| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from tests.utils import single_gpu_only |
| 7 | from vllm import SamplingParams |
| 8 | from vllm.config import CompilationConfig |
| 9 | from vllm.platforms import current_platform |
| 10 | |
| 11 | from ..utils import ( |
| 12 | evaluate_llm_for_gsm8k, |
| 13 | get_spec_decode_metric_value, |
| 14 | get_test_prompts, |
| 15 | ) |
| 16 | |
| 17 | |
| 18 | @pytest.fixture |
| 19 | def disable_vllm_compile_cache_on_rocm(request: pytest.FixtureRequest) -> None: |
| 20 | if current_platform.is_rocm(): |
| 21 | request.getfixturevalue("disable_vllm_compile_cache") |
| 22 | |
| 23 | |
| 24 | @pytest.mark.parametrize( |
| 25 | "speculative_config", |
| 26 | [ |
| 27 | { |
| 28 | "method": "ngram", |
| 29 | "prompt_lookup_max": 5, |
| 30 | "prompt_lookup_min": 3, |
| 31 | "num_speculative_tokens": 3, |
| 32 | }, |
| 33 | { |
| 34 | "method": "suffix", |
| 35 | "suffix_decoding_max_spec_factor": 2.0, |
| 36 | }, |
| 37 | ], |
| 38 | ) |
| 39 | @pytest.mark.usefixtures("disable_vllm_compile_cache_on_rocm") |
| 40 | @single_gpu_only |
| 41 | def test_ngram_and_suffix_correctness( |
| 42 | speculative_config: dict, |
| 43 | model_name: str, |
| 44 | vllm_runner, |
| 45 | ): |
| 46 | with vllm_runner( |
| 47 | model_name, |
| 48 | block_size=None, |
| 49 | # Keep LLM defaults; VllmRunner only provides lifecycle cleanup here. |
| 50 | trust_remote_code=False, |
| 51 | enable_chunked_prefill=None, |
| 52 | speculative_config=speculative_config, |
| 53 | max_model_len=4096, |
| 54 | # Preserve LLM's default compilation/cudagraph configuration. Without |
| 55 | # this, VllmRunner injects its reduced test-only capture sizes. |
| 56 | compilation_config=CompilationConfig(), |
| 57 | ) as runner: |
| 58 | evaluate_llm_for_gsm8k(runner.llm) |
| 59 | |
| 60 | |
| 61 | @pytest.mark.parametrize("async_scheduling", [True], ids=["async"]) |
| 62 | @single_gpu_only |
| 63 | def test_ngram_gpu_default_with_async_scheduling( |
| 64 | async_scheduling: bool, |
| 65 | vllm_runner, |
| 66 | ): |
| 67 | """Test ngram_gpu speculative decoding (k=3) correctness with and without |
| 68 | async scheduling, validated via GSM8K accuracy. |
| 69 | Uses Qwen/Qwen3-8B (ref GSM8K accuracy: 87%-92%). |
| 70 | """ |
| 71 | qwen3_model = "Qwen/Qwen3-8B" |
| 72 | with vllm_runner( |
| 73 | qwen3_model, |
| 74 | block_size=None, |
| 75 | trust_remote_code=False, |
| 76 | speculative_config={ |
| 77 | "method": "ngram_gpu", |
| 78 | "prompt_lookup_max": 3, |
| 79 | "prompt_lookup_min": 2, |
| 80 | "num_speculative_tokens": 2, |
| 81 | }, |
| 82 | max_model_len=4096, |
| 83 | async_scheduling=async_scheduling, |
| 84 | enable_chunked_prefill=None, |
| 85 | compilation_config=CompilationConfig(), |
| 86 | ) as spec_runner: |
| 87 | # Assert the resolved async_scheduling config matches what was requested. |
| 88 | assert ( |
| 89 | spec_runner.llm.llm_engine.vllm_config.scheduler_config.async_scheduling |
| 90 | == async_scheduling |
| 91 | ) |
| 92 | evaluate_llm_for_gsm8k(spec_runner.llm, expected_accuracy_threshold=0.8) |
| 93 | |
| 94 | |
| 95 | @single_gpu_only |
| 96 | def test_suffix_decoding_acceptance( |
| 97 | sampling_config: SamplingParams, |
| 98 | model_name: str, |
| 99 | vllm_runner, |
| 100 | ): |
| 101 | """Check that suffix decoding caching takes effect and improves acceptance |
| 102 | lengths and acceptance rates over multiple runs of the same prompts. |
| 103 | """ |
| 104 | test_prompts = get_test_prompts(mm_enabled=False) |
| 105 | |
| 106 | with vllm_runner( |
| 107 | model_name, |
| 108 | block_size=None, |
| 109 | trust_remote_code=False, |
| 110 | speculative_config={ |
| 111 | "method": "suffix", |
| 112 | "suffix_decoding_max_spec_factor": 2.0, |
| 113 | "suffix_decoding_max_cached_requests": 1000, |
| 114 | }, |
| 115 | max_model_len=1024, |
| 116 | disable_log_stats=False, |
| 117 | enable_chunked_prefill=None, |
| 118 | compilation_config=CompilationConfig(), |
| 119 | ) as spec_runner: |
| 120 | # Run several times and check that the accepted tokens increase. |
| 121 | num_draft = [] |
| 122 | num_accept = [] |
| 123 | for _ in range(10): # Run multiple times to warm up the cache. |
| 124 | spec_runner.llm.chat(test_prompts, sampling_config) |
| 125 | # Collect draft and acceptance stats. |
| 126 | metrics = spec_runner.llm.get_metrics() |
| 127 | num_draft.append( |
| 128 | get_spec_decode_metric_value( |
| 129 | metrics, "vllm:spec_decode_num_draft_tokens" |
| 130 | ) |
| 131 | ) |
| 132 | num_accept.append( |
| 133 | get_spec_decode_metric_value( |
| 134 | metrics, "vllm:spec_decode_num_accepted_tokens" |
| 135 | ) |
| 136 | ) |
| 137 | |
| 138 | # Calculate the acceptance rates for the first and last runs. |
| 139 | first_accept_tokens = num_accept[0] |
| 140 | first_draft_tokens = num_draft[0] |
| 141 | assert first_draft_tokens > 0, ( |
| 142 | "Suffix decoder produced no draft tokens on the first run: " |
| 143 | f"accepted={first_accept_tokens}, drafted={first_draft_tokens}" |
| 144 | ) |
| 145 | first_accept_rate = first_accept_tokens / first_draft_tokens |
| 146 | |
| 147 | # Take the diff since the stats are cumulative. |
| 148 | last_accept_tokens = num_accept[-1] - num_accept[-2] |
| 149 | last_draft_tokens = num_draft[-1] - num_draft[-2] |
| 150 | assert last_draft_tokens > 0, ( |
| 151 | "Suffix decoder produced no draft tokens on the last run: " |
| 152 | f"accepted_delta={last_accept_tokens}, drafted_delta={last_draft_tokens}; " |
| 153 | f"cumulative_drafted={num_draft[-2:]}" |
| 154 | ) |
| 155 | last_accept_rate = last_accept_tokens / last_draft_tokens |
| 156 | summary = ( |
| 157 | f"first accepted/drafted={first_accept_tokens}/{first_draft_tokens} " |
| 158 | f"(rate={first_accept_rate:.3f}); last delta accepted/drafted=" |
| 159 | f"{last_accept_tokens}/{last_draft_tokens} (rate={last_accept_rate:.3f})" |
| 160 | ) |
| 161 | |
| 162 | # Expect the acceptance length to improve. |
| 163 | assert first_accept_tokens < last_accept_tokens, ( |
| 164 | f"Expected accepted tokens to increase after cache warmup; {summary}" |
| 165 | ) |
| 166 | |
| 167 | # Expect the acceptance rate to improve. |
| 168 | assert first_accept_rate < last_accept_rate, ( |
| 169 | f"Expected acceptance rate to increase after cache warmup; {summary}" |
| 170 | ) |
| 171 | |
| 172 | # Heuristic: expect at least 80.0% acceptance rate at the end. |
| 173 | assert last_accept_rate > 0.80, f"Expected final acceptance rate > 0.80; {summary}" |