compileiq-search-space

작성자: nvidia

Search()의 search_space= 인자를 선택할 때 사용합니다. 세 가지 제공자 클래스(PtxasSearchSpace, NvccSearchSpace, LocalSearchSpaceBin)를 다루며, 방법은…

npx skills add https://github.com/nvidia/compileiq --skill compileiq-search-space

compileiq-search-space

CompileIQ ships compiler search spaces as release-backed binary blobs that the providers in compileiq.search_spaces.compilers fetch on demand. This skill covers the three provider classes, the variants they expose today, how to use them offline, and how to define a custom search space for non-compiler tuning.

When

  • Choosing a search space for a new project.
  • Pinning a specific release for a paper or production deployment.
  • Working on an air-gapped or corporate-firewalled host.
  • Stress-testing a one-off .bin you have on hand.

The three provider classes

Reference: compileiq/search_spaces/compilers.py:66-102.

from compileiq.search_spaces.compilers import (
    PtxasSearchSpace,
    NvccSearchSpace,
    LocalSearchSpaceBin,
)

# Default — auto-fetch latest PTXAS 13.3 default variant from GitHub releases
ss = PtxasSearchSpace()

# Pinned for reproducibility
ss = PtxasSearchSpace(version="13.3", variant="default", tag="search-spaces-2026.05")

# **Attention** workloads (FlashAttention / GQA / MHA / MLA / FlashInfer Batch Decode)
ss = PtxasSearchSpace(version="13.3", variant="att")

# NVCC variant for full-pipeline tuning
ss = NvccSearchSpace(version="13.3")

# Single .bin you already have on disk — skips manifest + network
ss = LocalSearchSpaceBin("/path/to/ptxas13.3_search_space.bin")

# Pass to Search
from compileiq.ciq import Search
tuner = Search(objective_function=..., search_space=ss, search_config=...)

Variants available today

From release/search-spaces/manifest-source.yaml:

CompilerVersionVariantFileWhen to use
ptxas13.3defaultptxas13.3_search_space.binGeneric PTXAS tuning; the right starting point for most kernels.
ptxas13.3attptxas13.3_att_search_space.binAttention workloads. att is short for attention, not "attribute". This variant is curated for FlashAttention, GQA, MHA, MLA, FlashInfer Batch Decode, and similar attention kernels. Prefer this whenever the kernel is attention-shaped.
nvcc13.3defaultnvcc13.3_search_space.binFull-compiler (front-end + back-end) tuning when you want NVCC-level knobs, not just PTXAS.

To enumerate variants in the latest release at any time:

gh release view --json assets --jq '.assets[].name' -R NVIDIA/CompileIQ <tag>

Air-gapped / offline mirror

Pre-download the manifest plus all .bin files on a connected host, then point CompileIQ at the local mirror via CIQ_SEARCH_SPACES_DIR:

# On a connected host
mkdir -p /shared/ciq-search-spaces
gh release download search-spaces-latest -R NVIDIA/CompileIQ -D /shared/ciq-search-spaces

# Move the directory to the air-gapped host (rsync, scp, sneaker-net, …)

# On the air-gapped host
export CIQ_SEARCH_SPACES_DIR=/shared/ciq-search-spaces
python -c "from compileiq.search_spaces.compilers import PtxasSearchSpace; print(PtxasSearchSpace().retrieve())"

Other env-var knobs:

  • CIQ_SEARCH_SPACES_REPO (default NVIDIA/CompileIQ): override the GitHub repo the resolver queries — useful for staging or forks.
  • CIQ_SS_TAG_PREFIX (default search-spaces-): tag prefix used when resolving tag="latest". Rarely needs changing.

Cache location

Resolved binaries are cached at ~/.cache/compileiq/<tag>/<sha256_prefix>_<filename>. The resolver verifies cached files by sha256; a corrupted entry is re-downloaded on the next call. Safe to wipe — wiping just costs one re-download.

Custom search spaces (non-compiler tuning)

For hyperparameter tuning, autotuner knobs, or any user-defined space, pass a dict (or list-of-dicts) directly instead of a provider. Primitives live in compileiq/search_spaces/base.py:

import compileiq.search_spaces.base as ss

search_space = {
    "block_size":  ss.choice([64, 128, 256, 512]),
    "unroll":      ss.range(start=1, end=8, step=1),
    "use_shmem":   ss.literal(True, knockout_prob=0.5),
    "lr":          ss.log_sampling(start=1e-5, end=1e-1, total=20),
}
PrimitivePurpose
choice([...])Sample uniformly from a discrete list.
range(start, end, step)Range-like sampling (also supports float steps).
literal(value, knockout_prob=...)Constant value; knockout_prob lets the GA disable this parameter.
log_sampling(start, end, total)Logarithmic distribution between start and end with total discrete buckets.

Mixed user + compiler search space (list shape):

search_space = [
    {"config_idx": ss.range(0, len(CONFIGS) - 1)},   # user-defined
    PtxasSearchSpace(version="13.3"),                # compiler-side
]

When the search space is a list, the objective receives a list of the same length — see compileiq-author-objective for the unpacking pattern.

Self-test

# Default variant resolves
python -c "
from compileiq.search_spaces.compilers import PtxasSearchSpace
p = PtxasSearchSpace().retrieve()
assert p.exists() and p.stat().st_size > 0, p
print(f'default OK: {p}')
"

# Attention variant resolves
python -c "
from compileiq.search_spaces.compilers import PtxasSearchSpace
p = PtxasSearchSpace(version='13.3', variant='att').retrieve()
assert p.exists() and p.stat().st_size > 0, p
print(f'att OK: {p}')
"

# Misconfigured air-gap mirror produces a clear error
CIQ_SEARCH_SPACES_DIR=/nonexistent python -c "
from compileiq.search_spaces.compilers import PtxasSearchSpace
try:
    PtxasSearchSpace().retrieve()
    print('UNEXPECTED: should have failed')
except Exception as e:
    print(f'expected failure: {type(e).__name__}')
"

Gotchas

  • att is attention, not "attribute". Earlier docs and a test fixture mislabeled the variant as "attribute-based register allocation" — corrected in this same branch. When recommending the variant to a user, say "attention" explicitly.
  • Providers vs Booster Packs are different things. Providers return one binary search space the optimization core consumes during a search; Booster Packs are pre-built .acf candidates to apply outside a search. Don't feed a .acf from a Booster Pack to PtxasSearchSpace(...).
  • LocalSearchSpaceBin doesn't validate .bin contents. It only validates that the file exists. A malformed file will fail downstream when the core tries to parse it; the resulting error message points at the resolver, not the provider.

Next

  • Writing the objective function that consumes the search space: compileiq-author-objective.
  • Running the search: compileiq-run-search.

nvidia의 다른 스킬

compileiq-debug
nvidia
무언가 잘못되었을 때 사용: Search()가 멈추거나, 모든 평가가 INVALID_SCORE를 반환하거나, 점수가 개선되지 않거나, 모든 설정이 동일한 숫자를 반환하거나, ptxas 오류 등이 발생할 때
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
compileiq-validate-result
nvidia
검색이 완료된 후, 속도 향상을 청구하거나 ACF를 발송하기 전에 사용합니다. dump_results CSV를 로드하고, 상위 K개 후보(단일 목표)를 추출합니다…
changelog-audit
nvidia
릴리스 전에 Warp CHANGELOG.md를 감사합니다: 누락된 항목 복구, 사용자 영향별 정렬, 항목 언어 다듬기, 줄 바꿈, (릴리스 브랜치 모드) 비교 업데이트…
maintain-dynamic-plugins
nvidia
NeMo Relay 동적 플러그인 로더, 매니페스트, Rust 네이티브 SDK, gRPC 워커 프로토콜, Python 워커 SDK, 문서, 테스트 및 릴리스 워크플로 커버리지를 유지 관리합니다.
dgx-diagnose
nvidia
일반적인 DGX Station GB300 문제 진단 — CUDA 충돌, 잘못된 GPU 타겟팅, vLLM/SGLang 컨테이너 버그, MIG 상태 문제, NVLink/Fabric Manager 오류,…