Group: GitHub Pedronahum

Sort by:Popular▼

✓Popular✓Newest

Typed tensors for Kotlin: rank, dtype and axis names live in the type system, so a transposed weight is a compile error.

DXIR — Tlaloc's differentiable IR, its reference interpreter, and the phi-calculus coarsener that rewrites loops before differentiation.

Typed tensors for Kotlin: rank, dtype and axis names live in the type system, so a transposed weight is a compile error.

DXIR — Tlaloc's differentiable IR, its reference interpreter, and the phi-calculus coarsener that rewrites loops before differentiation.

StableHLO and Shardy emission from DXIR: the portable artifact Tlaloc programs compile to.

The grad / forward / higher-order transformation surface, and the Tracer that captures a DXIR function at runtime. One AD engine, no value tape.

The grad / forward / higher-order transformation surface, and the Tracer that captures a DXIR function at runtime. One AD engine, no value tape.

The Tlaloc Gradle plugin (id io.github.pedronahum.tlaloc): applies the Tlaloc compiler plugin of the same version to every Kotlin/JVM compilation and maps the tlaloc { } block to its options.

KPTX: a Kotlin DSL, emitter, parser and ISA table for NVIDIA PTX, plus the kernel library built with it.

Typed multi-step workflows over Tlaloc programs, and export of decode graphs as a serving artifact: a manifest plus StableHLO bodies, read at serve time without a JVM.

CUDA driver bindings via the Foreign Function & Memory API — no JNI.

Bill of materials that aligns every Tlaloc artifact on one version.

The Tlaloc K2 compiler plugin: it rewrites grad { } into synthesized gradient code at compile time and turns shape and differentiability misuse into compile errors.

KPTX: a Kotlin DSL, emitter, parser and ISA table for NVIDIA PTX, plus the kernel library built with it.

Typed multi-step workflows over Tlaloc programs, and export of decode graphs as a serving artifact: a manifest plus StableHLO bodies, read at serve time without a JVM.

Layers, initializers, losses, optimizers and parameter handling over Tlaloc's typed tensors, with DiffKT's model surface as the parity target.

Layers, initializers, losses, optimizers and parameter handling over Tlaloc's typed tensors, with DiffKT's model surface as the parity target.

CUDA driver bindings via the Foreign Function & Memory API — no JNI.

IREE execution backend for Tlaloc's StableHLO artifacts.

IREE execution backend for Tlaloc's StableHLO artifacts.