| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import importlib.util |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | |
| 9 | from ....conftest import VllmRunner |
| 10 | |
| 11 | pytestmark = pytest.mark.skipif( |
| 12 | importlib.util.find_spec("terratorch") is None, |
| 13 | reason="terratorch unavailable while PyPI has `lightning` quarantined; see #41376", |
| 14 | ) |
| 15 | |
| 16 | |
| 17 | def _run_test( |
| 18 | vllm_runner: type[VllmRunner], |
| 19 | model: str, |
| 20 | ) -> None: |
| 21 | prompt = [ |
| 22 | { |
| 23 | # This model deals with no text input |
| 24 | "prompt_token_ids": [1], |
| 25 | "multi_modal_data": { |
| 26 | "image": { |
| 27 | "pixel_values": torch.ones((6, 512, 512), dtype=torch.float16), |
| 28 | "location_coords": torch.ones((1, 2), dtype=torch.float16), |
| 29 | } |
| 30 | }, |
| 31 | } |
| 32 | for _ in range(10) |
| 33 | ] |
| 34 | |
| 35 | with vllm_runner( |
| 36 | model, |
| 37 | runner="pooling", |
| 38 | dtype="half", |
| 39 | enforce_eager=True, |
| 40 | skip_tokenizer_init=True, |
| 41 | enable_mm_embeds=True, |
| 42 | # Limit the maximum number of sequences to avoid the |
| 43 | # test going OOM during the warmup run |
| 44 | max_num_seqs=32, |
| 45 | default_torch_num_threads=1, |
| 46 | ) as vllm_model: |
| 47 | vllm_model.llm.encode(prompt, pooling_task="plugin") |
| 48 | |
| 49 | |
| 50 | MODELS = ["ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11"] |
| 51 | |
| 52 | |
| 53 | @pytest.mark.core_model |
| 54 | @pytest.mark.parametrize("model", MODELS) |
| 55 | def test_models_image( |
| 56 | hf_runner, |
| 57 | vllm_runner, |
| 58 | image_assets, |
| 59 | model: str, |
| 60 | ) -> None: |
| 61 | _run_test( |
| 62 | vllm_runner, |
| 63 | model, |
| 64 | ) |