onw 0.2: renamed from npue; onw command; Qwen3.5 (dense) / Gemma 4 / vision support; LM head segment
49c3379 verified | """Qwen3.5/3.6 vision: numpy patchify vs the HF processor, and the static OV tower vs the HF tower (CPU). | |
| usage: python check_vision.py HF_DIR IMAGE [DEVICE=CPU]""" | |
| import json, os, sys | |
| import numpy as np | |
| import openvino as ov | |
| import torch | |
| from PIL import Image | |
| from transformers import AutoProcessor | |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) | |
| from onw.models import qwen_vision as QV | |
| from onw.models.lfm2_moe import Checkpoint | |
| def main(): | |
| src, img_path = sys.argv[1], sys.argv[2] | |
| dev = sys.argv[3] if len(sys.argv) > 3 else "CPU" | |
| grid = (32, 32) | |
| img = Image.open(img_path).convert("RGB").resize((512, 512), Image.BICUBIC) | |
| vcfg = {k: v for k, v in json.load(open(os.path.join(src, "config.json")))["vision_config"].items() if k != "model_type"} | |
| proc = AutoProcessor.from_pretrained(src) | |
| ref = proc.image_processor(images=[img], return_tensors="pt") | |
| ref_px, thw = ref["pixel_values"].numpy(), ref["image_grid_thw"] | |
| pp = json.load(open(os.path.join(src, "preprocessor_config.json"))) | |
| mine = QV.patchify(img, grid, mean=pp["image_mean"], std=pp["image_std"]) | |
| print("grid", thw.tolist(), "patchify rel err", float(np.linalg.norm(mine - ref_px) / np.linalg.norm(ref_px))) | |
| from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5VisionModel | |
| from transformers.models.qwen3_5.configuration_qwen3_5 import Qwen3_5VisionConfig | |
| hf = Qwen3_5VisionModel._from_config(Qwen3_5VisionConfig(**vcfg)) | |
| W = Checkpoint(src) | |
| sd = {k[len(QV.VP):]: torch.tensor(W[k]) for k in W.where if k.startswith(QV.VP)} | |
| print("load:", hf.load_state_dict(sd, strict=False)) | |
| with torch.no_grad(): | |
| want = hf(torch.tensor(ref_px), thw).pooler_output.float().numpy() | |
| m = QV.build(src, vcfg, W, grid) | |
| got = ov.Core().compile_model(m, dev, {"INFERENCE_PRECISION_HINT": "f32"} if dev == "CPU" else {}).create_infer_request() \ | |
| .infer([mine.astype(np.float16)])[0].astype(np.float32) | |
| print(f"{dev} tower vs HF: shape {got.shape} vs {want.shape}, rel err {np.linalg.norm(got - want) / np.linalg.norm(want):.4f}") | |
| if __name__ == "__main__": | |
| main() | |