| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | from transformers import Qwen3Config |
| 6 | from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig |
| 7 | |
| 8 | |
| 9 | class PixelShuffleSiglip2VisionConfig(Siglip2VisionConfig): |
| 10 | """Vision configuration for Isaac with Pixel Shuffle support. |
| 11 | |
| 12 | Extends Siglip2VisionConfig with additional fields for pixel shuffle. |
| 13 | """ |
| 14 | |
| 15 | model_type = "pixel_shuffle_siglip2" |
| 16 | base_config_key = "vision_config" |
| 17 | |
| 18 | def __init__( |
| 19 | self, |
| 20 | pixel_shuffle_scale_factor: int = 1, |
| 21 | num_patches: int = 256, |
| 22 | **kwargs, |
| 23 | ): |
| 24 | super().__init__(**kwargs) |
| 25 | |
| 26 | # Add our custom fields |
| 27 | self.pixel_shuffle_scale_factor = pixel_shuffle_scale_factor |
| 28 | self.num_patches = num_patches |
| 29 | |
| 30 | |
| 31 | class IsaacConfig(Qwen3Config): |
| 32 | """Configuration class for Isaac multimodal model.""" |
| 33 | |
| 34 | model_type = "isaac" |
| 35 | sub_configs = { |
| 36 | "vision_config": PixelShuffleSiglip2VisionConfig, |
| 37 | "text_config": Qwen3Config, |
| 38 | } |
| 39 | |
| 40 | def __init__( |
| 41 | self, |
| 42 | text_config=None, |
| 43 | vision_config=None, |
| 44 | vision_patch_size: int = 16, |
| 45 | vision_max_num_patches: int = 256, |
| 46 | vision_min_num_patches: int | None = None, |
| 47 | pixel_shuffle_scale: int = 1, |
| 48 | max_sequence_length: int = 16384, |
| 49 | vision_token: str = "<image>", |
| 50 | vision_attn_implementation: str | None = None, |
| 51 | **kwargs, |
| 52 | ): |
| 53 | if isinstance(text_config, dict): |
| 54 | # from HF config |
| 55 | self.text_config = self.sub_configs["text_config"](**text_config) |
| 56 | elif text_config is None: |
| 57 | # For BC use all kwargs to init text config. |
| 58 | self.text_config = self.sub_configs["text_config"](**kwargs) |
| 59 | else: |
| 60 | # from Qwen3Config |
| 61 | self.text_config = text_config |
| 62 | |
| 63 | # EventStreamProcessor parameters (for backward compatibility) |
| 64 | self.video_patch_size = vision_patch_size |
| 65 | self.vision_max_num_patches = vision_max_num_patches |
| 66 | self.vision_min_num_patches = vision_min_num_patches |
| 67 | self.pixel_shuffle_scale = pixel_shuffle_scale |
| 68 | |
| 69 | # Processing parameters |
| 70 | self.max_sequence_length = max_sequence_length |
| 71 | self.vision_token = vision_token |
| 72 | |
| 73 | # Handle vision config - PixelShuffleSiglip2VisionConfig instance |
| 74 | if isinstance(vision_config, dict): |
| 75 | self.vision_config = PixelShuffleSiglip2VisionConfig(**vision_config) |
| 76 | elif vision_config is None: |
| 77 | self.vision_config = PixelShuffleSiglip2VisionConfig() |
| 78 | else: |
| 79 | self.vision_config = vision_config |
| 80 | |
| 81 | # Ensure compatibility with pretrained checkpoints |
| 82 | self.vision_config.pixel_shuffle_scale_factor = getattr( |
| 83 | self.vision_config, |
| 84 | "pixel_shuffle_scale_factor", |
| 85 | pixel_shuffle_scale, |
| 86 | ) |
| 87 | self.vision_config.num_patches = getattr( |
| 88 | self.vision_config, |
| 89 | "num_patches", |
| 90 | vision_max_num_patches, |
| 91 | ) |
| 92 | self.vision_attn_implementation = vision_attn_implementation |
| 93 | super().__init__(**kwargs) |
| 94 | |
| 95 | |
| 96 | __all__ = [ |
| 97 | "IsaacConfig", |
| 98 | "PixelShuffleSiglip2VisionConfig", |
| 99 | ] |