blinkdl/rwkv-cuda

Analysis updated 2026-07-17 · repo last pushed 2025-12-10

★ 232CudaAudience · researcherComplexity · 4/5QuietSetup · hard

Mindmap

mindmap
  root((repo))
    What it does
      Speeds up RWKV
      Optimizes depthwise conv
      Runs on Nvidia GPUs
    Tech stack
      CUDA
      Python
      PyTorch
      Ninja
    Use cases
      Faster model training
      Faster chatbot responses
      GPU memory optimization
      Benchmark comparisons
    Audience
      ML researchers
      RWKV engineers
    Setup
      Run Python script
      Auto compiles kernels
      Requires Ninja

mindmap root((repo)) What it does Speeds up RWKV Optimizes depthwise conv Runs on Nvidia GPUs Tech stack CUDA Python PyTorch Ninja Use cases Faster model training Faster chatbot responses GPU memory optimization Benchmark comparisons Audience ML researchers RWKV engineers Setup Run Python script Auto compiles kernels Requires Ninja

What do people build with it?

USE CASE 1

Speed up RWKV model training by replacing the default depthwise convolution with a custom CUDA kernel up to 100x faster.

USE CASE 2

Reduce chatbot response latency by using optimized CUDA kernels during RWKV inference on Nvidia hardware.

USE CASE 3

Compare successive kernel optimization versions using the included raw benchmark numbers.

USE CASE 4

Cut hours off a full RWKV training run by shaving milliseconds off each forward pass.

What is it built with?

CUDAPythonPyTorchNinja

How does it compare?

	blinkdl/rwkv-cuda	yassa9/dvlt.cu	stablemarkk/hash256_miner
Stars	232	31	20
Language	Cuda	Cuda	Cuda
Last pushed	2025-12-10	—	—
Maintenance	Quiet	—	—
Setup difficulty	hard	hard	moderate
Complexity	4/5	5/5	4/5
Audience	researcher	researcher	developer

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · hard Time to first run · 1h+

Requires an Nvidia GPU, CUDA toolkit, and the Ninja build tool to compile kernels.

Not specified in the explanation.

Copy-paste prompts

Prompt 1

How do I install and compile the CUDA kernels in blinkdl/rwkv-cuda using the provided Python script and Ninja?

Prompt 2

Explain why the depthwise convolution kernel in blinkdl/rwkv-cuda is roughly 100x faster than PyTorch's default implementation.

Prompt 3

How do I swap the default RWKV depthwise convolution for blinkdl/rwkv-cuda's optimized kernel in my training script?

Prompt 4

What Nvidia GPU and CUDA toolkit version do I need to run blinkdl/rwkv-cuda's optimized kernels?

Frequently asked questions

What is rwkv-cuda?

A set of custom CUDA kernels that speed up RWKV language model operations on Nvidia GPUs, cutting a key computation from 94ms to under 1ms.

What language is rwkv-cuda written in?

Mainly Cuda. The stack also includes CUDA, Python, PyTorch.

Is rwkv-cuda actively maintained?

Quiet — no commits in 6-12 months (last push 2025-12-10).

What license does rwkv-cuda use?

Not specified in the explanation.

How hard is rwkv-cuda to set up?

Setup difficulty is rated hard, with roughly 1h+ to a first successful run.

Who is rwkv-cuda for?

Mainly researcher.

Open on GitHub → Explain another repo

This repo across BitVibe Labs

Verify against the repo before relying on details.