glm_9b_visual/test_vision_model.py

import torch
from PIL import Image
from transformers import AutoImageProcessor
from transformers.models.glm4v.modeling_glm4v import Glm4vVisionModel, Glm4vVisionConfig
from safetensors.torch import load_file

device = "cuda:0" if torch.cuda.is_available() else "cpu"

image_path = "/home/alic-li/图片/黛玉训狗_红楼之家_饺子和布莱恩的木石前盟.jpg"
vision_model_path = "/mnt/69043a6d-b152-4bd1-be10-e1130af6487f/glm_9b_visual/"

image_processor = AutoImageProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking", trust_remote_code=True)

config = Glm4vVisionConfig.from_pretrained(vision_model_path)
vision_model = Glm4vVisionModel(config).to(device)

state_dict = load_file(f"{vision_model_path}/model.safetensors")
missing_keys, unexpected_keys = vision_model.load_state_dict(state_dict, strict=False)
print("Missing keys:", missing_keys)
print("Unexpected keys:", unexpected_keys)

image = Image.open(image_path).convert("RGB")
inputs = image_processor(images=image, return_tensors="pt").to(device)

pixel_values = inputs["pixel_values"]
grid_thw = inputs["image_grid_thw"]  # 这个才是正确的 patch 分布！

with torch.no_grad():
    outputs = vision_model(
        hidden_states=pixel_values,
        grid_thw=grid_thw
    )

print("Image embedding shape:", outputs.shape)
print("First 5 features:", outputs[:, :5].cpu().numpy())
print("All features:", outputs.cpu().numpy())