Fix ROCM CI: skipped some tests, added checks, install torchvision

[Xia-Weiwen]

As the title.
These CPU tests require latest torch nightly. However, ROCM CI uses some old version of torch nightly without the latest changes these tests require.
An additional torch version check is also added to scaled_int_mm.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4218

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 1 Pending

As of commit 5e040d2 with merge base 6807454 ():
💚 Looks good so far! There are no failures yet. 💚

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

[pytorch-bot]

Warning: Unknown label ciflow/rocm-mi300.
Currently recognized labels are

• ciflow/benchmark
• ciflow/tutorials
• ciflow/rocm
• ciflow/4xh100
• ciflow/xpu

Please add the new label to .github/pytorch-probot.yml

[Xia-Weiwen]

@claude review

[claude]

Claude finished @Xia-Weiwen's task in 5m 44s —— View job

---

Review of PR #4218: Fix CI: skip some CPU tests if torch is ROCM

• Read PR diff and changed files
• Analyze skip conditions and patterns
• Post review feedback

---

Overall this is a reasonable approach to unblock ROCm CI. The changes are straightforward — skip CPU tests that depend on recent torch nightly features not available in the ROCm torch build. A few observations:

1. Misleading skip message on test_dynamic_quant_gpu_singleline

In test/quantization/test_quant_api.py:167-169, the test test_dynamic_quant_gpu_singleline is skipped with the message "Don't test CPU for ROCM version of torch", but the test name says "gpu". Looking at the implementation, the model defaults to CPU (device="cpu" in example_inputs), so the skip is functionally correct — the test does run on CPU despite the name. However, this is confusing. Consider either:

• Using a more accurate skip message like "Skipping due to incompatibility with ROCm torch nightly", or
• Noting that this test actually runs on CPU (it's not a "GPU" test in practice).

2. Inconsistent skip patterns across the PR

The PR uses two different styles for the same purpose:

• Decorator style (test_integration.py:259): @unittest.skipIf(is_ROCM(), "Don't test CPU for ROCM version of torch")
• Inline style (everywhere else): if device == "cpu" and is_ROCM(): self.skipTest(...)

The inline style makes sense for parametrized tests where the device varies, but test_per_token_linear_cpu uses the decorator style while test_dynamic_quant_gpu_singleline (also non-parametrized, always CPU) uses the inline style. It would be more consistent to use the decorator for both non-parametrized cases.

3. Existing skip_if_rocm decorator is available but unused

There's an existing skip_if_rocm decorator in torchao/testing/utils.py:74 (already imported in test_quant_api.py). For non-parametrized tests like test_per_token_linear_cpu and test_dynamic_quant_gpu_singleline, you could use @skip_if_rocm("CPU test requires latest torch nightly") for consistency with the rest of the codebase. However, skip_if_rocm uses pytest.skip() while the current code uses unittest.skipIf/self.skipTest — so this may be a deliberate choice for unittest compatibility.

4. Minor: Consider a comment explaining why

The skip message "Don't test CPU for ROCM version of torch" tells us what is happening but not why. The PR description explains it well (ROCm CI uses an older torch nightly), but a brief code comment would help future maintainers understand when these skips can be removed (e.g., # ROCm CI torch nightly doesn't include the latest CPU-related changes yet).

Summary

The changes are correct and will achieve the goal of unblocking ROCm CI. The issues above are minor style/consistency nits that don't affect correctness. The PR looks good to merge as-is, but addressing the inconsistencies would improve maintainability.

---

[Copilot]

Pull request overview

This PR adjusts the test suite to avoid running a subset of CPU-targeted quantization/kernel tests when the installed PyTorch build is ROCm, because ROCm CI is pinned to an older nightly that lacks required changes.

Changes:

• Added ROCm-based skips for specific CPU test variants in quantization workflow tests.
• Added ROCm-based skips for specific CPU test variants in kernel autotuner tests.
• Added ROCm-based skips for specific CPU integration tests (including one decorator-based skip).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
test/quantization/test_quant_api.py	Skips a CPU-executed quantization test when running under ROCm.
test/quantization/quantize_/workflows/int8/test_int8_tensor.py	Skips multiple CPU parametrizations for int8 workflow tests under ROCm.
test/kernel/test_autotuner.py	Skips the CPU parametrization of an autotuner-related test under ROCm.
test/integration/test_integration.py	Skips specific CPU integration tests/paths under ROCm (decorator + runtime skip).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

is_ROCM() currently requires torch.cuda.is_available(), so a ROCm-built torch running without a visible GPU would not be detected and this skip would not trigger. If the intent is to skip when torch is a ROCm build (as the PR description suggests), prefer checking torch.version.hip is not None or using the existing skip_if_rocm helper so the condition matches the message/reason.

Also consider making the skip reason more specific (e.g., depends on newer torch nightly) to avoid implying that CPU testing is unsupported on ROCm in general.

Suggested change

	def test_dynamic_quant_gpu_singleline(self):
	if is_ROCM():
	self.skipTest("Don't test CPU for ROCM version of torch")
	@skip_if_rocm(
	"test_dynamic_quant_gpu_singleline not supported on ROCm builds (uses CUDA-only path)"
	)
	def test_dynamic_quant_gpu_singleline(self):

[Copilot]

Same concern as elsewhere: is_ROCM() is gated on torch.cuda.is_available(), so it won’t skip on a ROCm build if no GPU is visible. If the goal is to skip CPU variants specifically for ROCm torch nightlies, consider using torch.version.hip is not None / skip_if_rocm for a build-based check, and keep the skip reason aligned with the actual failure mode (old ROCm nightly).

Suggested change

	if device == "cpu" and is_ROCM():
	if device == "cpu" and torch.version.hip is not None:

[Copilot]

is_ROCM() depends on torch.cuda.is_available(), so this won’t skip on a ROCm build if the GPU is not visible to the process. If the skip is meant to apply to ROCm builds (per PR description), prefer torch.version.hip is not None or the existing skip_if_rocm helper for consistency and correctness.

Suggested change

	if device == "cpu" and is_ROCM():
	if device == "cpu" and (torch.version.hip is not None):

[Copilot]

Using @unittest.skipIf(is_ROCM(), ...) inherits is_ROCM()’s requirement that torch.cuda.is_available() is true. If this is intended to skip on ROCm torch builds (even when no GPU is visible), use a build-level check like torch.version.hip is not None or the repo’s skip_if_rocm helper so the decorator matches the intended condition.

Suggested change

	@unittest.skipIf(is_ROCM(), "Don't test CPU for ROCM version of torch")
	@skip_if_rocm("Don't test CPU for ROCM version of torch")

[Copilot]

This skip uses is_ROCM() which requires torch.cuda.is_available(). If the intent is to skip CPU export tests whenever torch is built with ROCm (regardless of GPU visibility), switch to a build-based condition such as torch.version.hip is not None / skip_if_rocm so the skip reliably triggers in all ROCm environments.

Suggested change

	if test_device == "cpu" and is_ROCM():
	if test_device == "cpu" and getattr(torch.version, "hip", None) is not None:

[jerryzh168]

I feel we should just have an allowlist of tests that runs in rocm, currently it runs all tests: https://github.com/pytorch/ao/blob/main/.github/workflows/regression_test_rocm.yml#L51

[jerryzh168]

OK we can land this to unblock CI first, we'll figure allowlist for rocm a bit later

[pytorch-bot]

Warning: Unknown label ciflow/rocm-mi300.
Currently recognized labels are

• ciflow/benchmark
• ciflow/tutorials
• ciflow/rocm
• ciflow/4xh100
• ciflow/xpu

Please add the new label to .github/pytorch-probot.yml

[jerryzh168]

can you fix the torchvision issue in rocm CI following da257b5 as well

should be fixed by adding torchvision here: https://github.com/pytorch/ao/blob/main/.github/workflows/regression_test_rocm.yml#L25 I think

[Xia-Weiwen]

@jerryzh168 Sure.