| 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 |
| 8 | from tests.compile.backend import TestBackend |
| 9 | from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass |
| 10 | from vllm.config import CompilationConfig, CompilationMode, PassConfig, VllmConfig |
| 11 | from vllm.platforms import current_platform |
| 12 | |
| 13 | DEVICE_TYPE = current_platform.device_type |
| 14 | |
| 15 | |
| 16 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32]) |
| 17 | # Important edge case is when `num_tokens == buffer_size` |
| 18 | @pytest.mark.parametrize( |
| 19 | ("num_tokens", "buffer_size"), [(256, 256), (256, 512), (1024, 1024), (1024, 1025)] |
| 20 | ) |
| 21 | @pytest.mark.parametrize("hidden_size", [64, 4096]) |
| 22 | def test_noop_elimination(dtype, num_tokens, hidden_size, buffer_size): |
| 23 | torch.set_default_device(DEVICE_TYPE) |
| 24 | torch.set_default_dtype(dtype) |
| 25 | torch.manual_seed(1) |
| 26 | |
| 27 | class Model(torch.nn.Module): |
| 28 | def __init__(self) -> None: |
| 29 | super().__init__() |
| 30 | # Avoid using empty, since on rocm torch.empty |
| 31 | # does not initialize the memory. |
| 32 | self.pos_embed = torch.randn(buffer_size, hidden_size, dtype=dtype) |
| 33 | |
| 34 | def forward(self, x): |
| 35 | # Avoid += to prevent inplace addition. |
| 36 | x = x + self.pos_embed[: x.shape[0]] |
| 37 | # Chain of reshapes |
| 38 | y = x.reshape(-1, 128, 32) |
| 39 | z = y.reshape(-1, 4096) |
| 40 | # No-op reshape |
| 41 | a = z.reshape(-1, 4096) |
| 42 | # Final reshape that should remain |
| 43 | b = a.reshape(-1, 128, 32) |
| 44 | # No-op slice |
| 45 | c = b[0 : b.shape[0]] |
| 46 | # The pass should replace the result of this op with `c`. |
| 47 | d = torch.slice_scatter( |
| 48 | torch.ones_like(c), # Dummy tensor to be scattered into |
| 49 | c, # Source tensor |
| 50 | 0, # dim |
| 51 | 0, # start |
| 52 | c.shape[0], # end |
| 53 | ) |
| 54 | return d |
| 55 | |
| 56 | vllm_config = VllmConfig( |
| 57 | compilation_config=CompilationConfig( |
| 58 | mode=CompilationMode.VLLM_COMPILE, |
| 59 | pass_config=PassConfig(eliminate_noops=True), |
| 60 | ) |
| 61 | ) |
| 62 | with vllm.config.set_current_vllm_config(vllm_config): |
| 63 | noop_pass = NoOpEliminationPass(vllm_config) |
| 64 | |
| 65 | backend = TestBackend(noop_pass) |
| 66 | |
| 67 | model = Model() |
| 68 | # First dimension dynamic |
| 69 | x = torch.rand(num_tokens, hidden_size) |
| 70 | torch._dynamo.mark_dynamic(x, 0) |
| 71 | |
| 72 | result = model(x) |
| 73 | |
| 74 | model2 = torch.compile(model, backend=backend) |
| 75 | result2 = model2(x) |
| 76 | |
| 77 | ATOL, RTOL = (2e-3, 2e-3) |
| 78 | torch.testing.assert_close(result, result2, atol=ATOL, rtol=RTOL) |
| 79 | |
| 80 | # The no-op reshape and slice should be eliminated. |
| 81 | # The initial slice on the positional embedding should remain. |
| 82 | # The chain of reshapes should be fused into a single reshape. |
| 83 | assert backend.op_count(torch.ops.aten.reshape.default) == 1 |
| 84 | assert backend.op_count(torch.ops.aten.slice.Tensor) == 1 |
| 85 | assert backend.op_count(torch.ops.aten.slice_scatter.default) == 0 |
| 86 | |
| 87 | |
| 88 | def test_non_noop_slice_preserved(): |
| 89 | """Ensure that a slice with end=-1 (dropping last row) is NOT eliminated. |
| 90 | |
| 91 | Regression test for a bug where end=-1 was treated like an inferred |
| 92 | dimension (reshape semantics) leading to incorrect elimination. |
| 93 | """ |
| 94 | torch.set_default_device(DEVICE_TYPE) |
| 95 | x = torch.randn(16, 16) |
| 96 | |
| 97 | class SliceModel(torch.nn.Module): |
| 98 | def forward(self, x): |
| 99 | base = x.clone() |
| 100 | src = torch.ones(15, 16) |
| 101 | y = torch.slice_scatter(base, src, dim=0, start=0, end=-1) |
| 102 | return x[0:-1, :], y |
| 103 | |
| 104 | vllm_config = VllmConfig( |
| 105 | compilation_config=CompilationConfig( |
| 106 | mode=CompilationMode.VLLM_COMPILE, |
| 107 | pass_config=PassConfig(eliminate_noops=True), |
| 108 | ) |
| 109 | ) |
| 110 | with vllm.config.set_current_vllm_config(vllm_config): |
| 111 | noop_pass = NoOpEliminationPass(vllm_config) |
| 112 | backend = TestBackend(noop_pass) |
| 113 | model = SliceModel() |
| 114 | ref = model(x) |
| 115 | compiled = torch.compile(model, backend=backend) |
| 116 | out = compiled(x) |
| 117 | torch.testing.assert_close(ref, out) |
| 118 | # The slice should remain (not a no-op). |
| 119 | assert backend.op_count(torch.ops.aten.slice.Tensor) == 1 |
| 120 | assert backend.op_count(torch.ops.aten.slice_scatter.default) == 1 |