From 13537f8e207ed9d970128667142e6f297bc21f32 Mon Sep 17 00:00:00 2001 From: Brayden Zhong Date: Tue, 16 Jun 2026 11:58:22 -0700 Subject: [PATCH] Unskip Marlin NVFP4 tests (#27589) Co-authored-by: Brayden Zhong Co-authored-by: shaunkotek --- test/manual/models/test_nvidia_nemotron_3_nano_archived.py | 1 - test/registered/jit/test_gptq_marlin.py | 2 -- test/registered/jit/test_moe_wna16_marlin.py | 3 --- 3 files changed, 6 deletions(-) diff --git a/test/manual/models/test_nvidia_nemotron_3_nano_archived.py b/test/manual/models/test_nvidia_nemotron_3_nano_archived.py index edb9f6c54..39a06b27e 100644 --- a/test/manual/models/test_nvidia_nemotron_3_nano_archived.py +++ b/test/manual/models/test_nvidia_nemotron_3_nano_archived.py @@ -44,7 +44,6 @@ class TestNvidiaNemotron3Nano30BBF16FlashInfer(LMEvalMixin, DefaultServerBase): ] + NEMOTRON_3_NANO_THINKING_ARGS -@unittest.skip("Skip, test pass locally but compiling takes too long in CI") @unittest.skipIf( not (is_sm80_supported() or is_sm90_supported()), "NVFP4 Marlin fallback test requires CUDA SM8X/SM9X", diff --git a/test/registered/jit/test_gptq_marlin.py b/test/registered/jit/test_gptq_marlin.py index 495bfcf14..32c482c15 100644 --- a/test/registered/jit/test_gptq_marlin.py +++ b/test/registered/jit/test_gptq_marlin.py @@ -115,7 +115,6 @@ def test_gptq_marlin_gemm( assert max_diff < 0.04 -@pytest.mark.skip(reason="Skip, test pass locally but compiling takes too long in CI") @pytest.mark.skipif( not (is_sm80_supported() or is_sm90_supported()), reason="NVFP4 Marlin fallback tests require CUDA SM8X/SM9X", @@ -142,7 +141,6 @@ def test_nvfp4_marlin_support_and_scale_transforms_sm80_sm90(dtype): assert actual_global_scale.item() == 2.0**119 -@pytest.mark.skip(reason="Skip, test pass locally but compiling takes too long in CI") @pytest.mark.skipif( not (is_sm80_supported() or is_sm90_supported()), reason="NVFP4 Marlin dense numeric test requires CUDA SM80, SM86, or SM90", diff --git a/test/registered/jit/test_moe_wna16_marlin.py b/test/registered/jit/test_moe_wna16_marlin.py index 96e6a0ba3..320cf95aa 100644 --- a/test/registered/jit/test_moe_wna16_marlin.py +++ b/test/registered/jit/test_moe_wna16_marlin.py @@ -349,7 +349,6 @@ def test_moe_wna16_marlin_gemm( torch.testing.assert_close(c_jit, c_aot, rtol=0, atol=0) -@pytest.mark.skip(reason="Skip, test pass locally but compiling takes too long in CI") @pytest.mark.skipif( not (is_sm80_supported() or is_sm90_supported()), reason="Non-gated NVFP4 Marlin fallback test requires CUDA SM8X/SM9X", @@ -413,7 +412,6 @@ def test_fused_marlin_moe_non_gated_relu2(): torch.testing.assert_close(output, output_ref, rtol=0.04, atol=0.04) -@pytest.mark.skip(reason="Skip, test pass locally but compiling takes too long in CI") @pytest.mark.skipif( not (is_sm80_supported() or is_sm90_supported()), reason="NVFP4 Marlin MoE padding test requires CUDA SM8X/SM9X", @@ -510,7 +508,6 @@ def test_fused_marlin_moe_nvfp4_non_gated_padded_intermediate_launches(): assert out.shape == (m, hidden_size) -@pytest.mark.skip(reason="Skip, test pass locally but compiling takes too long in CI") @pytest.mark.skipif( not (is_sm80_supported() or is_sm90_supported()), reason="NVFP4 Marlin MoE numeric test requires CUDA SM80, SM86, or SM90",