| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | import torch.nn.functional as F |
| 7 | |
| 8 | from vllm.model_executor.layers import utils as layer_utils |
| 9 | from vllm.utils.flashinfer import is_flashinfer_cutedsl_bf16_gemm_supported |
| 10 | |
| 11 | |
| 12 | @pytest.fixture(scope="module", autouse=True) |
| 13 | def require_flashinfer_bf16_cutedsl() -> None: |
| 14 | if not torch.accelerator.is_available(): |
| 15 | pytest.skip("CUDA is required") |
| 16 | if not is_flashinfer_cutedsl_bf16_gemm_supported(): |
| 17 | pytest.skip("FlashInfer BF16 cute-dsl backend is unavailable") |
| 18 | |
| 19 | |
| 20 | @pytest.mark.parametrize("n,k", [(1024, 1024), (2048, 3072)]) |
| 21 | @pytest.mark.parametrize( |
| 22 | "m,use_bias,pdl", |
| 23 | [ |
| 24 | (1, False, False), |
| 25 | (8, True, False), |
| 26 | (16, False, True), |
| 27 | (32, True, True), |
| 28 | ], |
| 29 | ) |
| 30 | def test_flashinfer_bf16_cutedsl_correctness( |
| 31 | m: int, |
| 32 | n: int, |
| 33 | k: int, |
| 34 | use_bias: bool, |
| 35 | pdl: bool, |
| 36 | ) -> None: |
| 37 | torch.manual_seed(0) |
| 38 | x = torch.randn(m, k, device="cuda", dtype=torch.bfloat16) * 0.1 |
| 39 | weight = torch.randn(n, k, device="cuda", dtype=torch.bfloat16) * 0.1 |
| 40 | bias = ( |
| 41 | torch.randn(n, device="cuda", dtype=torch.bfloat16) * 0.1 if use_bias else None |
| 42 | ) |
| 43 | |
| 44 | actual = layer_utils.cuda_flashinfer_bf16_gemm_impl( |
| 45 | x, weight, bias, pdl, "flashinfer_cutedsl" |
| 46 | ) |
| 47 | expected = F.linear(x, weight, bias) |
| 48 | |
| 49 | torch.testing.assert_close(actual, expected, rtol=2e-2, atol=2e-1) |