Crate pepita 

Pepita: Sovereign AI Distributed Computing Primitives

Pepita provides minimal kernel interfaces and distributed computing primitives for running Sovereign AI workloads. It combines low-level kernel interfaces (ublk, io_uring, blk-mq) with high-level distributed execution (scheduler, executor, pool).

Design Principles (Iron Lotus Framework)

• First-Principles Rust: Zero external dependencies in kernel mode
• Pure Rust Sovereignty: 100% auditable, zero C/C++ dependencies
• Work-Stealing Scheduler: Blumofe-Leiserson algorithm [1]
• Toyota Way Quality: Jidoka, Poka-yoke, Genchi Genbutsu
• Certeza Testing: 95% coverage, 80% mutation score

Features

• std (default): Standard library support for testing and distributed mode
• kernel: True #![no_std] mode for kernel integration

Example (Kernel Interfaces)

use pepita::ublk::{UblkCtrlCmd, UblkIoDesc, UblkIoCmd};
use pepita::io_uring::{IoUringSqe, IoUringCqe};

// Verify ABI compatibility with Linux kernel
assert_eq!(core::mem::size_of::<UblkCtrlCmd>(), 32);
assert_eq!(core::mem::size_of::<UblkIoDesc>(), 24);
assert_eq!(core::mem::size_of::<IoUringSqe>(), 64);

Example (Distributed Computing)

use pepita::pool::Pool;
use pepita::task::Task;
use pepita::executor::Backend;

// Create execution pool
let pool = Pool::builder()
    .cpu_workers(4)
    .build()?;

// Submit task
let task = Task::binary("./worker")
    .args(vec!["--input", "data.bin"])
    .backend(Backend::Cpu)
    .build();

let result = pool.submit(task)?;

References

[1] Blumofe & Leiserson (1999). “Scheduling Multithreaded Computations by Work Stealing.” Journal of the ACM.

Re-exports

pub use blk_mq::BlockOps;

pub use blk_mq::Request;

pub use blk_mq::RequestOp;

pub use blk_mq::TagSetConfig;

pub use error::KernelError;

pub use error::Result;

pub use io_uring::IoUringCqe;

pub use io_uring::IoUringSqe;

pub use io_uring::IORING_OP_URING_CMD;

pub use memory::DmaBuffer;

pub use memory::DmaDirection;

pub use memory::PageAllocator;

pub use memory::Pfn;

pub use memory::PhysAddr;

pub use memory::VirtAddr;

pub use ublk::UblkCtrlCmd;

pub use ublk::UblkIoCmd;

pub use ublk::UblkIoDesc;

pub use ublk::UBLK_F_CMD_IOCTL_ENCODE;

pub use ublk::UBLK_F_SUPPORT_ZERO_COPY;

pub use ublk::UBLK_F_USER_COPY;

pub use ublk::UBLK_U_CMD_ADD_DEV;

pub use ublk::UBLK_U_CMD_DEL_DEV;

pub use ublk::UBLK_U_CMD_GET_PARAMS;

pub use ublk::UBLK_U_CMD_SET_PARAMS;

pub use ublk::UBLK_U_CMD_START_DEV;

pub use ublk::UBLK_U_CMD_STOP_DEV;

pub use ublk::UBLK_U_IO_COMMIT_AND_FETCH_REQ;

pub use ublk::UBLK_U_IO_FETCH_REQ;

Modules

blk_mq

Block multi-queue (blk-mq) interface.

constants

Kernel constants

error

Error types for Pepita kernel interfaces.

executorstd

Executor backends for task execution.

faultstd

Fault tolerance mechanisms for distributed execution.

gpustd

GPU: Compute Shader Execution

io_uring

io_uring kernel interface.

memory

Memory management types.

poolstd

High-level execution pool for distributed task execution.

schedulerstd

Work-stealing scheduler implementing the Blumofe-Leiserson algorithm.

simdstd

SIMD: Vectorized Operations

taskstd

Task types for distributed execution.

transportstd

Message transport for distributed task execution.

ublk

ublk (Userspace Block Device) kernel interface.

virtiostd

Virtio: Virtual I/O Devices

vmmstd

VMM: Lightweight Virtual Machine Monitor

zramstd

zram: Compressed RAM Block Device