| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # flake8: noqa |
| 4 | """Tests Model Optimizer nvfp4 models against ground truth generation |
| 5 | Note: these tests will only pass on B200 |
| 6 | """ |
| 7 | |
| 8 | import os |
| 9 | from typing import List |
| 10 | |
| 11 | import pytest |
| 12 | from transformers import AutoTokenizer |
| 13 | |
| 14 | from tests.quantization.utils import is_quant_method_supported |
| 15 | from vllm import LLM, SamplingParams |
| 16 | |
| 17 | from vllm.platforms import current_platform |
| 18 | |
| 19 | os.environ["TOKENIZERS_PARALLELISM"] = "true" |
| 20 | |
| 21 | MAX_MODEL_LEN = 1024 |
| 22 | |
| 23 | MODELS = ["nvidia/Llama-3.3-70B-Instruct-FP4"] |
| 24 | |
| 25 | EXPECTED_STRS_MAP = { |
| 26 | "nvidia/Llama-3.3-70B-Instruct-FP4": [ |
| 27 | "vLLM (Vectorized Large Language Model) is indeed a high-throughput and memory-efficient inference", |
| 28 | "Here are the major milestones in the development of artificial intelligence (AI) from 1950 to ", |
| 29 | "Artificial intelligence (AI) and human intelligence (HI) are two distinct forms of intelligence that process", |
| 30 | "A neural network is a type of machine learning model inspired by the structure and function of the human brain", |
| 31 | "In the heart of a cutting-edge robotics lab, a team of engineers had been working tirelessly to push", |
| 32 | "The COVID-19 pandemic has had a profound impact on global economic structures and future business models, leading", |
| 33 | "The Mona Lisa, painted by Leonardo da Vinci in the early 16th century, is one of", |
| 34 | "Here are the translations:\n\n* Japanese: (Sasuga no tori ga miwa o ts", |
| 35 | ] |
| 36 | } |
| 37 | |
| 38 | |
| 39 | # This test compares against golden strings for exact match since |
| 40 | # there is no baseline implementation to compare against |
| 41 | # and is unstable w.r.t specifics of the fp4 implementation or |
| 42 | # the hardware being run on. |
| 43 | # Disabled to prevent it from breaking the build |
| 44 | @pytest.mark.skip( |
| 45 | reason="Prevent unstable test based on golden strings from breaking the build " |
| 46 | " and test input model being too large and hanging the system." |
| 47 | ) |
| 48 | @pytest.mark.skipif( |
| 49 | not is_quant_method_supported("modelopt_fp4"), |
| 50 | reason="modelopt_fp4 is not supported on this GPU type.", |
| 51 | ) |
| 52 | @pytest.mark.parametrize("model_name", MODELS) |
| 53 | def test_models(example_prompts, model_name) -> None: |
| 54 | llm = LLM( |
| 55 | model=model_name, |
| 56 | max_model_len=MAX_MODEL_LEN, |
| 57 | trust_remote_code=True, |
| 58 | enforce_eager=True, |
| 59 | quantization="modelopt_fp4", |
| 60 | ) |
| 61 | |
| 62 | tokenizer = AutoTokenizer.from_pretrained(model_name) |
| 63 | formatted_prompts = [ |
| 64 | tokenizer.apply_chat_template( |
| 65 | [{"role": "user", "content": prompt}], |
| 66 | tokenize=False, |
| 67 | add_generation_prompt=True, |
| 68 | ) |
| 69 | for prompt in example_prompts |
| 70 | ] |
| 71 | params = SamplingParams(max_tokens=20, temperature=0) |
| 72 | generations: List[str] = [] |
| 73 | # Note: these need to be run 1 at a time due to numerical precision, |
| 74 | # since the expected strs were generated this way. |
| 75 | for prompt in formatted_prompts: |
| 76 | outputs = llm.generate(prompt, params) |
| 77 | generations.append(outputs[0].outputs[0].text) |
| 78 | del llm |
| 79 | |
| 80 | print(model_name, generations) |
| 81 | expected_strs = EXPECTED_STRS_MAP[model_name] |
| 82 | for i in range(len(example_prompts)): |
| 83 | generated_str = generations[i] |
| 84 | expected_str = expected_strs[i] |
| 85 | assert expected_str == generated_str, ( |
| 86 | f"Test{i}:\nExpected: {expected_str!r}\nvLLM: {generated_str!r}" |
| 87 | ) |
| 88 | |
| 89 | |
| 90 | EAGER = [True, False] |
| 91 | |
| 92 | SM_100_NVFP4_BACKENDS = [ |
| 93 | "flashinfer_cutedsl", |
| 94 | "flashinfer_cudnn", |
| 95 | "flashinfer_trtllm", |
| 96 | "flashinfer_cutlass", |
| 97 | "torch", |
| 98 | ] |
| 99 | |
| 100 | |
| 101 | @pytest.mark.parametrize("model", ["nvidia/Llama-3.1-8B-Instruct-NVFP4"]) |
| 102 | @pytest.mark.parametrize("eager", EAGER) |
| 103 | @pytest.mark.parametrize( |
| 104 | "backend", |
| 105 | [ |
| 106 | "emulation", |
| 107 | "flashinfer_cutedsl", |
| 108 | "flashinfer_cudnn", |
| 109 | "flashinfer_trtllm", # the small seq_len ensures trtllm_8x4_layout backend is used |
| 110 | "flashinfer_cutlass", |
| 111 | "torch", |
| 112 | ], |
| 113 | ) |
| 114 | def test_nvfp4(vllm_runner, model, eager, backend): |
| 115 | if backend == "flashinfer_cutedsl" and not ( |
| 116 | current_platform.is_device_capability_family(100) |
| 117 | ): |
| 118 | pytest.skip("The flashinfer_cutedsl backend is only supported on SM10x") |
| 119 | |
| 120 | if ( |
| 121 | not current_platform.has_device_capability(100) |
| 122 | and backend in SM_100_NVFP4_BACKENDS |
| 123 | ): |
| 124 | pytest.skip( |
| 125 | f"The backend {backend} is not supported with current_platform.has_device_capability(100) == False" |
| 126 | ) |
| 127 | |
| 128 | with vllm_runner(model, enforce_eager=eager, linear_backend=backend) as llm: |
| 129 | output = llm.generate_greedy(["1 2 3 4 5"], max_tokens=2) |
| 130 | assert output[0][1] == "1 2 3 4 5 6" |
| 131 | |
| 132 | |
| 133 | @pytest.mark.parametrize( |
| 134 | "model", |
| 135 | [ |
| 136 | "nvidia/Qwen3-30B-A3B-NVFP4", |
| 137 | "RedHatAI/Qwen3-30B-A3B-NVFP4", |
| 138 | ], |
| 139 | ) |
| 140 | @pytest.mark.parametrize("backend", ["emulation"]) |
| 141 | @pytest.mark.skipif( |
| 142 | not current_platform.is_rocm(), |
| 143 | reason="NVFP4 MOE emulation is only useful on AMD Instinct MI3xx", |
| 144 | ) |
| 145 | def test_nvfp4_moe(vllm_runner, model, backend): |
| 146 | with vllm_runner( |
| 147 | model, |
| 148 | moe_backend=backend, |
| 149 | linear_backend=backend, |
| 150 | load_format="dummy", |
| 151 | hf_overrides={"num_hidden_layers": 2}, |
| 152 | ) as llm: |
| 153 | _ = llm.generate_greedy(["1 2 3 4 5"], max_tokens=2) |