11 +284,11 @@ def test_nvfatbin_add_cubin(CUBIN, arch):
     nvfatbin.destroy(handle)
 
 
-@pytest.mark.parametrize("arch", ["sm_80"], indirect=True)
-def test_nvfatbin_add_cubin_ELF_ARCH_MISMATCH(CUBIN, arch):
+def test_nvfatbin_add_cubin_ELF_ARCH_MISMATCH():
+    cubin = _build_cubin("sm_80")
     handle = nvfatbin.create([], 0)
     with pytest.raises(nvfatbin.nvFatbinError, match="ERROR_ELF_ARCH_MISMATCH"):
-        nvfatbin.add_cubin(handle, CUBIN, len(CUBIN), "75", "inc")
+        nvfatbin.add_cubin(handle, cubin, len(cubin), "75", "inc")
 
     nvfatbin.destroy(handle)
 

cuda_core/tests/test_utils.py

@@ -148,7 +148,7 @@ def _cpu_array_samples():
     return samples
 
 
-@pytest.mark.parametrize("in_arr,", _cpu_array_samples())
+@pytest.mark.parametrize("in_arr", _cpu_array_samples())
 class TestViewCPU:
     def test_args_viewable_as_strided_memory_cpu(self, in_arr):
         @args_viewable_as_strided_memory((0,))