| 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 tests.kernels.quantization.nvfp4_utils import ( |
| 7 | FLOAT4_E2M1_MAX, |
| 8 | FLOAT8_E4M3_MAX, |
| 9 | dequantize_nvfp4_to_dtype, |
| 10 | ) |
| 11 | from vllm._custom_ops import scaled_fp4_quant |
| 12 | from vllm.model_executor.layers.activation import SiluAndMul |
| 13 | from vllm.platforms import current_platform |
| 14 | from vllm.utils.torch_utils import set_random_seed |
| 15 | |
| 16 | if not current_platform.has_device_capability(100): |
| 17 | pytest.skip( |
| 18 | reason="Nvfp4 Requires compute capability of 10 or above.", |
| 19 | allow_module_level=True, |
| 20 | ) |
| 21 | |
| 22 | FP4_DTYPE = torch.uint8 |
| 23 | FP8_DTYPE = current_platform.fp8_dtype() |
| 24 | |
| 25 | DTYPES = [torch.float16, torch.bfloat16] |
| 26 | SHAPES = [(128, 256), (128, 128), (256, 256), (256, 128), (1, 256), (64, 256)] |
| 27 | BLOCK_SIZE = 16 |
| 28 | |
| 29 | |
| 30 | @pytest.mark.parametrize("dtype", DTYPES) |
| 31 | @pytest.mark.parametrize("shape", SHAPES) |
| 32 | @torch.inference_mode() |
| 33 | def test_silu_mul_nvfp4_quant( |
| 34 | default_vllm_config, |
| 35 | dtype: torch.dtype, |
| 36 | shape: tuple[int, int], |
| 37 | ) -> None: |
| 38 | set_random_seed(42) |
| 39 | device = "cuda:0" |
| 40 | torch.set_default_device(device) |
| 41 | |
| 42 | x = torch.randn(shape, dtype=dtype) |
| 43 | |
| 44 | # ref op |
| 45 | ref_output = SiluAndMul().forward_native(x) |
| 46 | ref_global_scale = (FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX) / torch.abs( |
| 47 | ref_output |
| 48 | ).max().to(torch.float32) |
| 49 | ref_output_quant, ref_block_scale = scaled_fp4_quant(ref_output, ref_global_scale) |
| 50 | |
| 51 | # fused op |
| 52 | fused_output_quant = torch.empty_like(ref_output_quant) |
| 53 | fused_block_scale = torch.empty_like(ref_block_scale) |
| 54 | torch.ops._C.silu_and_mul_nvfp4_quant( |
| 55 | fused_output_quant, fused_block_scale, x, ref_global_scale |
| 56 | ) |
| 57 | |
| 58 | # check dtype |
| 59 | assert ref_output_quant.dtype == FP4_DTYPE |
| 60 | assert fused_output_quant.dtype == FP4_DTYPE |
| 61 | assert ref_output_quant.shape == fused_output_quant.shape |
| 62 | |
| 63 | assert ref_block_scale.dtype == FP8_DTYPE |
| 64 | assert fused_block_scale.dtype == FP8_DTYPE |
| 65 | assert ref_block_scale.shape == fused_block_scale.shape |
| 66 | |
| 67 | # check dequantized output |
| 68 | ref_output_dequant = dequantize_nvfp4_to_dtype( |
| 69 | ref_output_quant, ref_block_scale, ref_global_scale, dtype, device |
| 70 | ) |
| 71 | fused_output_dequant = dequantize_nvfp4_to_dtype( |
| 72 | fused_output_quant, fused_block_scale, ref_global_scale, dtype, device |
| 73 | ) |
| 74 | |
| 75 | atol, rtol = 3e-1, 3e-1 |
| 76 | torch.testing.assert_close( |
| 77 | ref_output_dequant, fused_output_dequant, atol=atol, rtol=rtol |
| 78 | ) |