Install

The bootstrap installer detects your platform (Mac / Jetson Orin / NVIDIA GPU / CPU) and chooses the right extras automatically. It also fails fast on unsupported hardware (e.g. original 4 GB Jetson Nano).

Recommended (auto)

curl -fsSL https://fastcrest.com/install | sh

GPU production

pip install 'reflex-vla[serve,gpu,monolithic]'

Mac / CPU

pip install 'reflex-vla[serve,onnx,monolithic]'

Core only

pip install reflex-vla

Requires Python ≥ 3.10.

Upgrading

We ship patches frequently. After a major version bump, stale export caches can prevent the server from starting:

pip install --upgrade reflex-vla              # pip
uv add --refresh reflex-vla                   # uv (the --refresh flag is required;
                                              # uv caches the package index aggressively
                                              # and won't see new releases without it)

If reflex go fails after an upgrade, your export cache may be stale (built by an older version, schema mismatch). v0.5.4+ auto-detects and rebuilds; for older caches:

rm -rf ~/.cache/reflex/exports/<model_id>

Verify the install

reflex doctor

On a GPU box, look for green ✓ on these checks:

• TensorRT runtime (libnvinfer.so.10) — loadable
• CUDA cuBLAS (libcublas.so.12) — loadable
• CUDA cuDNN (libcudnn.so.9) — loadable
• ORT-TRT EP active — session created with TRT EP in active providers

If any are red ✗, the remediation hint says exactly which pip install to run.

GPU install needs FULL cuDNN 9

The pip wheel alone is not enough for production. Easiest path: NVIDIA’s container image:

docker run --gpus all -it nvcr.io/nvidia/tensorrt:24.10-py3
apt-get install -y clang   # for lerobot → evdev
pip install 'reflex-vla[serve,gpu,monolithic]'

Docker (zero-install serve)

# x86_64 CUDA runtime (cloud GPUs, dev workstations)
docker pull ghcr.io/fastcrest/reflex-vla:latest
docker run --gpus all \
  -v $(pwd)/p0:/exports \
  -p 8000:8000 \
  ghcr.io/fastcrest/reflex-vla:latest

# Jetson Orin / Orin Nano / Thor (arm64 + nvidia container runtime)
docker pull ghcr.io/fastcrest/reflex-vla:latest-arm64
docker run --runtime=nvidia \
  -v $(pwd)/p0:/exports \
  -p 8000:8000 \
  ghcr.io/fastcrest/reflex-vla:latest-arm64

The container’s default command is reflex serve /exports --host 0.0.0.0 --port 8000.