fpvla-fsdp-stage2-828-32k

3CTeam/fpvla-fsdp-stage2-828-32k

FP-VLA (Vision-Language-Action) model with trace prediction capability, fine-tuned from openvla/openvla-7b.

Model Architecture

Linear Trace Head

• Direct regression via nn.Linear for trajectory prediction
• Deterministic inference (no sampling noise)
• Input: 66 learnable query tokens (conditioned on image + past trajectory)
• Output: 33-point normalized trajectory [33, 2]

Training Info

• Checkpoint step: 32000

Usage

from prismatic.extern.hf.configuration_prismatic import OpenVLAConfig
from prismatic.extern.hf.modeling_prismatic import OpenVLAForActionPrediction

# Load config + model (constructor + manual weight loading)
config = OpenVLAConfig.from_pretrained("YOUR_REPO_ID")
model = OpenVLAForActionPrediction(config)

import glob
from safetensors.torch import load_file
state_dict = {}
for shard in sorted(glob.glob("YOUR_REPO_ID/*.safetensors")):
    state_dict.update(load_file(shard))
model.load_state_dict(state_dict, strict=False)

Evaluation

python vla_scripts/eval_trace_mix.py \
    --model_path YOUR_REPO_ID \
    --base_vla_path openvla/openvla-7b \
    --data_root_dir data \
    --dataset_name trace_mix \
    --output_dir runs/eval \
    --max_samples 64

License

Apache-2.0