← LeaderboardThe Skill Leaderboard

compileiq-author-objective

Use when writing the objective_function= passed to Search(). Covers the two legal signatures (compiler-only str vs mixed list), the baseline-knockout branch, per-eval cache busting, framework-specific --apply-controls injection (raw PTXAS, NVCC, Triton, Helion, cuTeDSL/FA4, FlashInfer), correctness-before-timing, INVALID_SCORE handling, and the Debug-pack O0/O3 ACF-injection canary that must pass before launching a search. Triggers on "objective function", "apply-controls", "INVALID_SCORE", "save_compiler_config", "baseline knockout", "BASELINE_CONFIG", "every config returns the same score", "

Published by NVIDIA ↗NVIDIA/skills

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

name: compileiq-author-objectivedescription: >  Use when writing the objective_function= passed to Search(). Covers the two  legal signatures (compiler-only str vs mixed list), the baseline-knockout  branch, per-eval cache busting, framework-specific --apply-controls  injection (raw PTXAS, NVCC, Triton, Helion, cuTeDSL/FA4, FlashInfer),  correctness-before-timing, INVALID_SCORE handling, and the Debug-pack  O0/O3 ACF-injection canary that must pass before launching a search.  Triggers on "objective function", "apply-controls", "INVALID_SCORE",  "save_compiler_config", "baseline knockout", "BASELINE_CONFIG", "every config  returns the same score", "TypeError fromhex".when_to_use: |  - About to write or modify the function passed as objective_function=.  - Search runs but every config returns an identical score (canary fails).  - Getting TypeError around fromhex or bytes (legacy pattern).  - Search hangs on individual configs (timeout/correctness handling missing).license: CC-BY-4.0 AND Apache-2.0metadata:   version: "1.2.0"  author: CompileIQ Team <compileiq@nvidia.com>  domain: compiler-optimization  tags: [compileiq, cuda, ptxas, compiler-optimization, autotuning]allowed-tools: Bash Readpaths: ["**/*.py", "**/*.cu", "**/*.cuh"]

compileiq-author-objective

The objective function is where ~80% of CompileIQ user errors happen. This skill tells you the exact shape it must have for current CompileIQ, how to inject --apply-controls for each supported compile path, and how to verify the whole pipeline works before paying for a full search.

For paste-ready full-file templates per framework, see references/templates.md.

When

  • Writing a brand-new objective function.
  • Migrating an older objective off the legacy bytes.fromhex(config_blob) pattern.
  • Diagnosing "every config returns the same score" or "TypeError: fromhex".

The two legal signatures

Shape of search_space=Objective signatureWhat config is
Single provider, e.g. PtxasSearchSpace()def objective(config: str) -> floatA hex string. Pass it straight to save_compiler_config(acf_path, config).
List, e.g. [{"k": ss.choice(...)}, PtxasSearchSpace()]def objective(mixed: list) -> floatA list of the same length. Unpack: user_space, ptxas_config = mixed, but check for BASELINE_CONFIG first (see below).

Mixed-space results keep the same list shape in best["params"]. Unpack it before saving the ACF, for example: user_space, ptxas_config = best["params"]. (Pattern reference: examples/compilers/triton_example/mixed_triton.py.)

For multi-objective, return tuple[float, ...] of length num_objectives.

Canonical imports

from compileiq.types import INVALID_SCORE, BASELINE_CONFIG
from compileiq.utils.helpers import save_compiler_config

INVALID_SCORE is CompileIQ's sentinel — return it on any failure (compile, hang, wrong answer, exception). Do not redefine it as float('inf').

BASELINE_CONFIG is the empty-dict sentinel ({}) CompileIQ passes for a baseline run. With normalize=True the first task of the search is always a baseline run; without normalization it can still appear when knockout removes every parameter. It is {} whatever the search-space shape, including mixed (list) spaces.

save_compiler_config(path, hex_str) writes the binary blob to disk; it handles the bytes.fromhex internally (compileiq/utils/helpers.py:128-137). Users never need to touch fromhex themselves.

Self-contained for IsoMultiProcessWorker and Ray

Heavy library imports (torch, triton, helion, cute) go inside the function so the process IsoMultiProcessWorker spawns — or the remote Ray task — can re-import them in a clean state. Cheap module-level constants (paths, regexes) are fine.

Per-eval cache busting (non-negotiable)

import os, tempfile
# Set at the start of each evaluation, inside the objective's process.
os.environ["TRITON_ALWAYS_COMPILE"] = "1"
os.environ["HELION_SKIP_CACHE"]     = "1"
os.environ["TRITON_CACHE_DIR"]      = tempfile.mkdtemp(prefix="ciq_triton_")

If the benchmark runs in a subprocess instead, prefix its command with the same three variables (for example TRITON_ALWAYS_COMPILE=1 TRITON_CACHE_DIR=... python bench.py).

For FlashInfer, additionally confirm the prebuilt cubin cache packages are absent — flashinfer_cubin and flashinfer_jit_cache. See docs/flashinfer_booster.md for the import-time check.

Per-framework --apply-controls injection

TargetInjection
Raw PTXAS (you have a .ptx file)ptxas --apply-controls candidate.acf kernel.ptx -arch=sm_100 -o kernel.cubin
NVCC source (CUDA .cu)nvcc --apply-controls candidate.acf -arch=sm_100 kernel.cu -o exe, which applies the ACF to every stage it lists (nvcc and ptxas). This form needs a Blackwell or later target (-arch=sm_100 and up): for earlier targets such as sm_90, nvcc rejects --apply-controls and the build fails, even for a ptxas-only ACF. There, use nvcc -Xptxas --apply-controls=candidate.acf ..., which reaches ptxas only and silently ignores any nvcc-stage controls in the ACF.
Triton kernellaunch keyword argument ptx_options=f"--apply-controls={acf_path}" on the kernel launch (Triton 3.4+) plus TRITON_ALWAYS_COMPILE=1, os.environ["TRITON_PTXAS_PATH"] = shutil.which("ptxas"), and os.environ["TRITON_PTXAS_BLACKWELL_PATH"] = shutil.which("ptxas") when Blackwell-specific PTXAS selection may apply. Use this inside objectives; see the PTXAS_OPTIONS gotcha below.
HelionHelion's official ACF API. See https://helionlang.com/examples/acfs/softmax_acf.html. Always set HELION_SKIP_CACHE=1.
cuTeDSL / FA4 (TVM-FFI)cute.compile(..., options=f"{existing_options} --ptxas-options '--apply-controls {acf_path}'"). If you can't reach the call site, patch CompileCallable.__call__ to splice in the option string.
FlashInferFLASHINFER_EXTRA_CUDAFLAGS="--ptxas-options=--apply-controls=$ACF_FILE" (see docs/flashinfer_booster.md).

Baseline knockout branch

def objective(config):
    if isinstance(config, dict) and not config:   # config == BASELINE_CONFIG
        return measure_without_acf()              # establish baseline run
    # config is a hex string: apply the ACF
    ...

Mixed (list) search spaces get the same {}, so check before unpacking, and measure your default user parameters with no ACF:

def objective(mixed):
    if mixed == BASELINE_CONFIG:
        return measure(DEFAULT_USER_CONFIG, acf_path=None)
    user_space, ptxas_config = mixed               # safe to unpack now
    ...

Without this check, unpacking {} raises ValueError on the first task when normalize=True.

Correctness-before-timing (mandatory)

The optimizer rewards whatever you measure. If you only measure latency, the algorithm will happily reward configs that compile faster by producing wrong answers. Always verify against a reference first:

if not torch.allclose(actual, reference, atol=1e-2, rtol=0):
    return INVALID_SCORE
return triton.testing.do_bench(lambda: kernel(...), warmup=100, rep=1000, return_mode="mean")

(Pattern from examples/compilers/triton_example/mixed_triton.py.)

Catch everything → return INVALID_SCORE

try:
    ...
except (subprocess.TimeoutExpired, RuntimeError, FileNotFoundError, ValueError, OSError) as e:
    return INVALID_SCORE

When in doubt, catch broadly. CompileIQ's workers do catch an exception that escapes the objective, print the traceback, and record INVALID_SCORE with a warning, so one bad config does not stop the search. Catching it yourself is still better: you control cleanup (temp ACFs, build dirs), and you can tell expected failures (compile error, wrong answer) from bugs in the objective. What does stop a search: under IsoMultiProcessWorker, a child process that dies abnormally (for example a segfault, or a return shape that doesn't match num_objectives) raises RuntimeError.

Pre-search canary (mandatory before tuner.start())

Two cheap calls that catch ~90% of "every score is the same" bugs:

# Shape check — does the objective even run?
sample = tuner.sample(1)[0]
score = objective(sample)
print(f"sample run: {score}")
assert isinstance(score, (int, float)) and score == score   # not NaN

# ACF-injection canary using the Debug pack (downloaded once)
from compileiq.utils.helpers import load_compiler_config
O0_HEX = load_compiler_config("booster-pack-debug/ptxas_opt0.acf")
O3_HEX = load_compiler_config("booster-pack-debug/ptxas_opt3.acf")

baseline = objective({})                  # BASELINE_CONFIG path
score_O0 = objective(O0_HEX)              # mixed space: objective([default_user_params, O0_HEX])
score_O3 = objective(O3_HEX)              # mixed space: objective([default_user_params, O3_HEX])

assert score_O0 > baseline * 1.05, (
    f"O0 should regress (got {score_O0} vs baseline {baseline}). "
    "ACF is NOT reaching PTXAS — fix the cache-bust."
)
assert abs(score_O3 - baseline) / baseline < 0.05, (
    f"O3 should match baseline (got {score_O3} vs {baseline})."
)
print("ACF injection canary PASSED — safe to start the search.")

If either assertion fails, stop and fix the cache-bust before launching the search. Otherwise every generation's score is measurement noise on a stale binary.

Self-test

A 3-line "smoke" objective inside the SKILL author's repo, used to verify the scaffolding before plugging in a real kernel:

def smoke_objective(config):
    return 1.0   # constant; useful to verify Search() shape, not measurement

Drop it into the Search(...) call and run 2 generations; if that completes and results.get_best_result() returns a dict, your scaffold is correct.

Gotchas

  • Don't inject Triton ACFs through PTXAS_OPTIONS inside an objective. Triton 3.6+ reads PTXAS_OPTIONS once, when its NVIDIA backend is imported, as the default for ptx_options; earlier versions ignore it. Changing os.environ["PTXAS_OPTIONS"] between evaluations therefore does nothing. Pass ptx_options= on the kernel call instead. The env var is only safe as a prefix on a whole benchmark process (PTXAS_OPTIONS=... python bench.py, Triton 3.6+).
  • Triton caches on the ACF path, not its contents. Rewriting the same ACF file each evaluation hits the cache and reuses the old binary unless TRITON_ALWAYS_COMPILE=1 is set.
  • Mixed search spaces require list unpacking. If you pass search_space=[user_dict, PtxasSearchSpace()], your objective must accept a list, not a string. Results keep that list in best["params"]; unpack it before saving the compiler config.
  • Don't redefine INVALID_SCORE. Import it from compileiq.types. If you redefine it locally as float('inf'), the value happens to work today but is not guaranteed to in future releases.
  • config_blob is no longer a parameter name. The old skill set used def objective(config_blob) and called bytes.fromhex(config_blob). Both are stale. Use def objective(config) and save_compiler_config(path, config).

Next

  • Sizing SearchConfiguration and picking a Worker: compileiq-run-search.
  • After the search: compileiq-validate-result.
  • If something's wrong: compileiq-debug.