| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | import torch |
| 5 | |
| 6 | from vllm import _custom_ops as ops |
| 7 | from vllm.platforms import current_platform |
| 8 | from vllm.utils.flashinfer import flashinfer_scaled_fp8_mm |
| 9 | from vllm.utils.torch_utils import set_random_seed |
| 10 | |
| 11 | if not current_platform.has_device_capability(100): |
| 12 | pytest.skip( |
| 13 | reason="Flashinfer FP8 gemms requires compute capability of 10.0 or above.", |
| 14 | allow_module_level=True, |
| 15 | ) |
| 16 | |
| 17 | DTYPES = [torch.float16, torch.bfloat16] |
| 18 | # m, n, k |
| 19 | SHAPES = [(128, 128, 64), (128, 128, 128), (256, 128, 64), (128, 256, 128)] |
| 20 | PAD_SHAPES = [(150, 128, 64), (128, 128, 96)] |
| 21 | SHAPES.extend(PAD_SHAPES) |
| 22 | |
| 23 | SEEDS = [42] |
| 24 | CUDA_DEVICES = ["cuda:0"] |
| 25 | |
| 26 | |
| 27 | @pytest.mark.parametrize("dtype", DTYPES) |
| 28 | @pytest.mark.parametrize("shape", SHAPES) |
| 29 | @pytest.mark.parametrize("use_bias", [True, False]) |
| 30 | @pytest.mark.parametrize("seed", SEEDS) |
| 31 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 32 | @pytest.mark.parametrize("autotune", [False, True]) |
| 33 | @torch.inference_mode() |
| 34 | def test_flashinfer_fp8_gemm( |
| 35 | dtype: torch.dtype, |
| 36 | shape: tuple[int, int, int], |
| 37 | use_bias: bool, |
| 38 | seed: int, |
| 39 | device: str, |
| 40 | autotune: bool, |
| 41 | ) -> None: |
| 42 | set_random_seed(seed) |
| 43 | m, n, k = shape |
| 44 | a = torch.randn((m, k), dtype=dtype, device=device) |
| 45 | b = torch.randn((n, k), dtype=dtype, device=device) / k |
| 46 | |
| 47 | a_fp8, a_scale = ops.scaled_fp8_quant(a) |
| 48 | b_fp8, b_scale = ops.scaled_fp8_quant(b) |
| 49 | |
| 50 | expected_out = torch.mm( |
| 51 | a_scale * a_fp8.to(dtype=torch.float32), |
| 52 | b_scale * b_fp8.to(dtype=torch.float32).t(), |
| 53 | ).to(dtype=dtype) |
| 54 | |
| 55 | if use_bias: |
| 56 | bias = torch.randn((n,), dtype=dtype, device=device) |
| 57 | expected_out = expected_out + bias |
| 58 | else: |
| 59 | bias = None |
| 60 | |
| 61 | import flashinfer |
| 62 | |
| 63 | with flashinfer.autotune(autotune): |
| 64 | out = flashinfer_scaled_fp8_mm( |
| 65 | a_fp8, |
| 66 | b_fp8.t(), |
| 67 | a_scale, |
| 68 | b_scale, |
| 69 | dtype, |
| 70 | bias=bias, |
| 71 | ) |
| 72 | |
| 73 | torch.testing.assert_close(out, expected_out, atol=1e-2, rtol=1e-2) |