chenglongy/glassvla-4b-sam2-lora-percent10-30k-sigma-12-sft

SpatialVLA Merged Model

This model is a merged version of:

• Base Model: /remote-home/share/chenglong/Workplace/SpatialVLA/ckpts_pretrained/spatialvla-4b-224-sft-fractal
• LoRA Adapter: /remote-home/share/chenglong/Workplace/SpatialVLA/outputs/spatialvla_4b_finetune/2025-11-19/06-04-15_glasses_sigma12_dataset_spatialvla-4b-224-sft-fractal_lr5e-6_bs16_node1_gpu4_r64_a64_ep10_linear+emb+h/checkpoint-31740

Merge Details

• LoRA Rank (r): 64
• LoRA Alpha: 64
• Target Modules: linear, position_embedding_head.3, q_proj, lm_head, up_proj, k_proj, o_proj, fc2, position_embedding_head.0, out_proj, fc1, down_proj, spatial_embed_tokens, v_proj, gate_proj
• Merge Date: <function get_file_binaries_from_pathnames at 0x7fe08977f640>

Usage

This merged model can be used directly without PEFT:

import torch
from transformers import AutoModel, AutoProcessor

model_path = "/remote-home/share/chenglong/Workplace/SpatialVLA/ckpts_merged/glassvla-4b-sam2-lora-percent10-30k-sigma-12-sft"
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_path, 
    trust_remote_code=True, 
    torch_dtype=torch.bfloat16
).eval().cuda()

# Use the model for inference
# ... your inference code here ...

Notes

• This is a fully merged model, so the LoRA adapter is no longer needed.
• The model can be used just like the original base model.
• All weights have been merged into a single set of parameters.