ADu2021/cuda-l2-kernel-optimization
skills/skillxiv-v0.0.2-claude-opus-4.6/cuda-l2-kernel-optimization/SKILL.md
Uses LLMs with RL to automatically optimize HGEMM CUDA kernels across 1,000 configurations, systematically outperforming NVIDIA's cuBLAS and cuBLASLt through continued pretraining, general RL, and specialized HGEMM RL stages.
$ install --global
skillsauthnpx skillsauth add ADu2021/skillXiv cuda-l2-kernel-optimizationInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 17, 2026, 5:31 AM16.2s1 file scanned
SKILL.md
- name:
- cuda-l2-kernel-optimization
- title:
- CUDA-L2: Surpassing cuBLAS via Reinforcement Learning for Matrix Multiplication
- version:
- 0.0.2
- engine:
- skillxiv-v0.0.2-claude-opus-4.6
- license:
- MIT
- url:
- https://arxiv.org/abs/2512.02551
- keywords:
- [cuda-optimization, reinforcement-learning, kernel-tuning, matrix-multiplication, systems-ml]
- description:
- Uses LLMs with RL to automatically optimize HGEMM CUDA kernels across 1,000 configurations, systematically outperforming NVIDIA's cuBLAS and cuBLASLt through continued pretraining, general RL, and specialized HGEMM RL stages.
Summary
CUDA-L2 combines large language models with reinforcement learning to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. The system employs a three-stage training approach: continued pretraining on diverse CUDA code, general kernel RL training, and specialized HGEMM RL training using execution speed as the reward signal.
Core Technique
Multi-Stage LLM Training:
- Continued Pretraining: Fine-tune LLM on high-quality CUDA kernel code to understand optimization patterns
- General RL: Teach the model to generate working kernels and improve via speed rewards
- Specialized HGEMM RL: Focus RL on matrix multiplication variants and configurations
Kernel Generation: The LLM generates complete CUDA kernel source code as text. Each generation is compiled, executed, and evaluated on speed.
Reward Signal: Execution time on representative workloads:
reward = baseline_speed / optimized_speed
Higher reward indicates better optimization.
Implementation
Continued pretraining data: Collect CUDA kernel implementations:
# Dataset: [kernel_code, optimization_notes, performance_hints]
pretrain_data = [(kernel1, notes1), (kernel2, notes2), ...]
General RL training:
for iteration in range(num_iterations):
# Generate kernel code
kernel_code = llm.generate(prompt=problem_spec)
# Compile and execute
compiled = compile_cuda(kernel_code)
speed = measure_execution_time(compiled, workload)
# Compute reward
reward = baseline_speed / speed if compiled_successfully else -1.0
# Update LLM
llm.update_with_rl(trajectory, reward)
HGEMM specialization: Create specialized prompts:
task_prompt = """
Optimize HGEMM for:
- Matrix size: 4096x4096
- Batch size: 16
- Hardware: A100 GPU
Requirements: maximize throughput
"""
When to Use
- Kernel optimization when manual tuning is insufficient
- Scenarios where a large space of CUDA configurations needs exploration
- Tasks where execution speed directly impacts system performance
- Applications with diverse workload patterns requiring adaptive kernels
When NOT to Use
- Simple kernels where hand-tuning or cuBLAS is sufficient
- Scenarios without access to LLMs or RL infrastructure
- Real-time compilation where LLM generation latency is prohibitive
- Tasks requiring guaranteed mathematical correctness (LLMs may generate incorrect kernels)
Key References
- CUDA kernel optimization and performance tuning
- Reinforcement learning for code generation
- Language models for system optimization
- Matrix multiplication and GEMM operations
Related Skills
ADu2021/flow-map-trajectory-tilting
testing
VerifiedTrustedCommunity
Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexible-data-mixture-of-experts
testing
VerifiedTrustedCommunity
Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexibility-trap-diffusion-reasoning
data-ai
VerifiedTrustedCommunity
Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flex-continuous-agent-evolution
devops
VerifiedTrustedCommunity
Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.
2SKILL.mdUpdated Apr 17, 2026