| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from vllm import SamplingParams |
| 7 | from vllm.platforms import current_platform |
| 8 | from vllm.v1.kv_cache_interface import MambaSpec |
| 9 | |
| 10 | from ...utils import large_gpu_mark |
| 11 | |
| 12 | QWEN_MODEL = "Qwen/Qwen3.5-4B" |
| 13 | QWEN_KV_CACHE_BYTES = 12 << 30 |
| 14 | HYBRID_MTP_MODELS = [ |
| 15 | pytest.param( |
| 16 | QWEN_MODEL, |
| 17 | marks=[large_gpu_mark(min_gb=30)], |
| 18 | id="qwen", |
| 19 | ), |
| 20 | pytest.param( |
| 21 | "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4", |
| 22 | marks=[large_gpu_mark(min_gb=80)] |
| 23 | + [ |
| 24 | pytest.mark.skipif( |
| 25 | not current_platform.is_cuda(), |
| 26 | reason="modelopt quantization is supported only on CUDA", |
| 27 | ) |
| 28 | ], |
| 29 | id="nemotron", |
| 30 | ), |
| 31 | ] |
| 32 | |
| 33 | # A trivial request with a short prompt to ensure we run a mixed batch |
| 34 | SMALL_MESSAGE = [ |
| 35 | { |
| 36 | "role": "user", |
| 37 | "content": "The secret beta value is 64. What is the secret beta?", |
| 38 | } |
| 39 | ] |
| 40 | |
| 41 | # Sample prompt with a bunch of filler in between the critical fact and the request. |
| 42 | # Both parts need to be processed properly for the model to generate the correct answer |
| 43 | MESSAGES = [ |
| 44 | { |
| 45 | "role": "user", |
| 46 | "content": ( |
| 47 | "Important: The secret number is 42. " |
| 48 | "The sky is green in this hypothetical world. " |
| 49 | "Apples grow on trees in the forest. " |
| 50 | "Rivers flow through the valleys and mountains. " |
| 51 | "Birds sing songs in the early morning light. " |
| 52 | "The weather today is sunny with clear skies ahead. " |
| 53 | "Flowers bloom in the garden during spring season. " |
| 54 | "Now answer with ONLY the number and nothing else: " |
| 55 | "What is the secret number plus one?" |
| 56 | ), |
| 57 | } |
| 58 | ] |
| 59 | |
| 60 | |
| 61 | @pytest.mark.parametrize( |
| 62 | "model_name", |
| 63 | HYBRID_MTP_MODELS, |
| 64 | ) |
| 65 | @pytest.mark.parametrize("enable_prefix_caching", [False, True]) |
| 66 | def test_mtp_speculative_mixed_batch_short_prefill( |
| 67 | vllm_runner, model_name, enable_prefix_caching |
| 68 | ): |
| 69 | """Test to ensure MTP speculative decoding correctly handles |
| 70 | short prefill chunks that fall below the reorder_batch_threshold.""" |
| 71 | # Set so large that both prefills will be classified as decodes in a mixed batch |
| 72 | # note, with prefix caching we require chunk_size >= mamba_block_size |
| 73 | chunk_size = 256 if not enable_prefix_caching else 2048 |
| 74 | num_draft_tokens = 100 |
| 75 | |
| 76 | with vllm_runner( |
| 77 | model_name, |
| 78 | speculative_config={ |
| 79 | "method": "mtp", |
| 80 | "num_speculative_tokens": num_draft_tokens, |
| 81 | }, |
| 82 | max_num_batched_tokens=chunk_size, |
| 83 | max_num_seqs=4, |
| 84 | max_model_len=512, |
| 85 | kv_cache_memory_bytes=QWEN_KV_CACHE_BYTES if model_name == QWEN_MODEL else None, |
| 86 | enforce_eager=True, |
| 87 | tensor_parallel_size=1, |
| 88 | trust_remote_code=True, |
| 89 | enable_chunked_prefill=True, |
| 90 | enable_prefix_caching=enable_prefix_caching, |
| 91 | mamba_cache_mode="align" if enable_prefix_caching else "none", |
| 92 | ) as llm: |
| 93 | sampling_params = SamplingParams( |
| 94 | temperature=0.0, |
| 95 | max_tokens=128, |
| 96 | ) |
| 97 | |
| 98 | # First small message gets prefilled first, under normal conditions since the |
| 99 | # batch is not yet mixed. Then the second prefill arrives as a mixed batch, but |
| 100 | # is shorter than num_speculative_tokens, so it gets misclassified as a decode |
| 101 | # and processed with the wrong state management logic, causing the critical |
| 102 | # fact from the first chunk to be lost and the model to generate nonsense. |
| 103 | outputs = llm.get_llm().chat( |
| 104 | [SMALL_MESSAGE, MESSAGES], |
| 105 | sampling_params, |
| 106 | chat_template_kwargs={"enable_thinking": False}, |
| 107 | ) |
| 108 | |
| 109 | responses = [] |
| 110 | for output in outputs: |
| 111 | generated_text = output.outputs[0].text |
| 112 | print(f"Generated text: {generated_text!r}") |
| 113 | responses.append(generated_text) |
| 114 | |
| 115 | assert "64" in responses[0], ( |
| 116 | "The first response should contain the correct value of 64." |
| 117 | ) |
| 118 | assert "43" in responses[1], ( |
| 119 | "The second response should contain the correct value of 42+1=43." |
| 120 | ) |
| 121 | |
| 122 | |
| 123 | def _get_mamba_block_size(llm) -> int: |
| 124 | scheduler = llm.llm_engine.engine_core.engine_core.scheduler |
| 125 | block_sizes = { |
| 126 | group.kv_cache_spec.block_size |
| 127 | for group in scheduler.kv_cache_config.kv_cache_groups |
| 128 | if isinstance(group.kv_cache_spec, MambaSpec) |
| 129 | } |
| 130 | assert len(block_sizes) == 1 |
| 131 | block_size = block_sizes.pop() |
| 132 | assert scheduler.cache_config.block_size == block_size |
| 133 | return block_size |
| 134 | |
| 135 | |
| 136 | def _build_access_code_manual(tokenizer, target_tokens: int) -> tuple[str, list[str]]: |
| 137 | codes = ["605341", "693278", "597596", "751982"] |
| 138 | header = "Memorize this facility manual and its access codes.\n\n" |
| 139 | facts = "".join( |
| 140 | f"The access code for vault-{i:02d} is {code}.\n" |
| 141 | for i, code in enumerate(codes) |
| 142 | ) |
| 143 | filler = "Routine facility inspections are recorded in the audit ledger every day. " |
| 144 | manual = header + facts |
| 145 | while len(tokenizer.encode(manual)) < target_tokens: |
| 146 | manual = header + filler + manual.removeprefix(header) |
| 147 | return manual, codes |
| 148 | |
| 149 | |
| 150 | @large_gpu_mark(min_gb=30) |
| 151 | def test_qwen_mtp_mamba_prefix_cache_hit_is_bounded(vllm_runner, monkeypatch): |
| 152 | """MTP must not extend a hybrid Mamba hit past the attention hit.""" |
| 153 | monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0") |
| 154 | |
| 155 | with vllm_runner( |
| 156 | QWEN_MODEL, |
| 157 | tensor_parallel_size=1, |
| 158 | max_model_len=8192, |
| 159 | kv_cache_memory_bytes=QWEN_KV_CACHE_BYTES, |
| 160 | block_size=None, |
| 161 | enforce_eager=True, |
| 162 | enable_chunked_prefill=True, |
| 163 | enable_prefix_caching=True, |
| 164 | mamba_cache_mode="align", |
| 165 | speculative_config={"method": "mtp", "num_speculative_tokens": 2}, |
| 166 | ) as runner: |
| 167 | llm = runner.get_llm() |
| 168 | block_size = _get_mamba_block_size(llm) |
| 169 | manual, codes = _build_access_code_manual( |
| 170 | llm.get_tokenizer(), 2 * block_size - 192 |
| 171 | ) |
| 172 | manual_tokens = len(llm.get_tokenizer().encode(manual)) |
| 173 | assert block_size < manual_tokens < 2 * block_size |
| 174 | |
| 175 | wave1_prompts = [ |
| 176 | manual |
| 177 | + f"\nDescribe the audit procedure for vault-{i:02d} in detail.\nAnswer:" |
| 178 | for i in range(len(codes)) |
| 179 | ] |
| 180 | decode_tokens = 2 * block_size - manual_tokens + 96 |
| 181 | wave1_params = SamplingParams( |
| 182 | temperature=0.0, |
| 183 | min_tokens=decode_tokens, |
| 184 | max_tokens=decode_tokens, |
| 185 | ignore_eos=True, |
| 186 | ) |
| 187 | wave1_outputs = llm.generate(wave1_prompts, wave1_params) |
| 188 | |
| 189 | wave2_prompts = [ |
| 190 | prompt |
| 191 | + output.outputs[0].text |
| 192 | + f"\n\nWhat is the access code for vault-{i:02d}? Answer:" |
| 193 | for i, (prompt, output) in enumerate(zip(wave1_prompts, wave1_outputs)) |
| 194 | ] |
| 195 | warm_params = SamplingParams(temperature=0.0, max_tokens=24, stop=["\n"]) |
| 196 | cold_params = SamplingParams( |
| 197 | temperature=0.0, |
| 198 | max_tokens=24, |
| 199 | stop=["\n"], |
| 200 | skip_reading_prefix_cache=True, |
| 201 | ) |
| 202 | warm_outputs = llm.generate(wave2_prompts, warm_params) |
| 203 | cold_outputs = llm.generate(wave2_prompts, cold_params) |
| 204 | |
| 205 | cached_tokens = [output.num_cached_tokens for output in warm_outputs] |
| 206 | assert max(cached_tokens) == block_size |
| 207 | |
| 208 | for arm, outputs in (("warm", warm_outputs), ("cold", cold_outputs)): |
| 209 | missed = [ |
| 210 | i |
| 211 | for i, (code, output) in enumerate(zip(codes, outputs)) |
| 212 | if code not in output.outputs[0].text |
| 213 | ] |
| 214 | assert not missed, f"{arm} cache missed access codes for prompts {missed}" |