Review install commands
The default output is a plan. Nothing installs without --execute.
terminal
rigsolve solve \
--want 'flash-attn==2.8.3' \
--target 'RTX 4090,driver=580.65,python=3.12,linux'Open source, offline first, built for NVIDIA GPU stacks
Resolve your GPU stackbefore you install.
rigsolve checks NVIDIA drivers, CUDA, Python, PyTorch, and native extensions together, then returns a sourced install or repair plan.
Detection avoids importing torch. Solving is offline and never installs packages by default.
Terminal/RTX 4090 example
$
rigsolve solve \
--want 'flash-attn==2.8.3' \
--target 'RTX 4090,driver=580.65,python=3.12,linux'# Generated by rigsolve; review before running.
# Matrix 2026.08.15 (1e066bd53f01); evidence: metadata-backed.
# WARNING: selected versions are metadata-backed; use --execute to install and verify them on this machine
# WARNING: flash-attn's wheel filename does not establish GPU kernel coverage for sm_89
python -m pip install --index-url https://download.pytorch.org/whl/cu126 torch==2.9.0
python -m pip install 'https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3%2Bcu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl#sha256=4e2f9e39313266b1544b68138b15b91ee6221eccf14f7902b7c6620351340810'
Offlinebundled compatibility matrix
Sourcedprovenance on every fact
Reviewableno install without --execute
Verifiableisolated imports and GPU probes
The gap in ordinary installation
Available packages can still form an incompatible GPU stack.
Package installers resolve declared dependencies. They do not jointly reason about driver ceilings, CUDA build lines, GPU architecture, Python wheels, torch coupling, and C++ ABI.
- Package installer
- Can these declared package requirements be installed?
- rigsolve
- Does the full GPU stack satisfy the recorded constraints?
Unknown evidence remains unknown. A successful solve is not a blanket runtime guarantee.
Six compatibility dimensions
Evaluate the stack as one system.
Every selected artifact must satisfy the applicable machine, package, platform, and native-build constraints.
| Dimension | Signal | What is checked |
|---|---|---|
| Driver | Runtime ceiling | Driver support for the selected CUDA runtime line. |
| CUDA | Binary runtime | Toolkit and package build markers on one compatible line. |
| GPU | Architecture | Compute capability and recorded kernel coverage. |
| Python | Wheel availability | Interpreter, ABI, platform, and glibc constraints. |
| PyTorch | Release coupling | Version, package index, and CUDA build compatibility. |
| Extensions | Native coupling | Torch version, CUDA line, GPU architecture, and C++ ABI. |
One plan, several outputs
Use the result where the installation happens.
Every renderer uses the same selected artifacts and constraints. Changing the format does not run the plan.
Feed automation
Render the same resolved plan as structured JSON.
plan.json
rigsolve solve --want torch \
--output json > plan.jsonCreate a container plan
Render a Linux target as a reviewable Dockerfile.
Dockerfile
rigsolve solve --want torch \
--output docker > DockerfileExplicit trust boundary
Plan first. Execute only when asked.
rigsolve detects without importing torch, solves against sourced facts, and prints the plan before any optional installation.
Read the trust model- 01
Detect the machine
Read driver, GPU, toolkit, Python, platform, and installed metadata.
- 02
Resolve constraints
Evaluate requested packages and the target against sourced facts.
- 03
Review the plan
Inspect artifacts, evidence levels, warnings, and commands.
- 04
Verify locally
Run isolated imports and available GPU probes after installation.
No torch import during detectionNo telemetryNo default installNo hidden fallback
Install the release
Inspect this machine before changing it.
Install v1.0.0, detect the local environment, and review a plan before deciding whether to execute it.
terminal
python -m pip install rigsolve
rigsolve detectApache-2.0, Python 3.10+, offline detection and solving
View on PyPI