This repository hosts fine-tuned weights for FastVLA, an accelerated Vision-Language-Action (VLA) architecture pairing the authentic OpenVLA-7B backbone with a 2Γ2 Spatial Token Compressor and an O(1) Continuous Flow-Matching Action Head. Fine-tuned on the Arabic Push-T robotics dataset (hamzabouajila/ar-pusht-image), this model enables sub-200ms robotic manipulation policies on consumer/edge GPUs (NVIDIA Tesla T4).
[ Camera Observation (224x224 RGB) ]
β
DinoV2 + SigLIP (OpenVLA)
β
256 Vision Patches
β
βββββββββββββββββββββββββΌβββββββββββββββββββββββ
β FastVLA 2Γ2 Spatial Token Compressor β
β (Patch Pooling 2x2: 256 -> 64 Patches) β
β VRAM: -45.3% | Throughput: 2.05x speedup β
βββββββββββββββββββββββββ¬βββββββββββββββββββββββ
β
βββββββββββββββββββββββββΌβββββββββββββββββββββββ
β LLaMA-2 7B Backbone (OpenVLA-7B) β
β - 32 Layers, Hidden Dim 4096 β
β - 4-Bit NF4 Quantization (bitsandbytes) β
β - LoRA Fine-Tuned (r=16, alpha=32) β
β - Arabic Instruction Conditioning β
βββββββββββββββββββββββββ¬βββββββββββββββββββββββ
β
βββββββββββββββββββββββββΌβββββββββββββββββββββββ
β Continuous Flow Matching Head β
β - 16-Step Action Trajectory Chunk β
β - 32 Continuous DoF Regression β
β - ODE Euler Integration (10 Steps) β
βββββββββββββββββββββββββ¬βββββββββββββββββββββββ
β
[ Predicted Action Trajectory (16x2) ]
All metrics below are verified through empirical execution on NVIDIA Tesla T4 hardware via Modal.
| Metric | Value |
|---|---|
| Model Backbone | OpenVLA-7B (LLaMA-2 7B, 32 Layers, 4096 Hidden) |
| Precision | 4-bit NF4 Quantization + LoRA ($r=16, \alpha=32$) |
| Hardware | NVIDIA Tesla T4 (16GB VRAM, Turing sm_75) |
| Training Steps | 1,500 optimizer updates (12,000 samples $\approx 4$ epochs) |
| Peak Working VRAM | 7.19 GB / 14.57 GB (49.4% headroom remaining) |
| Update Latency | 2,545 ms / step |
| Total Training Time | 55.0 minutes (0.92 hours) |
| Compute Cost | $0.24 ($0.26/hr on Modal T4) |
| Flow Matching Loss | 1.0625 $\to$ 0.9136 |
PushT-v0 Gym)Evaluated across 10 independent episodes with receding-horizon action chunk execution (replan_interval=8, max 300 steps/ep):
| Metric | Result | Description |
|---|---|---|
| Peak Single-Episode Coverage | 46.8% | Best demonstration of block alignment |
| Mean Peak Coverage | 9.0% | Average across 10 evaluation trials |
| Mean Final Coverage | 0.0% | Coverage at terminal step (receding horizon) |
| Mean Physical Jerk | 253.6 | Smoothness metric for continuous action trajectory |
| Target Success Rate ($\ge 95\%$) | 0/10 | Baseline checkpoint at 1.5k steps (4 epochs) |
Note on Convergence: In imitation learning for Push-T, achieving >80% closed-loop success typically requires 10,000 to 20,000 optimization steps. This 1,500-step checkpoint serves as the foundational open-source baseline proving the full OpenVLA-7B pipeline runs end-to-end on budget T4 hardware within 7.2 GB VRAM.
| File | Size | Description |
|---|---|---|
action_head.pt |
6.4 MB | Continuous Flow-Matching ODE action head (512 hidden, 16x2 DoF) |
compressor.pt |
37.8 MB | FastVLA 2x2 Spatial Token Compressor (4608 $\to$ 4096) |
adapter_model.safetensors |
33.6 MB | LoRA weights across all 32 transformer layers of OpenVLA-7B |
adapter_config.json |
1.1 KB | PEFT LoRA configuration ($r=16, \alpha=32$) |
config.json |
780 B | Full FastVLA-7B model architecture manifest |
eval_results.json |
2.1 KB | Raw JSON telemetry from 10 closed-loop Gym evaluation episodes |
import torch
from transformers import AutoTokenizer, AutoModelForVision2Seq
from peft import PeftModel
# 1. Load Base OpenVLA-7B in 4-bit NF4
base_model = AutoModelForVision2Seq.from_pretrained(
"openvla/openvla-7b",
torch_dtype=torch.bfloat16,
load_in_4bit=True,
device_map="auto"
)
# 2. Attach FastVLA LoRA Adapter
model = PeftModel.from_pretrained(base_model, "hamzabouajila/openvla-pusht-arabic")
# 3. Load Spatial Compressor & Action Head
compressor = torch.load("compressor.pt", map_location="cuda")
action_head = torch.load("action_head.pt", map_location="cuda")
@misc{fastvla2026,
author = {Hamza Bouajila},
title = {FastVLA: Accelerated OpenVLA-7B with Spatial Token Compression and Flow Matching},
year = {2026},
publisher = {Hugging Face},
howpublished = {\\url{https://huggingface.co/hamzabouajila/openvla-pusht-arabic}}
}
compressor.pt 38 MBadapter_model.safetensors 34 MBaction_head.pt 6 MB