L0
Metadata-backed
The artifact or build axis is published upstream.
Does not claim local success before execution.
Open-source GPU compatibility resolver
Resolve your PyTorch GPU stack.
Detect the machine, resolve overlooked compatibility constraints, and get a sourced install or repair plan. Executed plans are verified locally.
python -m pip install rigsolvev0.1.2 alpha / Python 3.10+ / Linux x86_64 + NVIDIA CUDA / Apache-2.0
rigsolve / example output
$ rigsolve solve --want 'flash-attn==2.8.3' --target 'RTX 4090,driver=580.65,python=3.12,linux'matrix 2026.08.15 | evidence: metadata-backedtorch==2.9.0 index: cu126flash-attn==2.8.3 torch2.9 | cxx11abiTRUEwarning filename does not prove sm_89 kernel coverageDriver
runtime ceiling and minor-compatibility floor
CUDA
toolkit and binary runtime line
GPU
compute capability and architecture support
Python
interpreter, ABI, platform, and glibc
torch
release, index, and CUDA build
Extensions
torch coupling and C++ ABI
HOW IT WORKS
From failure to install plan.
Machine state, package requests, and sourced facts enter one deterministic solver.
- 01
Profile the environment
Read GPU, driver, toolkit, Python, platform, and installed package metadata without importing torch.
rigsolve detect - 02
Resolve the stack
Check driver, CUDA, Python, platform, torch, ABI, architecture, package coupling, and known broken constraints.
rigsolve solve - 03
Review the answer
Print an ordered plan or a reduced conflict with its evidence.
rigsolve why
EVIDENCE MODEL
Know what each result proves.
Artifact metadata, successful imports, and GPU runs are separate claims. rigsolve reports the evidence it has.
Current bundled matrix
2026.08.15 / 114 sourced facts / 1 blocked upstream edge
Plans start from published build metadata; executed plans are verified on the local machine.
LevelNameEstablishesLimit
L1
Install-tested
The exact artifact installed in a recorded environment.
Does not prove import.
L2
Import-tested
The package imported and its available build metadata was recorded.
Does not prove a GPU kernel ran.
L3
GPU-tested
A real kernel ran on the recorded GPU architecture.
Does not imply portability to other GPUs.
Offline by default
Detect, solve, check, and inspect the bundled matrix offline. No telemetry.
Plans before side effects
Review pip, uv, Dockerfile, TOML, JSON, or Colab output first. --execute installs and verifies.
Provenance is required
Every matrix fact includes its source, date, and evidence level.
OPEN SOURCE
Share a broken combination.
Precise reports and reviewed verification results extend hardware coverage.