| 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 tests.conftest import VllmRunner |
| 10 | from tests.utils import create_new_process_for_each_test |
| 11 | |
| 12 | pytestmark = pytest.mark.skipif( |
| 13 | importlib.util.find_spec("terratorch") is None, |
| 14 | reason="terratorch unavailable while PyPI has `lightning` quarantined; see #41376", |
| 15 | ) |
| 16 | |
| 17 | |
| 18 | @create_new_process_for_each_test() # Hangs otherwise |
| 19 | @pytest.mark.parametrize( |
| 20 | "model", |
| 21 | [ |
| 22 | "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", |
| 23 | "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-BurnScars", |
| 24 | ], |
| 25 | ) |
| 26 | def test_inference( |
| 27 | vllm_runner: type[VllmRunner], |
| 28 | model: str, |
| 29 | ) -> None: |
| 30 | pixel_values = torch.full((6, 512, 512), 1.0, dtype=torch.float16) |
| 31 | location_coords = torch.full((1, 2), 1.0, dtype=torch.float16) |
| 32 | prompt = dict( |
| 33 | prompt_token_ids=[1], |
| 34 | multi_modal_data={ |
| 35 | "image": { |
| 36 | "pixel_values": pixel_values, |
| 37 | "location_coords": location_coords, |
| 38 | } |
| 39 | }, |
| 40 | ) |
| 41 | |
| 42 | with vllm_runner( |
| 43 | model, |
| 44 | runner="pooling", |
| 45 | dtype="half", |
| 46 | enforce_eager=True, |
| 47 | skip_tokenizer_init=True, |
| 48 | enable_mm_embeds=True, |
| 49 | # Limit the maximum number of sequences to avoid the |
| 50 | # test going OOM during the warmup run |
| 51 | max_num_seqs=32, |
| 52 | default_torch_num_threads=1, |
| 53 | ) as vllm_model: |
| 54 | vllm_output = vllm_model.llm.encode(prompt, pooling_task="plugin") |
| 55 | assert torch.equal( |
| 56 | torch.isnan(vllm_output[0].outputs.data).any(), torch.tensor(False) |
| 57 | ) |