01
Submit a request
Copy the example, add your definition, workloads, and baseline, then open a pull request to wishlist. No issue is required.
Agentic CUDA kernel program
Community wants,
KDA provides!
Need a CUDA kernel for your training/inference but cannot find the experts to implement?
Submit your definitions and workloads to KDA-wishlist, and KDA will automatically optimize it w/ agentic workflows!
Part of NVIDIA agentic CUDA and RSI efforts
Currently supporting NVIDIA B200 and B300 GPUs only.
REQUEST ACCEPTING
01
YOU PROVIDEFlashInfer Trace
Definition · Workloads · Baseline
research · generate · verify
02
WE RUNKernel Design Agents
Explore and benchmark candidates
select strongest candidate
03
YOU GETValidated kernel
Code · Benchmarks · Reproduction
CUDA kernels are foundational to the modern LLM ecosystem, but building high-performance implementations requires deep expertise.
KDA Wishlist is an experimental program where the KDA team uses agentic workflows to build the kernels the community needs.
HOW IT WORKS
From definitions
to implementations
Kernel Design Agents (KDA) are agentic CUDA systems that research, implement, verify, and iterate on performance-sensitive kernel tasks. Each request gives them a measurable contract—not just an idea.
FlashInfer Trace is a standardized JSON format for describing a GPU kernel definition and its workload.
PUBLIC REQUEST LIFECYCLEGITHUB-MANAGED
02
Community votes
Others add a thumbs-up reaction to the pull request and contribute useful workload context.
03
KDA thinks and builds
Accepted tasks enter a measured loop of analysis, generation, correctness checks, and profiling.
04
Results return
We publish the strongest validated kernel with benchmarks, reproduction steps, and known limits.
READY TO CONTRIBUTE
Submit your
request
Push your request files to a branch based on wishlist, then enter your fork and branch.
SELECTED ACHIEVEMENTS
Built by agents.
Merged upstream.
KDA has already turned optimization targets into production contributions across LLM serving, image generation, and video workloads.
01 / FLASHINFER CONTEST SOTA
1.39×peak vs. human SOTA
KDA-0.5 surpassed human SOTA
MLSys’26 FlashInfer Kernel Contest · B300 · MoE 2.25× · DSA 29.95× · GDN 6.10×
02 / SOLEXEC 1ST PLACE
#1SOLExec L1
KDA boosts Databricks to rank #1 on SOL-Bench
L1 Single Operation track · score 0.7608 · previous best 0.7584
03 / FLASHINFER CONTEST TOP 3
TOP3all contest tracks
KDA-0.1 achieves TOP3 on every track
MLSys’26 FlashInfer Contest · B300 · MoE 0.67× · DSA 11.91× · GDN 1.16×
04 / FLASH-KMEANS OPEN SOURCE
6.1×balanced k-means
Flash-KMeans
KDA speedup for the Wan 2.2 workload on B200
05 / SGLANG MERGED
2.03×weighted B200 kernel group
Cosmos3 causal Conv3D cat-pad
10.621 → 5.240 ms weighted; model E2E +2.11%
06 / SGLANG MERGED
5.84×kernel geomean
LTX-2.3 QKNorm + split-RoPE
B200 · 4.22–7.34× across 14 production shapes; E2E −9.16%
More improvements
- 1.21× selected GEMM · KDA NVFP4 GEMM for Qwen3.8 on SM120 · SGLang #36865
- +4.45% E2E throughput · Restore SM121 QSA correctness with Humanize + KDA · SGLang #36845
- 4.24–19.86% faster · Blackwell fused MHC half-MMA kernel · TensorRT-LLM #16799
- 1.06× peak GEMM speedup · SM100 NVFP4 SVDQuant fused GEMM · FlashInfer #3858
- 5.84× kernel geomean · LTX-2.3 QKNorm + split-RoPE CUDA fast path · SGLang #29708
- 1.032× model E2E · LTX-2 residual-gate CUDA fast path · SGLang #29361
- 2.03× weighted kernel group · Cosmos3 causal Conv3D cat-pad CUDA fast path · SGLang #29281
- 1.06× denoise speedup · Sana packed self- and cross-attention GEMMs · SGLang #28393
- 1.015× denoise speedup · FLUX fused FeedForward GELU up-projection · SGLang #28166
- 1.13× denoise speedup · Qwen-Image norm-scale-shift · SGLang #27392
- 1.06× chat throughput · Qwen3-Next FP8 MoE · SGLang #26204
Results vary by hardware, model, and workload. See each linked pull request or repository for validation and reproduction details.
GOOD TO KNOW
Before you
request a kernel
Which GPUs are supported?
Currently, KDA only supports NVIDIA B200 and B300 GPUs. Document one or both targets in your request README and provide reproducible workloads for each selected target.
Should I open a pull request or an issue?
Files ready? Follow the submission guide and open a pull request to wishlist. Keep the details in your request README; the pull request description can be brief. If you need help preparing the files, discuss an idea in an issue. Complete files are optional for discussions, and an issue is never required before a pull request.
Does a merged request mean the kernel is ready?
Merging a pull request adds the request to the wishlist. It does not mean optimization is complete. Follow the original pull request for progress and result links, including after it is merged.
How do upvotes work?
Browse requests, find one you care about, and add a thumbs-up reaction to the pull request description. The list includes open and closed pull requests so merged requests remain visible. Comments are best used for new workload evidence or implementation context.
Will my request and results be public?
Yes—unless we agree otherwise in advance. For accepted requests, KDA may publish the submitted definitions, generated implementations, benchmarks, profiling data, and design notes. We credit submitters for their definitions and workloads. We are also collecting representative kernels and workloads to guide the development of the next generation of KDA. If your business requires private kernels, email ligengz@nvidia.com to discuss a collaboration through NVIDIA Kernel Factory.
How many kernels will KDA deliver?
KDA typically delivers one to three of the most-requested kernels each week, depending on the team’s workloads. Sihao Liu and Ligeng Zhu are currently self-paying the GPU and tokens, so capacity is limited. If you have a bunch of kernels to optimize, email ligengz@nvidia.com to setup an official collaboration through NVIDIA Kernel Factory.
Why might a request be returned or declined?
Requests may need revision when they cannot be reproduced, lack a reliable correctness reference, target unsupported hardware, are too vague for automated evaluation, or have unclear licensing.
YOUR BOTTLENECK, NEXT
What should
KDA build?
Bring the definition, real workloads, and your best-known baseline implementation.
Submit a request Discuss an idea Browse and upvote requests