flashinfer-b200-contest-optimizer

Optimize FlashInfer CUDA kernels for NVIDIA B200 using profiling and parity validation.

19|2|Updated May 14, 2026

Search Tags:#optimization #profiling #kernel #cuda #flashinfer #nsight-compute #b200

Ready to run

npx skills add https://github.com/syhya/mlsys26-flashinfer-contest --skill flashinfer-b200-contest-optimizer

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: flashinfer-b200-contest-optimizer
Source: https://github.com/syhya/mlsys26-flashinfer-contest/tree/main/agent-assisted/skills/flashinfer-b200-contest-optimizer
Command: npx skills add https://github.com/syhya/mlsys26-flashinfer-contest --skill flashinfer-b200-contest-optimizer

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill addresses the complexity of optimizing high-performance CUDA kernels for NVIDIA B200 hardware by providing a structured, reference-first, and data-driven workflow that prevents inefficient trial-and-error.

Core Features & Use Cases

Reference-First Reconnaissance: Automatically scans local and official repositories to identify reusable kernel primitives before starting new development.
Shape-Aware Optimization: Uses real contest workload data to build representative regimes, ensuring optimizations are effective across the entire workload distribution.
Blackwell-Specific Tactics: Provides a clear path for applying advanced CUDA 13.2 features like Cluster Launch Control, DSMEM, and L2 persistence when profiler evidence justifies them.
Use Case: When a specific MoE kernel shows high latency on large-batch workloads, use this Skill to profile the bottleneck, identify a reusable primitive from a reference repo, and validate the optimized kernel against official contest parity benchmarks.

Quick Start

Use the flashinfer-b200-contest-optimizer skill to refresh local references and begin the optimization round for the active operator definition.

Frequently Asked Questions about flashinfer-b200-contest-optimizer

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

FAQPage Schema

How do I optimize CUDA kernels for NVIDIA B200 hardware?▼

To optimize CUDA kernels for NVIDIA B200 hardware, use a reference-first reconnaissance workflow that scans local and official repositories for reusable primitives before profiling bottlenecks with Nsight Compute and validating parity against official benchmarks.

What is the best way to profile FlashInfer kernel bottlenecks on large-batch workloads?▼

Profiling FlashInfer kernel bottlenecks on large-batch workloads requires using Nsight Compute to gather profiler evidence, which then justifies applying Blackwell-specific tactics like Cluster Launch Control, DSMEM, and L2 persistence.

Do I need to validate optimized kernels against official contest benchmarks?▼

Yes, you must validate optimized kernels against official contest benchmarks to ensure parity. The workflow enforces paired-baseline gating and reproducible artifact archival to prevent inefficient trial-and-error during performance tuning.

How does shape-aware profiling improve FlashInfer contest operator performance?▼

Shape-aware profiling improves FlashInfer contest operator performance by using real contest workload data to build representative regimes, ensuring your CUDA optimizations remain effective across the entire workload distribution rather than isolated cases.

When should I apply Blackwell-specific CUDA 13.2 features like Cluster Launch Control?▼

You should apply Blackwell-specific CUDA 13.2 features like Cluster Launch Control and DSMEM only when concrete profiler evidence from Nsight Compute justifies their use, avoiding premature optimization without data-driven bottleneck analysis.

Why does my MoE kernel show high latency during CUDA kernel development?▼

High latency in MoE kernels during CUDA kernel development often stems from unoptimized bottlenecks on large-batch workloads, which you can identify by profiling with Nsight Compute and replacing inefficient code with reusable reference primitives.