| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """E2E tests for online MXFP8 quantization. |
| 5 | |
| 6 | Loads a BF16 model with ``--quantization mxfp8`` (online quantization) and |
| 7 | compares log-probabilities against the same model served in BF16 without |
| 8 | quantization. This exercises the full pipeline: config parsing, |
| 9 | ``Mxfp8OnlineLinearMethod``, ``Mxfp8OnlineMoEMethod``, weight loading, |
| 10 | online quantization / shuffling, and inference through ``apply_monolithic``. |
| 11 | |
| 12 | Layer skipping (``modules_to_not_convert``) is configured in the model's |
| 13 | ``config.json`` under ``quantization_config`` and is not tested here. |
| 14 | |
| 15 | ``example_prompts`` is a pytest fixture (from conftest.py) that loads 8 |
| 16 | diverse prompts from ``tests/prompts/example.txt``. |
| 17 | """ |
| 18 | |
| 19 | import pytest |
| 20 | import torch |
| 21 | |
| 22 | from tests.quantization.utils import is_quant_method_supported |
| 23 | from vllm.platforms import current_platform |
| 24 | |
| 25 | from ..utils import check_logprobs_close |
| 26 | |
| 27 | # A small MoE model that fits on a single GPU and has both linear + MoE layers. |
| 28 | MOE_MODEL = "allenai/OLMoE-1B-7B-0125-Instruct" |
| 29 | # A small dense model (no MoE) to validate the linear-only path. |
| 30 | DENSE_MODEL = "Qwen/Qwen3-0.6B" |
| 31 | |
| 32 | MAX_MODEL_LEN = 1024 |
| 33 | MAX_TOKENS = 4 |
| 34 | NUM_LOG_PROBS = 8 |
| 35 | |
| 36 | |
| 37 | @pytest.mark.skipif( |
| 38 | not is_quant_method_supported("mxfp8"), |
| 39 | reason="mxfp8 is not supported on this GPU type (requires sm_100+).", |
| 40 | ) |
| 41 | @pytest.mark.quant_model |
| 42 | @pytest.mark.parametrize("model", [DENSE_MODEL, MOE_MODEL], ids=["dense", "moe"]) |
| 43 | def test_mxfp8_logprobs( |
| 44 | vllm_runner, |
| 45 | quant_baseline_logprobs, |
| 46 | example_prompts, |
| 47 | model: str, |
| 48 | monkeypatch: pytest.MonkeyPatch, |
| 49 | ) -> None: |
| 50 | """Compare BF16 baseline logprobs against online MXFP8-quantized model. |
| 51 | |
| 52 | Runs the same model twice -- once in BF16 (baseline) and once with |
| 53 | online MXFP8 quantization -- then checks that the top log-probabilities |
| 54 | are close. Only 4 tokens are generated to keep the test fast while |
| 55 | still catching numerical divergence. |
| 56 | """ |
| 57 | with monkeypatch.context() as m: |
| 58 | m.setenv("TOKENIZERS_PARALLELISM", "true") |
| 59 | |
| 60 | baseline_outputs = quant_baseline_logprobs( |
| 61 | model, |
| 62 | example_prompts, |
| 63 | max_model_len=MAX_MODEL_LEN, |
| 64 | max_tokens=MAX_TOKENS, |
| 65 | num_logprobs=NUM_LOG_PROBS, |
| 66 | ) |
| 67 | |
| 68 | with vllm_runner( |
| 69 | model, |
| 70 | max_model_len=MAX_MODEL_LEN, |
| 71 | enforce_eager=True, |
| 72 | quantization="mxfp8", |
| 73 | ) as vllm_model: |
| 74 | test_outputs = vllm_model.generate_greedy_logprobs( |
| 75 | example_prompts, MAX_TOKENS, NUM_LOG_PROBS |
| 76 | ) |
| 77 | |
| 78 | check_logprobs_close( |
| 79 | outputs_0_lst=baseline_outputs, |
| 80 | outputs_1_lst=test_outputs, |
| 81 | name_0="bf16", |
| 82 | name_1="mxfp8", |
| 83 | ) |
| 84 | |
| 85 | |
| 86 | @pytest.mark.skipif( |
| 87 | not is_quant_method_supported("mxfp8"), |
| 88 | reason="mxfp8 is not supported on this GPU type (requires sm_100+).", |
| 89 | ) |
| 90 | @pytest.mark.skipif( |
| 91 | not current_platform.is_rocm(), |
| 92 | reason="AITER MXFP8 MoE backend is ROCm-only.", |
| 93 | ) |
| 94 | @pytest.mark.quant_model |
| 95 | def test_mxfp8_aiter_requires_swigluoai_activation( |
| 96 | monkeypatch: pytest.MonkeyPatch, |
| 97 | ) -> None: |
| 98 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 99 | from vllm.model_executor.layers.fused_moe.config import ( |
| 100 | FusedMoEConfig, |
| 101 | FusedMoEParallelConfig, |
| 102 | RoutingMethodType, |
| 103 | ) |
| 104 | from vllm.model_executor.layers.fused_moe.experts import aiter_mxfp8_moe |
| 105 | from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import ( |
| 106 | select_mxfp8_moe_backend, |
| 107 | ) |
| 108 | |
| 109 | monkeypatch.setattr( |
| 110 | aiter_mxfp8_moe.AiterMxfp8Experts, |
| 111 | "_supports_current_device", |
| 112 | staticmethod(lambda: True), |
| 113 | ) |
| 114 | |
| 115 | config = FusedMoEConfig( |
| 116 | num_experts=8, |
| 117 | experts_per_token=2, |
| 118 | hidden_dim=256, |
| 119 | intermediate_size=256, |
| 120 | num_local_experts=8, |
| 121 | num_logical_experts=8, |
| 122 | moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(), |
| 123 | activation=MoEActivation.SILU, |
| 124 | in_dtype=torch.bfloat16, |
| 125 | device="cuda", |
| 126 | routing_method=RoutingMethodType.Renormalize, |
| 127 | moe_backend="aiter", |
| 128 | ) |
| 129 | |
| 130 | with pytest.raises(ValueError, match="requires activation=swigluoai_uninterleave"): |
| 131 | select_mxfp8_moe_backend(config) |
| 132 | |
| 133 | |
| 134 | @pytest.mark.skipif( |
| 135 | not is_quant_method_supported("mxfp8"), |
| 136 | reason="mxfp8 is not supported on this GPU type (requires sm_100+).", |
| 137 | ) |
| 138 | @pytest.mark.skipif( |
| 139 | not current_platform.is_rocm(), |
| 140 | reason="AITER MXFP8 MoE backend is ROCm-only.", |
| 141 | ) |
| 142 | @pytest.mark.quant_model |
| 143 | def test_mxfp8_aiter_requires_swigluoai_params( |
| 144 | monkeypatch: pytest.MonkeyPatch, |
| 145 | ) -> None: |
| 146 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 147 | from vllm.model_executor.layers.fused_moe.config import ( |
| 148 | FusedMoEConfig, |
| 149 | FusedMoEParallelConfig, |
| 150 | RoutingMethodType, |
| 151 | ) |
| 152 | from vllm.model_executor.layers.fused_moe.experts import aiter_mxfp8_moe |
| 153 | from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import ( |
| 154 | select_mxfp8_moe_backend, |
| 155 | ) |
| 156 | |
| 157 | monkeypatch.setattr( |
| 158 | aiter_mxfp8_moe.AiterMxfp8Experts, |
| 159 | "_supports_current_device", |
| 160 | staticmethod(lambda: True), |
| 161 | ) |
| 162 | |
| 163 | config = FusedMoEConfig( |
| 164 | num_experts=8, |
| 165 | experts_per_token=2, |
| 166 | hidden_dim=256, |
| 167 | intermediate_size=256, |
| 168 | num_local_experts=8, |
| 169 | num_logical_experts=8, |
| 170 | moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(), |
| 171 | activation=MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 172 | in_dtype=torch.bfloat16, |
| 173 | device="cuda", |
| 174 | routing_method=RoutingMethodType.Renormalize, |
| 175 | moe_backend="aiter", |
| 176 | ) |
| 177 | |
| 178 | with pytest.raises(ValueError, match="hardcodes SwiGLU-OAI"): |
| 179 | select_mxfp8_moe_backend(config) |
| 180 | |
| 181 | |
| 182 | @pytest.mark.skipif( |
| 183 | not is_quant_method_supported("mxfp8"), |
| 184 | reason="mxfp8 is not supported on this GPU type (requires sm_100+).", |
| 185 | ) |
| 186 | @pytest.mark.skipif( |
| 187 | not current_platform.is_rocm(), |
| 188 | reason="AITER MXFP8 MoE backend is ROCm-only.", |
| 189 | ) |
| 190 | @pytest.mark.quant_model |
| 191 | def test_mxfp8_aiter_accepts_swigluoai_params( |
| 192 | monkeypatch: pytest.MonkeyPatch, |
| 193 | ) -> None: |
| 194 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 195 | from vllm.model_executor.layers.fused_moe.config import ( |
| 196 | FusedMoEConfig, |
| 197 | FusedMoEParallelConfig, |
| 198 | RoutingMethodType, |
| 199 | ) |
| 200 | from vllm.model_executor.layers.fused_moe.experts import aiter_mxfp8_moe |
| 201 | from vllm.model_executor.layers.fused_moe.oracle.fp8 import Fp8MoeBackend |
| 202 | from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import ( |
| 203 | select_mxfp8_moe_backend, |
| 204 | ) |
| 205 | |
| 206 | monkeypatch.setattr( |
| 207 | aiter_mxfp8_moe.AiterMxfp8Experts, |
| 208 | "_supports_current_device", |
| 209 | staticmethod(lambda: True), |
| 210 | ) |
| 211 | |
| 212 | config = FusedMoEConfig( |
| 213 | num_experts=8, |
| 214 | experts_per_token=2, |
| 215 | hidden_dim=256, |
| 216 | intermediate_size=256, |
| 217 | num_local_experts=8, |
| 218 | num_logical_experts=8, |
| 219 | moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(), |
| 220 | activation=MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 221 | in_dtype=torch.bfloat16, |
| 222 | device="cuda", |
| 223 | routing_method=RoutingMethodType.Renormalize, |
| 224 | moe_backend="aiter", |
| 225 | swiglu_alpha=aiter_mxfp8_moe._AITER_SWIGLU_ALPHA, |
| 226 | swiglu_beta=aiter_mxfp8_moe._AITER_SWIGLU_BETA, |
| 227 | ) |
| 228 | |
| 229 | backend, experts_cls = select_mxfp8_moe_backend(config) |
| 230 | |
| 231 | assert backend == Fp8MoeBackend.AITER_MXFP8 |
| 232 | assert experts_cls is aiter_mxfp8_moe.AiterMxfp8Experts |
| 233 | |
| 234 | |
| 235 | @pytest.mark.skipif( |
| 236 | not is_quant_method_supported("mxfp8"), |
| 237 | reason="mxfp8 is not supported on this GPU type (requires sm_100+).", |
| 238 | ) |
| 239 | @pytest.mark.quant_model |
| 240 | @pytest.mark.parametrize("model", [DENSE_MODEL, MOE_MODEL], ids=["dense", "moe"]) |
| 241 | def test_mxfp8_generation(vllm_runner, model: str) -> None: |
| 242 | """Smoke test: verify online MXFP8 model generates coherent text.""" |
| 243 | prompt = "1 2 3 4 5" |
| 244 | with vllm_runner( |
| 245 | model, |
| 246 | enforce_eager=True, |
| 247 | quantization="mxfp8", |
| 248 | max_model_len=MAX_MODEL_LEN, |
| 249 | ) as vllm_model: |
| 250 | output = vllm_model.generate_greedy([prompt], max_tokens=5) |
| 251 | |
| 252 | generated = output[0][1] |
| 253 | assert len(generated) > len(prompt), ( |
| 254 | f"MXFP8 model produced no new tokens. Output: {generated!r}" |
| 255 | ) |