Tacoin/openpi-pi0.5-libero-onnx
By Tacoin · ↓ 0 · ♥ 0 · Updated 2026-01-19
What's in this repo
This page lists the files in Tacoin/openpi-pi0.5-libero-onnx, including model.fp32.onnx, model.w4a4.onnx, model.w4a4.onnx.temp.onnx and more.
- 10 ONNX models (.onnx files)
- 9978 files, including 1523 .weight, 621 .bias, 10 .onnx, 8 .data, 6 .0_self_attn_q_proj_weight_quantizer_1_constant_attr__value, 6 .0_self_attn_q_proj_weight_quantizer_2_constant_attr__value
- Last updated 2026-01-19
ONNX files
| File | Size | Download |
|---|---|---|
| final_fp16/model.fp32.onnx | — | Download |
| final_w4a4/model.w4a4.onnx | — | Download |
| final_w4a4/model.w4a4.onnx.temp.onnx | — | Download |
| final_w8a16/model.w8a16.onnx | — | Download |
| final_w8a16_new/model.w8a16.onnx | — | Download |
| final_w8a16_new/model.w8a16.onnx.temp.onnx | — | Download |
| fp16/model.onnx | — | Download |
| fp32/model.onnx | — | Download |
| int4/model.onnx | — | Download |
| int8/model.onnx | — | Download |
How to use
These ONNX files can be loaded with ONNX Runtime. To inspect a model graph, upload the file to the Netron viewer. Download links point to the original files on Hugging Face.
README
OpenPi VLA Models (ONNX Export)
This repository contains ONNX-exported versions of the OpenPi VLA (Vision-Language-Action) models, specifically tuned for NVIDIA Jetson Thor (Blackwell) and GeForce RTX Ada/Blackwell GPUs.
These models are derived from the Pi0.5 Libero baseline (OpenVLA architecture).
📊 Performance Benchmarks (Jetson Thor)
| Variant | Latency (ms) | Throughput (QPS) | GPU Mem (Est) | Status | Notes |
|---|---|---|---|---|---|
| W8A16 (QDQ) | 181.81 | 6.37 | ~6.3 GiB | ✅ Recommended | Production Ready. Verified accuracy parity (MSE 0.0061). |
| FP16 | 184.54 | 6.26 | ~13.0 GiB | ✅ Verified | Baseline. Native FP16 execution. |
| W8A8 (Sim) | 128.37 | 8.38 | ~3.7 GiB | 🚧 Benchmark | Fastest (1.4x speedup), requires calibration. |
| INT4 (Sim) | 183.15 | 6.33 | ~6.3 GiB | 🚧 Experimental | Parity speed with W8A16. |
Memory Note: INT8 quantization reduces runtime VRAM usage by approximately 3.2x compared to FP32 (4GB vs 13GB measured on runtime).
📂 Repository Structure & "Why is this 231GB?"
We preserve all precision variants and intermediate export states for research reproducibility.
| Folder | Precision | Description | Size | Contents |
|---|---|---|---|---|
final_w8a16/ |
W8A16 | [USE THIS] Clean, collapsed ONNX model. | ~12GB | model.w8a16.onnx, .data |
final_fp16/ |
FP16 | Baseline high-precision export. | ~24GB | model.fp32.onnx |
final_w4a4/ |
W4A4 | Experimental INT4/INT4 export. | ~12GB | Granular (Many files) |
final_w8a16_new/ |
W8A16 | Uncollapsed dev export (Duplicate state). | ~12GB | Granular (Many files) |
checkpoints/ |
PyTorch | Raw Archive. Original JAX->PyTorch weights. | ~100GB | .pt files, config, raw layers |
Recommendation: Users should download ONLY final_w8a16/ for deployment.
⚠️ Environment & Compatibility Notes
1. NVIDIA Jetson Thor (SM 11.0)
- Driver Requirement: CUDA 13.0+ / Driver 560+
- Issue: Standard
onnxruntimeand PyTorch binaries may not fully support SM 11.0 (Blackwell) instructions yet. - Fix: Use the NVIDIA NGC containers or compile from source.
2. Quantization Support
- W4A4: Requires a custom TensorRT backend to execute
trt.DequantizeLinearops efficiently. CPU execution is not supported. - W8A16: Uses standard ONNX QDQ nodes (
QuantizeLinear/DequantizeLinear). Compatible with standard ONNX Runtime (GPU) and TensorRT.
🛠️ Deployment Code
Python (ONNX Runtime)
import onnxruntime as ort
import os
import tensorrt_libs
# 1. Register TensorRT Plugins (Critical for QDQ models)
os.environ["LD_LIBRARY_PATH"] = os.environ.get("LD_LIBRARY_PATH", "") + ":" + os.path.dirname(tensorrt_libs.__file__)
# 2. Load Model
model_path = "final_w8a16/model.w8a16.onnx"
providers = [
("TensorrtExecutionProvider", {
"device_id": 0,
"trt_fp16_enable": True,
"trt_int8_enable": True, # Enable INT8 Tensor Cores
"trt_engine_cache_enable": True,
}),
"CUDAExecutionProvider"
]
session = ort.InferenceSession(model_path, providers=providers)
print("Model loaded successfully!")
TensorRT CLI (trtexec)
# Compile W8A16 model to Engine
trtexec --onnx=final_w8a16/model.w8a16.onnx --fp16 --int8 --saveEngine=model.engine --verbose
Model info
Created: 2025-12-31
Updated: 2026-01-19
View on Hugging FaceFiles (9978)
Related models
More from Tacoin
No other models by this author.