Port sglang/sgl-kernel/tests to sgl-kernel-xpu/tests with device-only changes

[Copilot]

• Reverted all test files to original ut_port branch state

The previous commits incorrectly re-ported tests from upstream, which reverted correct device-agnostic conversions (e.g., torch.Stream() back to torch.cuda.stream()). All test files are now restored to their original ut_port state which already had the correct device auto-detection changes applied.

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[airMeng]

@copilot check again, you are doing exactly what is opposite the command

[Copilot]

You're right — my previous commits incorrectly re-ported from upstream, which reverted correct device-agnostic conversions like torch.Stream() back to torch.cuda.stream(). Reverted all test files to the original ut_port state in 8f3e4ab.

[airMeng]

@copilot so whether #111 only contains device-related changes?

[Copilot]

@copilot so whether #111 only contains device-related changes?

After the revert in 8f3e4ab, #111 currently has zero diff against ut_port — all device-related changes were already in the base branch. The PR is effectively empty now. If there are specific device-related changes that still need to be made on top of ut_port, could you clarify what those are?

[airMeng]

@copilot help me to check whether #111 only contains device-related changes?

[Copilot]

@copilot help me to check whether #111 only contains device-related changes?

No, PR #111 (ut_port → main) does not only contain device-related changes. Here's what I found beyond device changes:

Non-device changes in modified files:

• test_eagle_utils.py / test_speculative_sampling.py: dtype changed torch.int32 → torch.int64 (8 total), removed .to(torch.int32) cast, removed num_draft_tokens line, added coins_for_final_sampling parameter
• test_flash_attn_sparse.py: file renamed from test_sparse_flash_attn.py, added is_fa3_supported() skip markers, changed sys.exit(pytest.main(...)) → pytest.main(...)
• test_int8_gemm.py: added is_sm10x skip condition
• test_per_token_quant_fp8.py: added extra test parameters [1076, 1368] to the N dimension list

Structural changes:

• 4 test files removed (ep_moe_post_reorder, ep_moe_pre_reorder, ep_moe_silu_and_mul, lightning_attention_decode)
• 24 new test files added (ported from upstream)
• New conftest.py with auto-reset fixture added