Kernels on Hugging Face

Load Algenta compute kernels on Apple silicon straight from the Hugging Face Kernel Hub with the kernels library.

Algenta publishes its compute kernels to the Hugging Face Kernel Hub under the thyn-ai organization. Each repository is one kernel family: a compiled Mojo library, loaded in-process by the kernels library, with a published contract and documentation for every function.

Install

pip install kernels torch

from kernels import get_kernel

stats = get_kernel(
    "thyn-ai/algenta-stats",
    version=1,
    backend="cpu",
    trust_remote_code=["thyn-ai/algenta-stats"],
)
stats.ab_testing.z_test_proportions(50, 1000, 65, 1000)  # -> [-1.4408, 0.1496]  z statistic, p-value

Plain Python in, plain Python out. Lists, tuples, buffers and tensors are accepted wherever the contract expects a list; structured results are dictionaries. Every call is checked against the published contract before it reaches native code. An invalid call raises KernelError with a stable code, never a crash. help(stats.ab_testing) documents every function, with its defaults.

Families

Repository	What's inside	Backends
thyn-ai/algenta-stats	Statistics, probability, hypothesis testing, regression and sampling	CPU
thyn-ai/algenta-llm	LLM, attention, RoPE, embedding and KV-cache kernels, resident decode sessions	CPU · Metal

More families follow the same layout. Each card lists its modules with a line on each, and kernel.CONTRACT holds every signature.

Resident decode on Apple silicon

The Metal build of algenta-llm runs greedy Qwen3-family decode on the GPU. One session is open at a time; each step blocks until its token is computed and returns the token id.

llm = get_kernel("thyn-ai/algenta-llm", version=1, trust_remote_code=["thyn-ai/algenta-llm"])
with llm.DecodeSession("/path/to/converted-checkpoint", prompt_token_ids, max_new_tokens=32) as session:
    tokens = list(session)

prompt_token_ids come from the checkpoint's own tokenizer. The checkpoint directory is the converted form written by prepare_model_checkpoint(), described in Model-agnostic LLM inference. llm.decode(...) runs a whole session in one call; session.step() returns the full record for one token. Decode needs the Metal or CUDA build; on the CPU build DecodeSession raises KernelError("gpu_build_required").

Requirements

• Apple silicon: macOS 15 or later for CPU builds, macOS 26 or later for Metal.
• kernels 0.17 or later and PyTorch 2.5 to 2.14. PyTorch has to be installed: the loader picks the build for your PyTorch version. The kernels themselves never import it.
• trust_remote_code names the repositories you allow. kernels loads Hugging Face's trusted publishers by default; any other publisher needs its exact repository ID here.
• On a Mac the loader picks the Metal build when the family ships one and the system is macOS 26 or later. Pass backend="cpu" for CPU-only families, and for every family on macOS 15.

Linux x86-64 and arm64 builds ship from the Algenta release pipeline. Windows is not supported.

Terms

Every kernel repository carries the Algenta Community License 1.0: free for personal, research and open-source use, and for internal use at organizations with fewer than 50 employees and under $5M in annual revenue. Beyond that, a commercial license is required: [e-mail].

Next

Model-agnostic LLM inferenceRuntime library catalog