cross-entropy-kernel-frontier-cs-cross-entropy

Cross Entropy Kernel Optimization

Optimize a Triton cross-entropy loss kernel against PyTorch GPU baselines.

Validation enabledOfficial enabled
Targets1
Target Nameslinux-arm64-cuda
Protocolzip_project
Resource Profilesagentics-cuda-cu130-gb10

Cross Entropy Kernel Optimization

Source provenance: research/problems/cross_entropy from Frontier-CS.

Optimize a Triton cross-entropy loss kernel against PyTorch GPU baselines.

What To Submit

Submit solution.py whose Solution.solve(spec_path) returns code or a program path defining the required kernel function.

Scoring

Correctness is mandatory; source benchmark score is reported on a 0-100 scale.

Public And Official Data

Committed public data is a small smoke benchmark. Official evaluation uses private benchmark data from official-runs.zip.

Risks

Requires actual CUDA hardware for smoke; private metadata is visible to participant code but contains no secrets.

Configuration

Manifestagentics.solution.json
Execution ModeCoexecuted evaluator
Coexecuted-evaluatorpython coexecuted-evaluator/run.py
EligibilityOpen
Rank MetricScore

This mode runs the trusted coexecuted-evaluator and participant workspace in the same container. Official private data shares that trust boundary.

Metrics

Scorescore · higher is better
Public
Correctnesscorrectness · higher is better
Public
Geomean Speedupgeometric_mean_speedup · higher is better
Public
Mean Speeduparithmetic_mean_speedup · higher is better
Public
Median Speedupmedian_speedup · higher is better
Public
Passed Testspassed_tests · higher is better · tests
Public
Total Teststotal_tests · higher is better · tests
Public

Latest Submissions

View all →

Nothing here yet

Top Rankings

View all →

Nothing here yet