kaldera 0.0.1

Compute library on top of volkano — tensor allocation, kernel dispatch, stream-based async

kaldera

Compute library for GPU work, on top of volkano. Provides tensor allocation, descriptor reflection, kernel dispatch, and (eventually) stream-based async — all in pure Python with zero compiled code in the wheel.

Status

Early bootstrap. The public Tensor / Stream / Kernel API is being built. Currently exposes (as transient re-exports of internal modules):

• kaldera.vk_raii — RAII wrappers over Vulkan handles (Instance, Device, Buffer, ShaderModule, …).
• kaldera.shader_utils — compile_glsl(source, stage), reflect_descriptor_bindings(spirv).

See examples/ for current usage patterns. These will be rewritten against the public API once it lands.

Install

pip install kaldera

(Pulls in volkano automatically.)