| """Configuration for the SPEV multimodal decision model.""" |
|
|
| import math |
| from typing import ClassVar |
|
|
| from transformers import PretrainedConfig, Qwen3OmniMoeThinkerConfig |
|
|
|
|
| class SPEVConfig(PretrainedConfig): |
| model_type = "spev" |
| sub_configs: ClassVar[dict] = {"thinker_config": Qwen3OmniMoeThinkerConfig} |
|
|
| def __init__( |
| self, |
| thinker_config=None, |
| adapter_rank=32, |
| adapter_alpha=64, |
| head_width=256, |
| head_heads=8, |
| head_ff=1024, |
| head_depth=2, |
| head_architecture="isolated_evidence_v1", |
| temperature=1.0, |
| max_tokens=8192, |
| **kwargs, |
| ): |
| if isinstance(thinker_config, dict): |
| thinker_config = Qwen3OmniMoeThinkerConfig(**thinker_config) |
| self.thinker_config = thinker_config or Qwen3OmniMoeThinkerConfig() |
| for name, value in ( |
| ("adapter_rank", adapter_rank), |
| ("head_width", head_width), |
| ("head_heads", head_heads), |
| ("head_ff", head_ff), |
| ("head_depth", head_depth), |
| ("max_tokens", max_tokens), |
| ): |
| if not isinstance(value, int) or isinstance(value, bool) or value <= 0: |
| raise ValueError(f"{name} must be a positive integer") |
| if not math.isfinite(adapter_alpha) or adapter_alpha <= 0: |
| raise ValueError("adapter_alpha must be positive and finite") |
| if not math.isfinite(temperature) or temperature <= 0: |
| raise ValueError("temperature must be positive and finite") |
| self.adapter_rank = adapter_rank |
| self.adapter_alpha = adapter_alpha |
| self.head_width = head_width |
| if head_architecture != "isolated_evidence_v1": |
| raise ValueError("Unsupported SPEV decision head architecture") |
| if head_width % head_heads: |
| raise ValueError("head_width must be divisible by head_heads") |
| self.head_heads = head_heads |
| self.head_ff = head_ff |
| self.head_depth = head_depth |
| self.head_architecture = head_architecture |
| self.temperature = temperature |
| self.max_tokens = max_tokens |
| kwargs.setdefault("tie_word_embeddings", False) |
| super().__init__(**kwargs) |
|
|
| def get_text_config(self, decoder=False, encoder=False): |
| return self.thinker_config.text_config |
|
|
|
|
| SPEVConfig.register_for_auto_class() |
|
|
