"""Decision Index native-engine adapter using the released Transformers readout.""" from io import BytesIO from pathlib import Path from PIL import Image from decision_index.engines.base import Engine, Unsupported as KitUnsupported from decision_index.vision.images import load from .model import SingVLS1, Unsupported, PUBLIC_IMAGE_MAX_PIXELS class D3Engine(Engine): name = "sing-vl-s1" latency = "One native Transformers prediction including input preparation; model loading excluded." def __init__(self, model, device="cuda:0", image_root=None, image_max_pixels=PUBLIC_IMAGE_MAX_PIXELS, **options): super().__init__(**options) self.image_root = image_root self.native = SingVLS1(model, device=device, image_max_pixels=image_max_pixels) self.provenance = { "kind": "native-transformers", "model": str(Path(model).resolve()), "max_pixels": image_max_pixels, "image_resize": "native Sing VL S1 processor: preserve aspect ratio, patch-grid rounding, RGB bicubic resize, 1.6 MP cap", "attention": "prefix-bidirectional full attention, causal GDN, no KV cache", "max_images": 4, "text_max_tokens": 65536, "vision_max_tokens": 10000, } def __call__(self, state, questions, images=(), prompt=None): decoded = [] for ref in images: with Image.open(BytesIO(load(ref, self.image_root))) as image: decoded.append(image.copy()) try: response, details = self.native.predict(state, questions, decoded) except Unsupported as exc: raise KitUnsupported(str(exc)) from exc grids = next(iter(details.values())).get("image_grid_thw") factor = self.native.image_factor raw = {"details": details, "prompt_metadata": prompt, "image_pixels": [int(t*h*w*factor*factor/self.native.merge_size**2) for t,h,w in grids or []], "image_tokens": next(iter(details.values())).get("image_tokens", 0)} return response, raw def runtime(self): return self.native.runtime() def close(self): del self.native