[CI/Build] Restrict ERNIE fused RoPE tests to NVIDIA CUDA

[andyluo7]

Purpose

What broke

The AMD Diffusion · Model Test selected the ERNIE-Image fused RoPE tests and failed four cases because try_fused_qk_rotary_emb(...) correctly returned None on ROCm.

Observed in AMD builds #12244 and #12243, on different MI300X workers:

4 failed, 56 passed, 24 skipped, 3203 deselected

Reproduction

On a ROCm worker:

pytest -sv tests/diffusion/models/ \
  -m "core_model and cuda and not (cards_2 or cards_3 or cards_4 or cards_5 or cards_6 or cards_7 or cards_8)" \
  --run-level core_model

Root cause

PyTorch exposes ROCm devices through the torch.cuda namespace, so torch.cuda.is_available() is true on ROCm. The tests therefore ran even though the implementation intentionally enables this fused kernel only when current_omni_platform.is_cuda() is true.

The same tests execute successfully on NVIDIA CUDA in build #15561.

Fix

Use the project platform abstraction in all five GPU-dependent skip guards. This keeps the NVIDIA fused-kernel coverage on CUDA, skips it on ROCm, and leaves the platform-independent failed-key cache test enabled.

Declare the actual CI resources with hardware_test: the five NVIDIA-only tests require one CUDA L4, while the cache-bound test retains one-card coverage on CUDA L4 and ROCm MI325.

The production kernel guard is unchanged.

Test Plan

Local validation:

pre-commit run --files tests/diffusion/models/ernie_image/test_ernie_image_fused_rope.py
All applicable hooks passed.

python3 -m compileall -q tests/diffusion/models/ernie_image/test_ernie_image_fused_rope.py
Passed.

Static hardware-marker contract check
Passed.

Focused pytest cannot run on this macOS host because the available Python environments do not have the project vllm/PyTorch dependencies installed. AMD and CUDA runtime validation is delegated to CI.

vLLM Version: CI image version for the Buildkite runs cited below

vLLM-Omni Commit: 94822606be6d77dcdc2e1b432cb8eb122b1a5821

Test Result

• AMD #12282, mi300_1: Diffusion · Model Test: passed in 6m05s (50 passed, 34 skipped, 3203 deselected). All ten NVIDIA-only cases skipped with NVIDIA CUDA required; the platform-independent cache-bound test ran and passed.
• CUDA #15600, Simple · Diffusion Test: passed in 17m18s (5979 passed, 41 skipped, 555 deselected).
• Local pre-commit: passed.
• Local syntax and hardware-marker contract checks: passed.

AI assistance: Used Codex to investigate the cross-platform failure, make the focused test-only change, and draft this description. I reviewed the one-file diff and the cited CI evidence.

[andyluo7]

Self-review: I verified that this PR changes only the five NVIDIA-only skip guards in the ERNIE-Image fused RoPE test module. The implementation already rejects ROCm through current_omni_platform.is_cuda(), so the test now uses the same platform contract instead of torch.cuda.is_available(), which is true on ROCm. I compared the identical AMD failures in Buildkite #12243 and #12244 with the passing NVIDIA CUDA run #15561. All applicable pre-commit hooks and syntax checks pass; local pytest was unavailable because this macOS environment has no PyTorch installation.

[vllm-omni-review-bot]

This PR was classified as CI work.

CI owner: @yenuo26 @NickCao

Routing: @yenuo26 via CI owner, CODEOWNERS; @NickCao via CODEOWNERS

@andyluo7, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[andyluo7]

Updated exact-head CI evidence for 94822606be6d77dcdc2e1b432cb8eb122b1a5821 after the hardware-marker review change:

• AMD #12282 mi300_1: Diffusion · Model Test: passed in 6m05s (50 passed, 34 skipped, 3203 deselected). All ten NVIDIA-only ERNIE cases were collected and skipped with NVIDIA CUDA required; test_failed_runtime_key_cache_is_bounded remained collected and passed.
• CUDA #15600 Simple · Diffusion Test: passed in 17m18s (5979 passed, 41 skipped, 555 deselected).

This validates both the ROCm routing fix and the new explicit hardware_test resource metadata at the current PR head.

[yenuo26]

Please modify pytest.mark.cuda to hardware_test, to mark the actual machine type and card count it uses.

[andyluo7]

Updated in 9482260. I removed the module-level pytest.mark.cuda and added explicit hardware_test metadata: the five NVIDIA-only fused-RoPE tests use cuda/L4 with one card, while the platform-independent cache-bound test keeps one-card coverage on both cuda/L4 and rocm/MI325. I kept the current_omni_platform.is_cuda() guards because the shared AMD diffusion lane still selects cuda-marked tests. Pre-commit, compileall, and the static decorator-contract check pass. Fresh exact-head validation is also green: AMD #12282 Diffusion · Model Test passed with 50 passed, 34 skipped, 3203 deselected, and CUDA #15600 Simple · Diffusion Test passed with 5979 passed, 41 skipped, 555 deselected.

[andyluo7]

Thanks for the guidance. The requested hardware_test update is now in 9482260, and fresh exact-head AMD #12282 plus CUDA #15600 validation passed. Could you please take another look when you have a chance?

[yenuo26]

LGTM