About LoRA loading at inference time (关于几何生成的推理LoRA加载)

[liuyuxuan3060]

我在训练hy3dshape/configs/hunyuandit-finetuning-flowmatching-dinol518-bf16-lr1e5-4096.yaml中，加了
model:
target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
params:
ckpt_path: "/mnt/storage7t/wulong/3d-gen/Hunyuan3D-2.1-main/resources/Hunyuan3D-2.1/hunyuan3d-dit-v2-1/model.fp16.ckpt"
first_stage_key: "surface"
cond_stage_key: "image"
scale_by_std: false
z_scale_factor: &z_scale_factor 1.0039506158752403
torch_compile: false

lora_config:
  rank: 16
  target_modules:
    - to_q
    - to_k
    - to_v
    - out_proj
    - proj
    - net.2

成功训练了LoRA模型，保存了 ckpt-step=00050000.ckpt-output/pytorch_model.bin，请问怎么在推理代码中，加入LoRA的加载

[linweijiang]

可以参考 minimal_demo_with_ckpt.py 这个demo：