| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import Any |
| 4 | |
| 5 | from transformers import PreTrainedConfig |
| 6 | from transformers.models.gemma4.configuration_gemma4 import Gemma4VisionConfig |
| 7 | |
| 8 | |
| 9 | def _init_text_config(self: PreTrainedConfig, **kwargs: Any) -> None: |
| 10 | PreTrainedConfig.__init__(self, **kwargs) |
| 11 | # DiffusionGemma always uses MoE and K=V sharing for full_attention |
| 12 | # layers. The HF reference removed these config fields entirely. |
| 13 | if getattr(self, "num_experts", None): |
| 14 | self.enable_moe_block = True |
| 15 | self.attention_k_eq_v = True |
| 16 | |
| 17 | |
| 18 | class DiffusionGemmaTextConfig(PreTrainedConfig): |
| 19 | model_type = "diffusion_gemma_text" |
| 20 | |
| 21 | def __init__(self, **kwargs: Any): |
| 22 | _init_text_config(self, **kwargs) |
| 23 | |
| 24 | |
| 25 | class DiffusionGemmaConfig(PreTrainedConfig): |
| 26 | model_type = "diffusion_gemma" |
| 27 | |
| 28 | def __init__( |
| 29 | self, |
| 30 | text_config: dict[str, Any] | None = None, |
| 31 | canvas_length: int = 256, |
| 32 | self_conditioning_size: int | None = None, |
| 33 | **kwargs: Any, |
| 34 | ): |
| 35 | self.text_config = DiffusionGemmaTextConfig(**(text_config or {})) |
| 36 | self.canvas_length = canvas_length |
| 37 | self.self_conditioning_size = self_conditioning_size |
| 38 | vision_config = kwargs.pop("vision_config", None) |
| 39 | if isinstance(vision_config, dict): |
| 40 | self.vision_config = Gemma4VisionConfig(**vision_config) |
| 41 | else: |
| 42 | self.vision_config = vision_config |
| 43 | self.audio_config = None |
| 44 | PreTrainedConfig.__init__(self, **kwargs) |