| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | |
| 7 | import vllm.ir.ops |
| 8 | from tests.compile.backend import TestBackend |
| 9 | from vllm.compilation.passes.fusion.add_rms_fusion import ( |
| 10 | AddRMSNormFusionPass, |
| 11 | RMSNormReshapeFusionPass, |
| 12 | ) |
| 13 | from vllm.compilation.passes.fx_utils import find_op_nodes, is_func |
| 14 | from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass |
| 15 | from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass |
| 16 | from vllm.config import CompilationConfig, CompilationMode, VllmConfig |
| 17 | from vllm.platforms import current_platform |
| 18 | |
| 19 | pytestmark = pytest.mark.skipif( |
| 20 | not current_platform.is_cuda_alike(), reason="Requires CUDA or ROCm" |
| 21 | ) |
| 22 | |
| 23 | |
| 24 | class RMSNormModel(torch.nn.Module): |
| 25 | def __init__(self, hidden_size: int) -> None: |
| 26 | super().__init__() |
| 27 | self.weight = torch.nn.Parameter(torch.randn(hidden_size)) |
| 28 | |
| 29 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 30 | rms = vllm.ir.ops.rms_norm(x, self.weight, 1e-6) |
| 31 | return rms.reshape(-1, rms.shape[-1]) |
| 32 | |
| 33 | |
| 34 | class AddRMSNormModel(torch.nn.Module): |
| 35 | def __init__(self, hidden_size: int, residual_first: bool) -> None: |
| 36 | super().__init__() |
| 37 | self.weight = torch.nn.Parameter(torch.randn(hidden_size)) |
| 38 | self.residual_first = residual_first |
| 39 | |
| 40 | def forward( |
| 41 | self, x: torch.Tensor, residual: torch.Tensor |
| 42 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 43 | residual = residual + x if self.residual_first else x + residual |
| 44 | rms = vllm.ir.ops.rms_norm(residual, self.weight, 1e-6) |
| 45 | return rms.reshape(-1, rms.shape[-1]), residual |
| 46 | |
| 47 | def ops_in_model_before(self): |
| 48 | return [torch.ops.aten.add, torch.ops.vllm_ir.rms_norm] |
| 49 | |
| 50 | def ops_in_model_after(self): |
| 51 | return [torch.ops.vllm_ir.fused_add_rms_norm] |
| 52 | |
| 53 | |
| 54 | def _run_fusion_test(model, config, passes, *inputs): |
| 55 | backend = TestBackend(NoOpEliminationPass(config), *passes, PostCleanupPass(config)) |
| 56 | outputs_unfused = model(*inputs) |
| 57 | outputs_fused = torch.compile(model, backend=backend)(*inputs) |
| 58 | torch.testing.assert_close(outputs_fused, outputs_unfused) |
| 59 | return backend |
| 60 | |
| 61 | |
| 62 | def _is_reshape(node): |
| 63 | return isinstance(node, torch.fx.Node) and is_func( |
| 64 | node, torch.ops.aten.reshape.default |
| 65 | ) |
| 66 | |
| 67 | |
| 68 | @pytest.fixture |
| 69 | def vllm_config(): |
| 70 | config = VllmConfig( |
| 71 | compilation_config=CompilationConfig(mode=CompilationMode.VLLM_COMPILE) |
| 72 | ) |
| 73 | with vllm.config.set_current_vllm_config(config): |
| 74 | torch.set_default_device("cuda") |
| 75 | torch.set_default_dtype(torch.bfloat16) |
| 76 | torch.manual_seed(0) |
| 77 | yield config |
| 78 | |
| 79 | |
| 80 | def test_rmsnorm_reshape_fusion(vllm_config): |
| 81 | fusion_pass = RMSNormReshapeFusionPass(vllm_config) |
| 82 | model = RMSNormModel(hidden_size=32) |
| 83 | x = torch.randn(2, 7, 32) |
| 84 | backend = _run_fusion_test(model, vllm_config, [fusion_pass], x) |
| 85 | |
| 86 | assert fusion_pass.matched_count == 1 |
| 87 | (rms_node,) = find_op_nodes(torch.ops.vllm_ir.rms_norm, backend.graph_post_pass) |
| 88 | assert _is_reshape(rms_node.args[0]) |
| 89 | |
| 90 | |
| 91 | @pytest.mark.parametrize("residual_first", [True, False]) |
| 92 | def test_add_rmsnorm_reshape_fusion(vllm_config, residual_first): |
| 93 | add_fusion = AddRMSNormFusionPass(vllm_config) |
| 94 | reshape_fusion = RMSNormReshapeFusionPass(vllm_config) |
| 95 | model = AddRMSNormModel(hidden_size=32, residual_first=residual_first) |
| 96 | x = torch.randn(2, 7, 32) |
| 97 | residual = torch.randn_like(x) |
| 98 | backend = _run_fusion_test( |
| 99 | model, vllm_config, [add_fusion, reshape_fusion], x, residual |
| 100 | ) |
| 101 | |
| 102 | assert add_fusion.matched_count == 1 |
| 103 | assert reshape_fusion.matched_count == 1 |
| 104 | backend.check_before_ops(model.ops_in_model_before()) |
| 105 | backend.check_after_ops(model.ops_in_model_after()) |