test: skip int8/bf16 integration tests on pre-Ampere GPUs

[Suchitra-idu]

What

Adds two SM80OrLater skip decorators to tests in test/integration/test_integration.py that fail on pre-Ampere CUDA hardware:

1. test__int_mm_eager_and_torch_compile_numerics — uses torch._int_mm, which requires int8 tensor cores (Ampere+).
2. test_benchmark_model_cuda — compiles a bfloat16 model with torch.compile, but bf16 compilation requires Ampere+ (Inductor itself emits SkipFrame: BF16 is not supported and warns explicitly).

Uses the existing SM80OrLater helper from torch.testing._internal.common_cuda for consistency with how the rest of the PyTorch test suite gates Ampere-only features.

Why

Both tests have always required Ampere features, but the existing skip guards only check torch.cuda.is_available() — they don't check compute capability. As a result, on pre-Ampere CUDA hardware (Turing, Volta, Pascal) they hard-fail instead of skipping cleanly.

Reproduction

Running on a GTX 1650 (Turing, SM 7.5) without these skips:

test__int_mm_eager_and_torch_compile_numerics:

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasLtMatmul
  with ... abType 3 cType 10 computeType 72 scaleType 10

Root cause: _int_mm (int8 × int8 → int32) requires tensor cores, which the GTX 16-series lacks.

test_benchmark_model_cuda:

torch._inductor.exc.InductorError: SkipFrame: BF16 is not supported
UserWarning: NVIDIA GeForce GTX 1650 does not support bfloat16 compilation natively

Root cause: bf16 compilation requires SM 8.0+.

Bonus: also transitively fixes downstream test pollution

Skipping test__int_mm_eager_and_torch_compile_numerics also fixes test_int8_weight_only_quant_subclass_api_3 and _4, which were previously failing with:

AssertionError at torch/_inductor/cudagraph_trees.py:2648
    assert len(node.tensor_weakrefs) == len(node.stack_traces)

These tests are not actually broken on pre-Ampere hardware. Verified empirically by running them in isolation:

pytest test/integration/test_integration.py::TestSubclass::test_int8_weight_only_quant_subclass_api_3 -v
# 1 passed

pytest test/integration/test_integration.py::TestSubclass::test_int8_weight_only_quant_subclass_api_4 -v
# 1 passed

Reviewer note: I did not add explicit skips on _3 and _4 because they do pass on pre-Ampere hardware when not poisoned by an earlier failure.

Test plan

• Ran full pytest test/integration -v locally on a GTX 1650 — all four previously-failing tests now either skip cleanly or pass.
• Verified _3 and _4 pass in isolation (proving they're not actually broken on Turing).
• CPU variants (_0_cpu, _1_cpu, _2_cpu) continue to pass unaffected.
• CI will verify that the gated tests still run (and pass) on Ampere/Hopper runners.

No behavior change on supported hardware; this is a test-only change.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4528

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

This comment was automatically generated by Dr. CI and updates every 15 minutes.