| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Correctness tests for the fused Inkling vision/audio tower kernels. |
| 4 | |
| 5 | Each kernel is checked against a pure-PyTorch reference implementing the |
| 6 | towers' original op sequence (native ``rms_norm`` semantics: fp32 variance, |
| 7 | bf16-rounded weight multiply; exact-erf GELU; fp32-accumulated embedding sum). |
| 8 | The fused kernels keep the same accumulation dtype and rounding points, so |
| 9 | outputs differ from the reference only by reduction order — a few bf16 ulps. |
| 10 | """ |
| 11 | |
| 12 | import pytest |
| 13 | import torch |
| 14 | import torch.nn.functional as F |
| 15 | |
| 16 | from vllm.platforms import current_platform |
| 17 | |
| 18 | if not current_platform.is_cuda(): |
| 19 | pytest.skip("requires CUDA", allow_module_level=True) |
| 20 | |
| 21 | from vllm.models.inkling.common.towers import fold_timespace_to_depth |
| 22 | from vllm.models.inkling.nvidia.ops.mm_towers import dmel_embed_sum_norm, rmsnorm_gelu |
| 23 | |
| 24 | DTYPE = torch.bfloat16 |
| 25 | |
| 26 | |
| 27 | def _bf16_spacing(x: torch.Tensor) -> torch.Tensor: |
| 28 | return torch.exp2(torch.floor(torch.log2(x.float().abs().clamp(min=1e-30)))) * ( |
| 29 | 2**-7 |
| 30 | ) |
| 31 | |
| 32 | |
| 33 | def _assert_close_ulps( |
| 34 | got: torch.Tensor, |
| 35 | ref: torch.Tensor, |
| 36 | max_ulps: float = 4.0, |
| 37 | atol: float = 1e-3, |
| 38 | pre_act: torch.Tensor | None = None, |
| 39 | ) -> None: |
| 40 | """Fused vs reference differ only by fp32 reduction order, which shows up |
| 41 | as a few bf16 ulps at any magnitude — a fixed rtol misrepresents that, so |
| 42 | compare against the reference's local ulp spacing. ``pre_act`` (the |
| 43 | reference pre-activation) adds derivative-propagated slack: near |
| 44 | ``gelu(x) ~ 0`` an ulp-level input flip legitimately moves the output by |
| 45 | many of ITS (tiny) ulps, bounded by |gelu'| <= 1.13 times the input ulps.""" |
| 46 | g, r = got.float(), ref.float() |
| 47 | tol = torch.clamp(max_ulps * _bf16_spacing(ref), min=atol) |
| 48 | if pre_act is not None: |
| 49 | tol = tol + 2.5 * _bf16_spacing(pre_act) |
| 50 | bad = (g - r).abs() > tol |
| 51 | assert not bad.any(), ( |
| 52 | f"{int(bad.sum())}/{ref.numel()} elements beyond tolerance; " |
| 53 | f"max abs diff {(g - r).abs().max().item():.3e}" |
| 54 | ) |
| 55 | |
| 56 | |
| 57 | def _ref_rms_norm(x: torch.Tensor, weight: torch.Tensor, eps: float) -> torch.Tensor: |
| 58 | # Matches ir.ops.rms_norm: fp32 normalize, bf16 weight multiply. |
| 59 | x32 = x.float() |
| 60 | var = x32.pow(2).mean(dim=-1, keepdim=True) |
| 61 | xn = x32 * torch.rsqrt(var + eps) |
| 62 | return (xn.to(weight.dtype) * weight).to(x.dtype) |
| 63 | |
| 64 | |
| 65 | @pytest.mark.parametrize("rows", [1, 7, 1200, 38400]) |
| 66 | @pytest.mark.parametrize("dim", [128, 320, 4800, 6144]) |
| 67 | @pytest.mark.parametrize("gelu", [True, False]) |
| 68 | def test_rmsnorm_gelu(rows: int, dim: int, gelu: bool) -> None: |
| 69 | torch.manual_seed(rows * dim) |
| 70 | x = torch.randn(rows, dim, device="cuda", dtype=DTYPE) |
| 71 | w = torch.randn(dim, device="cuda", dtype=DTYPE) |
| 72 | |
| 73 | h = _ref_rms_norm(x, w, 1e-5) |
| 74 | ref = F.gelu(h) if gelu else h |
| 75 | got = rmsnorm_gelu(x, w, 1e-5, gelu=gelu) |
| 76 | |
| 77 | _assert_close_ulps(got, ref, pre_act=h if gelu else None) |
| 78 | |
| 79 | |
| 80 | @pytest.mark.parametrize("n", [1, 5, 64]) |
| 81 | @pytest.mark.parametrize( |
| 82 | "shape,fold", |
| 83 | [ |
| 84 | ((2, 8, 8, 128), (1, 2)), # the real L0 -> L1 vision transition |
| 85 | ((2, 4, 4, 320), (1, 2)), |
| 86 | ((2, 8, 8, 128), (2, 2)), # temporal + spatial fold |
| 87 | ], |
| 88 | ) |
| 89 | def test_rmsnorm_gelu_folded_store( |
| 90 | n: int, shape: tuple[int, ...], fold: tuple[int, int] |
| 91 | ) -> None: |
| 92 | torch.manual_seed(n) |
| 93 | x = torch.randn(n, *shape, device="cuda", dtype=DTYPE) |
| 94 | w = torch.randn(shape[-1], device="cuda", dtype=DTYPE) |
| 95 | |
| 96 | plain = rmsnorm_gelu(x, w, 1e-5, gelu=True) |
| 97 | ref = fold_timespace_to_depth(plain, *fold) |
| 98 | got = rmsnorm_gelu(x, w, 1e-5, gelu=True, fold=fold) |
| 99 | |
| 100 | # The folded store is a pure permutation of the unfolded output. |
| 101 | assert got.shape == ref.shape |
| 102 | torch.testing.assert_close(got, ref, rtol=0, atol=0) |
| 103 | |
| 104 | |
| 105 | @pytest.mark.parametrize("num_frames", [1, 3, 100, 4097]) |
| 106 | @pytest.mark.parametrize("with_norm", [True, False]) |
| 107 | def test_dmel_embed_sum_norm(num_frames: int, with_norm: bool) -> None: |
| 108 | torch.manual_seed(num_frames) |
| 109 | n_bins, vocab, dim = 80, 16, 6144 |
| 110 | idx = torch.randint( |
| 111 | 0, vocab, (num_frames, n_bins), device="cuda", dtype=torch.int32 |
| 112 | ) |
| 113 | table = torch.randn(n_bins * vocab, dim, device="cuda", dtype=DTYPE) |
| 114 | norm_w = torch.randn(dim, device="cuda", dtype=DTYPE) |
| 115 | |
| 116 | flat = (torch.arange(n_bins, device="cuda", dtype=torch.int32) * vocab).unsqueeze( |
| 117 | 0 |
| 118 | ) + idx |
| 119 | ref = ( |
| 120 | F.embedding(flat.reshape(-1).long(), table) |
| 121 | .reshape(num_frames, n_bins, dim) |
| 122 | .sum(dim=1) |
| 123 | ) |
| 124 | if with_norm: |
| 125 | ref = _ref_rms_norm(ref, norm_w, 1e-6) |
| 126 | |
| 127 | got = dmel_embed_sum_norm(idx, table, norm_w if with_norm else None, 1e-6) |
| 128 | _assert_close_ulps(got, ref) |
| 129 | |
| 130 | |
| 131 | def test_empty_inputs() -> None: |
| 132 | x = torch.empty(0, 320, device="cuda", dtype=DTYPE) |
| 133 | w = torch.randn(320, device="cuda", dtype=DTYPE) |
| 134 | assert rmsnorm_gelu(x, w, 1e-5).shape == (0, 320) |
| 135 | |
| 136 | idx = torch.empty(0, 80, device="cuda", dtype=torch.int32) |
| 137 | table = torch.randn(1280, 6144, device="cuda", dtype=DTYPE) |
| 138 | assert dmel_embed_sum_norm(idx, table, None, 0.0).shape == (0, 6144) |