skills.rest
KangOxford
KangOxfordCOMMUNITY
Explore all skills by this author →
GitHub

siso-cuda-kernel-opt-tradeoffs

Analyze SISO CUDA kernel tradeoffs for state-space kernels across d_state values.

1|Updated Apr 23, 2026
Search Tags:#optimization#throughput#kernel#cuda#state-space#occupancy#smem
One-click install
npx skills add https://github.com/KangOxford/auto-quant-research --skill siso-cuda-kernel-opt-tradeoffs
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: siso-cuda-kernel-opt-tradeoffs
Source: https://github.com/KangOxford/auto-quant-research/tree/main/execution-layer/skills/siso-cuda-kernel-opt-tradeoffs
Command: npx skills add https://github.com/KangOxford/auto-quant-research --skill siso-cuda-kernel-opt-tradeoffs

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Document SISO (Mamba3 state-space) CUDA FFI kernel optimization tradeoffs — N-split structural constraint (V buffer not split), union SMEM lifetimes, launch_bounds occupancy tuning, per-sample speedup saturation with BSZ. Use when tuning K3 kernel for different d_state values or similar state-scan kernels.

Core Features & Use Cases

  • N_SPLITS tradeoff analysis and guidance for SMEM traffic and occupancy
  • SMEM lifetime unions to reduce memory footprint without sacrificing correctness
  • Launch_bounds tuning and CS sizing recommendations for different d_state values
  • Per-sample speedup saturation curve methodology and verification workflow

Quick Start

Run the optimization workflow to identify the optimal N_SPLITS, SMEM lifetime unions, and CS settings for your d_state values.

Frequently Asked Questions about siso-cuda-kernel-opt-tradeoffs

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I optimize CUDA kernel occupancy and SMEM usage for state-space models?▼

Tune state-space kernel occupancy by adjusting N_SPLITS, SMEM lifetime unions, and launch_bounds hints to balance throughput and occupancy across varying d_state values.

How does N_SPLITS affect state-space kernel throughput and SMEM traffic?▼

N_SPLITS impacts state-space kernel throughput by partitioning structural constraints without splitting the V buffer, balancing SMEM traffic and occupancy across different d_state configurations.

What's the best way to tune launch_bounds and CS sizing for Mamba3 state-space kernels?▼

Tune launch_bounds and CS sizing for Mamba3 state-space kernels by analyzing per-sample speedup saturation curves across varying d_state values to identify optimal throughput and occupancy configurations.

Can I reduce SMEM footprint in CUDA state-space kernels without sacrificing numerical correctness?▼

Reduce SMEM footprint in state-space kernels by applying SMEM lifetime unions, decreasing memory usage while maintaining numerical correctness verified within tolerance.

What do I need to verify numerical correctness when tuning CUDA state-space kernels?▼

Verify numerical correctness when tuning CUDA state-space kernels by running benchmarking scripts and verification checks to ensure outputs remain within tolerance after adjusting N_SPLITS, SMEM unions, and CS sizing.

Why does per-sample speedup saturate with increasing batch size in state-space kernels?▼

Per-sample speedup saturates with increasing batch size in state-space kernels due to occupancy limits; saturation curve analysis determines optimal throughput configurations for specific d_state values.

Related Skills

Explore More Software Engineering Skills→
tile-aitile-ai
168

tune

Profile TileOPs GPU kernel latency and bottlenecks with CUDA-event timings.

Official
Advanced
drunkcodingdrunkcoding
4

add-cuda-kernel

Extend FlashInfer with new CUDA kernels from definition to Python API.

Community
Advanced
xchang1121xchang1121
6

cuda-c-optimization

Optimize CUDA C kernels for performance, memory access, and numerical stability.

Community
Advanced
mindspore-aimindspore-ai
258

tilelang-cuda-optimization

Optimize TileLang CUDA kernels and diagnose compilation and runtime issues.

Official
Advanced