|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +""" |
| 3 | +Test the piecewise compilation with a simple model so that we |
| 4 | +can exactly calculate the expected output and side effects. |
| 5 | +""" |
| 6 | + |
| 7 | +import pytest |
| 8 | +import torch |
| 9 | +from torch import nn |
| 10 | +from torch.library import Library |
| 11 | +from vllm.compilation.counter import compilation_counter |
| 12 | +from vllm.compilation.decorators import support_torch_compile |
| 13 | +from vllm.config import (CompilationConfig, CompilationLevel, VllmConfig, |
| 14 | + set_current_vllm_config) |
| 15 | +from vllm.utils import direct_register_custom_op |
| 16 | + |
| 17 | +global_counter = 0 |
| 18 | + |
| 19 | +# create a library to hold the custom op |
| 20 | +silly_lib = Library("silly", "FRAGMENT") # noqa |
| 21 | + |
| 22 | + |
| 23 | +def silly_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, |
| 24 | + out: torch.Tensor) -> None: |
| 25 | + global global_counter |
| 26 | + global_counter += 1 |
| 27 | + print(f"{global_counter=}") |
| 28 | + out.copy_(q) |
| 29 | + out[0] += 1 |
| 30 | + |
| 31 | + |
| 32 | +def silly_attention_fake(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, |
| 33 | + out: torch.Tensor) -> None: |
| 34 | + return |
| 35 | + |
| 36 | + |
| 37 | +direct_register_custom_op( |
| 38 | + op_name="attention", |
| 39 | + op_func=silly_attention, |
| 40 | + mutates_args=["out"], |
| 41 | + fake_impl=silly_attention_fake, |
| 42 | + dispatch_key="PrivateUse1", |
| 43 | + target_lib=silly_lib, |
| 44 | +) |
| 45 | + |
| 46 | + |
| 47 | +@support_torch_compile |
| 48 | +class SillyModel(nn.Module): |
| 49 | + |
| 50 | + def __init__(self, |
| 51 | + *, |
| 52 | + vllm_config: VllmConfig, |
| 53 | + prefix: str = "", |
| 54 | + **kwargs) -> None: |
| 55 | + super().__init__() |
| 56 | + |
| 57 | + def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 58 | + """ |
| 59 | + Overall effect: |
| 60 | + x += 1 |
| 61 | + x[0] += 2 |
| 62 | + global_counter += 2 |
| 63 | + """ |
| 64 | + x = x + 1 |
| 65 | + x = x + 2 |
| 66 | + out = torch.empty_like(x) |
| 67 | + torch.ops.silly.attention(x, x, x, out) |
| 68 | + x = out |
| 69 | + x = x - 2 |
| 70 | + x = x - 1 |
| 71 | + out = torch.empty_like(x) |
| 72 | + torch.ops.silly.attention(x, x, x, out) |
| 73 | + x = out |
| 74 | + x = x + 1 |
| 75 | + return x |
| 76 | + |
| 77 | + |
| 78 | +@pytest.mark.skipif(True, reason="requires unreleased components") |
| 79 | +def test_simple_piecewise_compile(): |
| 80 | + |
| 81 | + vllm_config = VllmConfig(compilation_config=CompilationConfig( |
| 82 | + level=CompilationLevel.PIECEWISE, |
| 83 | + use_inductor=False, |
| 84 | + use_cudagraph=True, |
| 85 | + splitting_ops=["silly.attention"], |
| 86 | + cudagraph_copy_inputs=True, |
| 87 | + cudagraph_capture_sizes=[1, 2], |
| 88 | + )) |
| 89 | + vllm_config.compilation_config.pass_config.enable_fusion = False |
| 90 | + with set_current_vllm_config(vllm_config): |
| 91 | + model = SillyModel(vllm_config=vllm_config, prefix="") |
| 92 | + |
| 93 | + inputs = torch.randn(100).npu() |
| 94 | + |
| 95 | + with compilation_counter.expect( |
| 96 | + num_graphs_seen=1, # one graph for the model |
| 97 | + num_piecewise_graphs_seen=5, # 2 * num_layers + 1 |
| 98 | + num_piecewise_capturable_graphs_seen=3, # 1 + num_layers |
| 99 | + num_backend_compilations=3, # num_piecewise_capturable_graphs_seen |
| 100 | + num_cudagraph_caputured= |
| 101 | + 6, # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen |
| 102 | + ): |
| 103 | + |
| 104 | + model(inputs) |
| 105 | + |
| 106 | + model(torch.randn(2).npu()) |
| 107 | + model(torch.randn(1).npu()) |
| 108 | + |
| 109 | + input = torch.zeros(2).npu() |
| 110 | + global global_counter |
| 111 | + global_counter = 0 |
| 112 | + output = model(input) |
| 113 | + assert global_counter == 2 |
| 114 | + assert torch.allclose(output.cpu(), torch.tensor([3.0, 1.0])) |
| 115 | + |
| 116 | + |
| 117 | +if __name__ == "__main__": |
| 118 | + test_simple_piecewise_compile() |
0 commit comments