Official Support Matrix#
This page lists supported software stacks. See Supported Models for checkpoint IDs and Installation for build commands.
Platforms#
Platform |
Level |
OS / SDK |
CUDA Toolkit |
TensorRT |
Build location |
Precision constraint |
|---|---|---|---|---|---|---|
Jetson Thor |
Official |
JetPack 7.0 / 7.1 |
13.0 |
JetPack package |
Device |
Model-dependent |
Jetson Thor |
Official |
JetPack 7.2 |
13.2 |
JetPack package |
Device |
Model-dependent |
NVIDIA DRIVE Thor |
Official |
DriveOS 7.2 |
13.3 |
DriveOS SDK package |
SDK container, then deploy |
Model-dependent |
NVIDIA DGX Spark (GB10) |
Official |
DGX Spark software stack |
13.0 |
System package |
Device |
Model-dependent |
Jetson Orin |
Official |
JetPack 7.2 |
13.2 |
JetPack package |
Device |
FP16, INT8, and INT4 only |
NVIDIA IGX Thor |
Official |
Current Linux stack |
13.0 |
10.13.3.9 |
Device |
Model-dependent; SM110 iGPU or SM120 dGPU |
x86-64 Linux GPU |
Developer |
Ubuntu 22.04 / 24.04 |
12.x or 13.x |
Compatible user package |
Workstation |
Development and validation |
Official combinations are release-tested deployment targets. Compatible
combinations are expected to work with the stated constraints. Developer
combinations support development but are not edge deployment targets.
Jetson Orin does not run FP8 or FP4 model engines. Edge deployments normally use the TensorRT version supplied by the platform SDK; x86 builds must use mutually compatible TensorRT and CUDA packages.
Wheel Packaging Matrix#
The wheel tooling is configured to assemble one x86_64 and one aarch64 wheel
for each supported CPython minor: 3.10, 3.11, and 3.12: six release artifacts,
not separate downloads per GPU or TensorRT version. The platform tags are
manylinux_2_35_x86_64 (glibc 2.35+) and manylinux_2_39_aarch64 (glibc 2.39+).
These are installation floors, not support for every newer Linux stack.
Native payload selection is exact; the loader does not guess a nearest SM or
TensorRT major. See published-wheel installation
for setup and optional Python dependencies.
Wheel architecture |
Configured runtime rows |
|---|---|
x86_64 |
Ubuntu 22.04, CUDA 13, SM80, TensorRT 10 |
x86_64 |
Ubuntu 24.04, CUDA 13, SM86/SM100/SM120, TensorRT 10 |
x86_64 |
Ubuntu 24.04, CUDA 13, SM80/SM86/SM100/SM120, TensorRT 11 |
aarch64 |
Jetson Orin: JetPack 7.2, CUDA 13, SM87, platform TensorRT 10 |
aarch64 |
Jetson Thor: JetPack 7.0/7.1/7.2, CUDA 13, SM110, platform TensorRT 10 |
aarch64 |
DRIVE Thor: DriveOS 7.2, CUDA 13, SM110, platform TensorRT 10 |
aarch64 |
IGX Thor current stack, CUDA 13, SM110/SM120, platform TensorRT 10 |
aarch64 |
DGX Spark current stack, CUDA 13, SM121, platform TensorRT 10 |
Release qualification first checks the minimal base workflow
in a clean environment without optional workflow packages, then installs
[server] and checks the high-level API. Both phases build a small model and
require generated text and tokens. The integration gate requires evidence for
the enabled qualification targets and their configured Python ABIs before
publication.
The wheel contract matches the observed platform release, CUDA and TensorRT SONAMEs, and GPU SM exactly. It does not claim NVIDIA driver-version ranges; driver compatibility remains part of the CUDA/platform support contract. The broader x86 developer-source row remains useful for source builds but is not a promise that every OS/CUDA/SM cross-product is present in version-1 wheels. Adding a wheel row requires a canonical exact-SM CuTe artifact and passing installed-wheel build and inference validation.