Open source projects

Two repos.
One execution boundary.

Tensor compiles kernels into portable artifacts. AKAS compiles whole models and serves them. Production installs the runtime and a hardware driver — not a compiler.

2 projects · MIT licensed · pre-1.0

Tensor markKernel layer

Tensor

Portable kernels. Small runtime.

A GPU toolkit that does not assume PyTorch or CUDA. Compile TileLang kernels into portable .tbin artifacts, then execute them with a small runtime and your hardware driver.

  1. 01Write
  2. 02Compile
  3. 03Ship
  4. 04Run
prepared host launch
4.20 µsprepared host launch
vs matched TileLang
1.92×vs matched TileLang
vs a warmed Triton JIT
4.75×vs a warmed Triton JIT
vs a fixed TileLang schedule
0.97–1.96×vs a fixed TileLang schedule
Explore Tensor
AKAS markModel layer

AKAS

Compiler-first model inference.

Compose models in pure Python, fuse a whole model at compile time, and serve AOT-compiled .tbin programs without PyTorch or TileLang at runtime.

  1. 01Declare
  2. 02Compile
  3. 03Optimize
  4. 04Serve
prefill vs llama.cpp
+74%prefill vs llama.cpp
context lengths led on decode
8 / 8context lengths led on decode
tok/s at full context
79.55tok/s at full context
peak GPU memory
7.71 GiBpeak GPU memory
Explore AKAS
How they fit together

A compiler on the build machine. A small runtime in production.

Tensor owns

Compilation, artifacts, device execution.

AKAS owns

Model definitions, generation, scheduling, serving.

Together

One buffer and executable contract across three providers.