taichi-forge 0.4.25

Taichi Forge - a community-maintained fork of the Taichi Programming Language (import name: taichi_forge).

Taichi Forge

中文版

Taichi Forge is a community-maintained fork of Taichi. It keeps the Python-embedded DSL model of vanilla Taichi, while carrying modern toolchain, backend, graph, native algorithm, cache, and display-path work for simulation and rendering workloads.

Install

pip install -U taichi-forge

import taichi_forge as ti

ti.init(arch=ti.cuda)

The package name is taichi-forge; the Python import name is taichi_forge. Forge does not overwrite the upstream taichi package. Code that must import vanilla taichi unchanged should keep using upstream Taichi or an explicit compatibility shim.

Compatibility Baseline

Vanilla Taichi 1.7.4 is the main public-API compatibility reference for this fork. Forge keeps the Taichi DSL programming model, but the release line is independent from upstream Taichi version numbers.

Area	Policy
Public DSL	Supported Taichi-style kernels, fields, ndarrays, sparse SNodes, graph builders, and AOT APIs keep source-compatible semantics.
Package identity	Forge installs as taichi-forge and imports as taichi_forge; it does not claim the upstream taichi package name.
Backends	CPU, CUDA, and Vulkan are first-class Forge targets. Backend-specific additions are documented explicitly.
Experimental paths	Experimental features are marked by API name, option, warning, or documentation. They are not treated as vanilla compatibility promises.
Bugfix-only uploads	If a PyPI upload only fixes packaging, crash, cache, or documentation problems without changing intended feature semantics, the latest fixed patch in that release line is the authoritative version.

Changelog-Oriented Highlights

Current release baseline: 0.4.2.

The README intentionally describes user-visible changes and compatibility boundaries instead of detailed benchmark numbers.

0.1.x: Toolchain Modernization

• Moved the fork onto a modern LLVM 20 based toolchain.
• Added support targets for Python 3.10 through 3.14.
• Updated the Windows build path for current MSVC/Visual Studio environments.
• Kept the vanilla Taichi DSL import style under the separate taichi_forge package name.

0.2.x: Compile and Cache Infrastructure

• Added compile-tier controls such as ti.init(compile_tier=...) and per-kernel @ti.kernel(opt_level=...).
• Added batch precompile helpers: ti.compile_kernels(...) and the ti.parallel_compile(...) alias.
• Added ti.compile_profile() and ti cache warmup ... for compile-time diagnosis and offline-cache warmup.
• Split reusable frontend/source parsing state from backend-specific compiled artifacts where reuse is safe. Backend-specific cache entries stay isolated so switching arch does not overwrite another backend's compiled artifact.

See Forge API reference, Compile and cache guide, and Forge options.

0.3.x: Vulkan Sparse SNode and Native Sort

• Added Vulkan sparse SNode support beyond vanilla 1.7.4's dense/root-only Vulkan path. The documented public target includes pointer, bitmasked, and dynamic SNodes on Vulkan.
• Added experimental hash SNode support on CPU, CUDA, and Vulkan with fixed capacity semantics.
• Kept quantized Vulkan paths behind explicit experimental gates.
• Added the Forge-only stable sort dispatcher ti.algorithms.sort(...), while keeping vanilla-compatible ti.algorithms.parallel_sort(...).
• Treated bugfix-only sparse-pool and package uploads as superseded by the latest fixed patch in the same release line.

See Vulkan sparse SNode, Hash SNode, and Parallel sort API.

0.4.x: Graph, Native Algorithms, Cache, and Display Submission

• Modernized graph execution below the public graph-builder API while keeping GraphBuilder.dispatch, sequential graphs, compile, Graph.run, and AOT CGraph as the user-facing model.
• Added support for Forge-defined DSL native algorithm nodes in graph replay. This is not a public arbitrary native callback API.
• Broadened native algorithm coverage beyond sort. Public algorithm entry points include PrefixSumExecutor.run(), experimental_compact(), experimental_reduce(), experimental_histogram(), experimental_transform(), experimental_gather(), experimental_scatter(), experimental_scatter_add(), experimental_bucket_builder(), and experimental_grouped_reduce(), with reusable workspace objects for repeated calls.
• Formalized canvas.set_image() as a display-frame submission path and added ti.ui.DisplayFrame plus Canvas.submit_frame(...) for display-ready host, texture, and packed u32 frame inputs.
• Added display statistics APIs so engines can distinguish accepted, submitted, dropped, and reused display frames.
• Fixed some issues in vanilla Taichi as well.

See Forge API reference, Graph upgrade notes, Native algorithms, and Display frame submission.

Public Documentation

English public docs:

• Building Forge wheels
• Bug fixes compared with vanilla Taichi
• Forge API reference
• Forge options
• Compile and cache guide
• Vulkan sparse SNode
• Hash SNode
• Parallel sort API
• Native algorithms
• Graph upgrade notes
• Display frame submission

Build From Source

Forge wheels are built with scikit-build-core, matching .github/workflows/publish_pypi.yml. The PyPI-style build supports Windows x86_64 and Ubuntu 22.04 x86_64 for Python 3.10 through 3.14, with Vulkan, OpenGL, CUDA, LLVM, and the C API enabled.

Use Building Forge wheels for the exact Windows and Ubuntu package list, LLVM 20 setup, Vulkan SDK setup, CMAKE_ARGS, and python -I -m build --wheel --no-isolation commands.

Known Boundaries

• Forge is a fork with its own release track. Do not assume a Forge version maps to an upstream Taichi release number.
• Native algorithm APIs with experimental_ in the name are public but still allowed to evolve more conservatively than long-standing vanilla APIs.
• Strict cross-device zero-copy rendering is not a blanket guarantee. Some display routes are near-zero-copy or staging-based depending on source backend and resource ownership.
• Public compatibility means source compatibility for supported paths, not preserving every upstream implementation detail.

License

Taichi Forge follows the Apache-2.0 license inherited from upstream Taichi. See LICENSE.