nvidia/compileiq-bootstrap
> Use when starting a fresh CompileIQ project, hitting a socket timeout, or before running any other compileiq-* skill. Verifies CUDA 13.3+, ptxas, GPU access, that `from compileiq.ciq import Search` and friends resolve, and that `PtxasSearchSpace().retrieve()` returns a real path. Documents the env vars that control timeouts, caching, and search-space mirroring. Triggers on "set up compileiq", "compileiq doesn't work", "socket timeout", "where do search spaces come from", "air-gapped compileiq".
npx skills add https://github.com/NVIDIA/CompileIQ --skill compileiq-bootstrap
Validate that a host can run CompileIQ. No installation of system libraries
(that's distro-specific and out of scope). No state files. Just a checklist
that either passes or prints the precise next fix.
Search().start() hangs on the first evaluation.compileiq wheel, or the Python venv.CompileIQ's --apply-controls mechanism requires CUDA 13.3 or later
(see examples/compilers/nvbench_example/optimize_reduction.py:302).
nvcc --version | grep -E "release (1[3-9]|[2-9][0-9])\.[3-9]"
ptxas --version | grep -E "V(1[3-9]|[2-9][0-9])\.[3-9]"
If either command exits non-zero, install or upgrade the CUDA toolkit and put
/usr/local/cuda/bin on PATH and /usr/local/cuda/lib64 on
LD_LIBRARY_PATH.
nvidia-smi --query-gpu=name,compute_cap --format=csv
CompileIQ targets compute capability 9.0 (Hopper / H100) and 10.0 (Blackwell /
B200) most aggressively, but works on any GPU PTXAS supports for the chosen
arch.
One shot that covers everything callers will need:
python -c "
from compileiq.ciq import Search
from compileiq.types import INVALID_SCORE, BASELINE_CONFIG, WorkerTypes, ProblemType, SearchConfiguration
from compileiq.search_spaces.compilers import PtxasSearchSpace, NvccSearchSpace, LocalSearchSpaceBin
from compileiq.utils.helpers import save_compiler_config, load_compiler_config
from compileiq.worker import MultiProcessWorker, IsoMultiProcessWorker, RayWorker, AsyncWorker
print('imports OK')
"
If this fails, run pip install compileiq (or pip install -e . from a
source checkout) and re-run.
This catches network or air-gapped issues that otherwise surface as
socket-timeout hangs deep inside the first evaluation:
python -c "
from compileiq.search_spaces.compilers import PtxasSearchSpace
p = PtxasSearchSpace().retrieve()
assert p.exists() and p.stat().st_size > 0, p
print(f'resolved: {p}')
"
The first run downloads from GitHub releases and caches under
~/.cache/compileiq/<tag>/. Subsequent runs hit the cache.
| Variable | Default | What it controls |
|---|---|---|
| CIQ_SOCKET_TIMEOUT | 20 | Seconds to wait on IPC with the core. Raise to 60-120 for large search spaces; raise much higher if you regularly see a hang on the first evaluation. |
| CIQ_KEEP_CACHE | unset (0) | Set to 1/true/yes to keep .cache files after a run for post-mortem replay. |
| CIQ_PROCESS_MODE | forkserver | forkserver, fork, or spawn. Switch to spawn if forkserver fails on a constrained host. IsoMultiProcessWorker defaults to fork regardless. |
| CIQ_SEARCH_SPACES_DIR | unset | Path to a local mirror containing manifest.json plus the referenced .bin files. Set this on air-gapped hosts to skip network. |
| CIQ_SEARCH_SPACES_REPO | NVIDIA/CompileIQ | Override the GitHub repo that release-backed search-space resolution queries. Useful for staging or forks. |
| CIQ_SS_TAG_PREFIX | search-spaces- | Tag prefix the resolver uses when tag="latest". Rarely needs changing. |
bash scripts/check_env.sh
Exits 0 if every step passes. Exits with a non-zero count of failures and
prints, for each failure, the precise command the user should run next.
not require BLAS/LAPACK. If a hang occurs:
CIQ_SOCKET_TIMEOUT=120 and retry.CIQ_SEARCH_SPACES_DIR=/path/to/mirror.
CIQ_PROCESS_MODE=spawn.installs require sudo and break in containers and on shared clusters. If
nvcc, ptxas, or a Python interpreter is missing, surface the error
and let the user decide how to install.
compileiq-search-space (variant="att" by default).compileiq-booster-pack.compileiq-author-objective.Take nvidia/compileiq-bootstrap from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference pip.
Without those the skill loads but fails at the first command.