| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | import torch |
| 5 | import torch.nn as nn |
| 6 | from transformers import AutoConfig, AutoModel, CLIPImageProcessor, PreTrainedModel |
| 7 | |
| 8 | from vllm.distributed import cleanup_dist_env_and_memory |
| 9 | from vllm.platforms import current_platform |
| 10 | from vllm.transformers_utils.repo_utils import hf_api |
| 11 | from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE |
| 12 | |
| 13 | from ....conftest import ImageTestAssets |
| 14 | |
| 15 | # we use snapshot_download to prevent conflicts between |
| 16 | # dynamic_module and trust_remote_code for hf_runner |
| 17 | DOWNLOAD_PATTERN = ["*.json", "*.py", "*.safetensors", "*.txt", "*.model"] |
| 18 | |
| 19 | DEVICE_TYPE = current_platform.device_type |
| 20 | |
| 21 | |
| 22 | @torch.inference_mode() |
| 23 | def run_intern_vit_test( |
| 24 | image_assets: ImageTestAssets, |
| 25 | model_id: str, |
| 26 | *, |
| 27 | dtype: str, |
| 28 | ): |
| 29 | model = hf_api().snapshot_download(model_id, allow_patterns=DOWNLOAD_PATTERN) |
| 30 | torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype] |
| 31 | |
| 32 | img_processor = CLIPImageProcessor.from_pretrained(model) |
| 33 | images = [asset.pil_image for asset in image_assets] |
| 34 | pixel_values = [ |
| 35 | img_processor(images, return_tensors="pt").pixel_values.to(torch_dtype) |
| 36 | for images in images |
| 37 | ] |
| 38 | |
| 39 | config = AutoConfig.from_pretrained(model, trust_remote_code=True) |
| 40 | if not getattr(config, "norm_type", None): |
| 41 | config.norm_type = "rms_norm" |
| 42 | |
| 43 | # Monkey-patch for transformers v5 compatibility: InternVisionModel's |
| 44 | # custom code doesn't call post_init(), so all_tied_weights_keys is |
| 45 | # never set. Remove once https://github.com/OpenGVLab/InternVL fixes it. |
| 46 | _orig_init = PreTrainedModel.__init__ |
| 47 | |
| 48 | def _patched_init(self, *args, **kwargs): |
| 49 | _orig_init(self, *args, **kwargs) |
| 50 | self.all_tied_weights_keys = {} |
| 51 | |
| 52 | PreTrainedModel.__init__ = _patched_init |
| 53 | |
| 54 | try: |
| 55 | hf_model = AutoModel.from_pretrained( |
| 56 | model, dtype=torch_dtype, trust_remote_code=True |
| 57 | ).to(DEVICE_TYPE) |
| 58 | finally: |
| 59 | PreTrainedModel.__init__ = _orig_init |
| 60 | hf_outputs_per_image = [ |
| 61 | hf_model(pixel_value.to(DEVICE_TYPE)).last_hidden_state |
| 62 | for pixel_value in pixel_values |
| 63 | ] |
| 64 | |
| 65 | from vllm.model_executor.models.intern_vit import InternVisionModel |
| 66 | |
| 67 | vllm_model = InternVisionModel(config) |
| 68 | vllm_model.load_weights(hf_model.state_dict().items()) |
| 69 | |
| 70 | del hf_model |
| 71 | cleanup_dist_env_and_memory() |
| 72 | |
| 73 | vllm_model = vllm_model.to(DEVICE_TYPE, torch_dtype) |
| 74 | vllm_outputs_per_image = [ |
| 75 | vllm_model(pixel_values=pixel_value.to(DEVICE_TYPE)) |
| 76 | for pixel_value in pixel_values |
| 77 | ] |
| 78 | del vllm_model |
| 79 | cleanup_dist_env_and_memory() |
| 80 | |
| 81 | cos_similar = nn.CosineSimilarity(dim=-1) |
| 82 | for vllm_output, hf_output in zip(vllm_outputs_per_image, hf_outputs_per_image): |
| 83 | assert cos_similar(vllm_output, hf_output).mean() > 0.99 |
| 84 | |
| 85 | |
| 86 | @pytest.mark.parametrize( |
| 87 | "model_id", |
| 88 | [ |
| 89 | "OpenGVLab/InternViT-300M-448px", |
| 90 | "OpenGVLab/InternViT-6B-448px-V1-5", |
| 91 | ], |
| 92 | ) |
| 93 | @pytest.mark.parametrize("dtype", ["half"]) |
| 94 | def test_models( |
| 95 | default_vllm_config, dist_init, image_assets, model_id, dtype: str |
| 96 | ) -> None: |
| 97 | run_intern_vit_test( |
| 98 | image_assets, |
| 99 | model_id, |
| 100 | dtype=dtype, |
| 101 | ) |