| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # flake8: noqa |
| 4 | """Tests Quark mxfp4 models against ground truth generation""" |
| 5 | |
| 6 | import pytest |
| 7 | |
| 8 | from vllm import LLM, SamplingParams |
| 9 | |
| 10 | MODELS = ["amd/Llama-2-7b-chat-hf-wmxfp4-amxfp4-kvfp8-scale-uint8"] |
| 11 | |
| 12 | EXPECTED_STRS_MAP = { |
| 13 | "amd/Llama-2-7b-chat-hf-wmxfp4-amxfp4-kvfp8-scale-uint8": [ |
| 14 | "\n### Key Features\n\n* **High-throughput Inference**: vLL", |
| 15 | "\nArtificial intelligence (AI) has evolved significantly since its inception in the 1", |
| 16 | "Artificial intelligence (AI) and human intelligence (HI) are two distinct concepts that have been", |
| 17 | "A neural network is a machine learning model inspired by the structure of the human brain. It consists of", |
| 18 | "\nTitle: The Dreaming Robot\n\nAs the sun set on the bustling metropol", |
| 19 | "\nThe COVID-19 pandemic has had a profound impact on global economic structures and business", |
| 20 | "The Mona Lisa painting, created by Leonardo da Vinci in the early 16th", |
| 21 | " everybody knows this proverbial saying, but did you know that it's not entirely accurate?", |
| 22 | ] |
| 23 | } |
| 24 | |
| 25 | |
| 26 | @pytest.mark.skip(reason="Model to be released in the future") |
| 27 | @pytest.mark.quant_model |
| 28 | @pytest.mark.parametrize("model_name", MODELS) |
| 29 | def test_models(example_prompts, model_name) -> None: |
| 30 | sampling_params = SamplingParams(max_tokens=20, temperature=0) |
| 31 | llm = LLM( |
| 32 | model=model_name, |
| 33 | kv_cache_dtype="fp8", |
| 34 | quantization="quark", |
| 35 | ) |
| 36 | outputs = llm.generate(example_prompts, sampling_params) |
| 37 | for i, output in enumerate(outputs): |
| 38 | output_str = output.outputs[0].text |
| 39 | expected_str = EXPECTED_STRS_MAP[model_name][i] |
| 40 | assert expected_str == output_str, ( |
| 41 | f"Expected: {expected_str!r}\nvLLM: {output_str!r}" |
| 42 | ) |