| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from vllm.compilation.passes.fusion.sequence_parallelism import ( |
| 7 | SP_MIN_HIDDEN_SIZE, |
| 8 | SP_MIN_PER_GPU_SIZE_MB, |
| 9 | get_sequence_parallelism_threshold, |
| 10 | ) |
| 11 | |
| 12 | |
| 13 | class TestGetSequenceParallelismThreshold: |
| 14 | """Tests for get_sequence_parallelism_threshold function.""" |
| 15 | |
| 16 | def test_non_cuda_returns_none(self, mock_cuda_platform): |
| 17 | """Non-CUDA platforms should return None.""" |
| 18 | with mock_cuda_platform(is_cuda=False): |
| 19 | result = get_sequence_parallelism_threshold( |
| 20 | hidden_size=8192, tp_size=2, element_size=2 |
| 21 | ) |
| 22 | assert result is None |
| 23 | |
| 24 | def test_unsupported_device_capability_returns_none(self, mock_cuda_platform): |
| 25 | """Unsupported device capabilities (e.g., sm80) should return None.""" |
| 26 | with mock_cuda_platform(capability=(8, 0)): |
| 27 | result = get_sequence_parallelism_threshold( |
| 28 | hidden_size=8192, tp_size=2, element_size=2 |
| 29 | ) |
| 30 | assert result is None |
| 31 | |
| 32 | def test_small_hidden_size_returns_none(self, mock_cuda_platform): |
| 33 | """H100 with hidden_size below threshold should return None.""" |
| 34 | with mock_cuda_platform(capability=(9, 0)): |
| 35 | result = get_sequence_parallelism_threshold( |
| 36 | hidden_size=4096, |
| 37 | tp_size=2, |
| 38 | element_size=2, # 4096 < 8192 |
| 39 | ) |
| 40 | assert result is None |
| 41 | |
| 42 | def test_h100_large_model_returns_threshold(self, mock_cuda_platform): |
| 43 | """H100 with large enough hidden_size should return calculated threshold.""" |
| 44 | with mock_cuda_platform(capability=(9, 0)): |
| 45 | hidden_size = 8192 |
| 46 | tp_size = 2 |
| 47 | element_size = 2 # float16/bfloat16 |
| 48 | |
| 49 | result = get_sequence_parallelism_threshold( |
| 50 | hidden_size=hidden_size, |
| 51 | tp_size=tp_size, |
| 52 | element_size=element_size, |
| 53 | ) |
| 54 | |
| 55 | # Verify calculation: (8 * 2 * 1024 * 1024) // (8192 * 2) = 1024 |
| 56 | MiB = 1024 * 1024 |
| 57 | expected = int( |
| 58 | (SP_MIN_PER_GPU_SIZE_MB[90] * tp_size * MiB) |
| 59 | // (hidden_size * element_size) |
| 60 | ) |
| 61 | assert result == expected |
| 62 | assert result == 1024 |
| 63 | |
| 64 | @pytest.mark.parametrize( |
| 65 | "hidden_size,tp_size,element_size,expected", |
| 66 | [ |
| 67 | # Boundary: exactly at min hidden size threshold, tp_size=1 |
| 68 | # (8 * 1 * 1024 * 1024) // (8192 * 2) = 512 |
| 69 | (8192, 1, 2, 512), |
| 70 | # Larger hidden size reduces token threshold |
| 71 | # (8 * 1 * 1024 * 1024) // (16384 * 2) = 256 |
| 72 | (16384, 1, 2, 256), |
| 73 | # Larger tp_size increases token threshold |
| 74 | # (8 * 4 * 1024 * 1024) // (8192 * 2) = 2048 |
| 75 | (8192, 4, 2, 2048), |
| 76 | # Larger element_size (fp32) reduces token threshold |
| 77 | # (8 * 2 * 1024 * 1024) // (8192 * 4) = 512 |
| 78 | (8192, 2, 4, 512), |
| 79 | ], |
| 80 | ) |
| 81 | def test_threshold_calculation_variations( |
| 82 | self, mock_cuda_platform, hidden_size, tp_size, element_size, expected |
| 83 | ): |
| 84 | """Test threshold calculation with various parameter combinations.""" |
| 85 | with mock_cuda_platform(capability=(9, 0)): |
| 86 | result = get_sequence_parallelism_threshold( |
| 87 | hidden_size=hidden_size, |
| 88 | tp_size=tp_size, |
| 89 | element_size=element_size, |
| 90 | ) |
| 91 | assert result == expected |
| 92 | |
| 93 | def test_hidden_size_boundary(self, mock_cuda_platform): |
| 94 | """Test behavior at the exact hidden_size boundary.""" |
| 95 | with mock_cuda_platform(capability=(9, 0)): |
| 96 | # Just below threshold |
| 97 | result = get_sequence_parallelism_threshold( |
| 98 | hidden_size=SP_MIN_HIDDEN_SIZE[90] - 1, |
| 99 | tp_size=2, |
| 100 | element_size=2, |
| 101 | ) |
| 102 | assert result is None |
| 103 | |
| 104 | # Exactly at threshold |
| 105 | result = get_sequence_parallelism_threshold( |
| 106 | hidden_size=SP_MIN_HIDDEN_SIZE[90], |
| 107 | tp_size=2, |
| 108 | element_size=2, |
| 109 | ) |
| 110 | assert result is not None |
| 111 | |
| 112 | |
| 113 | # XPU-specific constants (must match sequence_parallelism.py values) |
| 114 | _XPU_MIN_HIDDEN_SIZE = 4096 |
| 115 | _XPU_MIN_PER_GPU_SIZE_MB = 8.0 |
| 116 | |
| 117 | |
| 118 | class TestGetSequenceParallelismThresholdXPU: |
| 119 | """Tests for get_sequence_parallelism_threshold on XPU platform.""" |
| 120 | |
| 121 | def test_xpu_small_hidden_size_returns_none(self, mock_xpu_platform): |
| 122 | """XPU with hidden_size below threshold should return None.""" |
| 123 | with mock_xpu_platform(): |
| 124 | result = get_sequence_parallelism_threshold( |
| 125 | hidden_size=_XPU_MIN_HIDDEN_SIZE - 1, |
| 126 | tp_size=2, |
| 127 | element_size=2, |
| 128 | ) |
| 129 | assert result is None |
| 130 | |
| 131 | def test_xpu_large_model_returns_threshold(self, mock_xpu_platform): |
| 132 | """XPU with hidden_size >= threshold should return calculated value.""" |
| 133 | with mock_xpu_platform(): |
| 134 | hidden_size = _XPU_MIN_HIDDEN_SIZE |
| 135 | tp_size = 2 |
| 136 | element_size = 2 |
| 137 | result = get_sequence_parallelism_threshold( |
| 138 | hidden_size=hidden_size, |
| 139 | tp_size=tp_size, |
| 140 | element_size=element_size, |
| 141 | ) |
| 142 | # (8 * 2 * 1024 * 1024) // (4096 * 2) = 2048 |
| 143 | MiB = 1024 * 1024 |
| 144 | expected = int( |
| 145 | (_XPU_MIN_PER_GPU_SIZE_MB * tp_size * MiB) // (hidden_size * element_size) |
| 146 | ) |
| 147 | assert result == expected |
| 148 | assert result == 2048 |
| 149 | |
| 150 | @pytest.mark.parametrize( |
| 151 | "hidden_size,tp_size,element_size,expected", |
| 152 | [ |
| 153 | # (8 * 1 * 1024 * 1024) // (4096 * 2) = 1024 |
| 154 | (4096, 1, 2, 1024), |
| 155 | # (8 * 4 * 1024 * 1024) // (4096 * 2) = 4096 |
| 156 | (4096, 4, 2, 4096), |
| 157 | # (8 * 2 * 1024 * 1024) // (8192 * 2) = 1024 |
| 158 | (8192, 2, 2, 1024), |
| 159 | # (8 * 2 * 1024 * 1024) // (4096 * 4) = 1024 |
| 160 | (4096, 2, 4, 1024), |
| 161 | ], |
| 162 | ) |
| 163 | def test_xpu_threshold_calculation_variations( |
| 164 | self, mock_xpu_platform, hidden_size, tp_size, element_size, expected |
| 165 | ): |
| 166 | """Test XPU threshold calculation with various parameter combinations.""" |
| 167 | with mock_xpu_platform(): |
| 168 | result = get_sequence_parallelism_threshold( |
| 169 | hidden_size=hidden_size, |
| 170 | tp_size=tp_size, |
| 171 | element_size=element_size, |
| 172 | ) |
| 173 | assert result == expected |
| 174 | |
| 175 | def test_xpu_hidden_size_boundary(self, mock_xpu_platform): |
| 176 | """Test behavior at the exact XPU hidden_size boundary.""" |
| 177 | with mock_xpu_platform(): |
| 178 | # Just below threshold |
| 179 | result = get_sequence_parallelism_threshold( |
| 180 | hidden_size=_XPU_MIN_HIDDEN_SIZE - 1, |
| 181 | tp_size=2, |
| 182 | element_size=2, |
| 183 | ) |
| 184 | assert result is None |
| 185 | |
| 186 | # Exactly at threshold |
| 187 | result = get_sequence_parallelism_threshold( |
| 188 | hidden_size=_XPU_MIN_HIDDEN_SIZE, |
| 189 | tp_size=2, |
| 190 | element_size=2, |
| 191 | ) |
| 192 | assert result is not None |