Metadata-Version: 2.2
Name: taichi-forge-runtime
Version: 0.7.0
Summary: Platform native runtime libraries for Taichi Forge.
Author: Taichi Forge contributors
License: Apache-2.0
Classifier: Development Status :: 4 - Beta
Classifier: Topic :: Software Development :: Compilers
Classifier: Topic :: Multimedia :: Graphics
Classifier: Topic :: Games/Entertainment :: Simulation
Classifier: Intended Audience :: Science/Research
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Project-URL: Homepage, https://github.com/fancifulland2718/taichi-forge
Project-URL: Documentation, https://github.com/fancifulland2718/taichi-forge/blob/master/docs/forge/index.en.md
Project-URL: Releases, https://github.com/fancifulland2718/taichi-forge/releases
Project-URL: Issues, https://github.com/fancifulland2718/taichi-forge/issues
Requires-Python: <4.0,>=3.10
Description-Content-Type: text/markdown

# Taichi Forge Runtime

This distribution contains platform-native Taichi Forge runtime libraries and
runtime bitcode resources. It is installed as a dependency of the Python-facing
`taichi-forge` distribution.

Each release publishes one runtime wheel per supported platform, not one wheel
per CUDA version. Official wheels use the CUDA driver API and Vulkan loader;
they do not bundle a CUDA Toolkit runtime or vendor libraries such as cuBLAS,
cuSPARSE, cuSOLVER, or cuFFT. Those libraries are optional user-provided
providers and are loaded only when the corresponding hardware operation is
selected. Installing `taichi-forge` does not require a local CUDA Toolkit.

Standard Windows/Linux wheels also include the basic AMDGPU backend. Selecting
`ti.amdgpu` requires a compatible user-installed ROCm/HIP runtime, driver and
linker. HIP libraries are loaded optionally, never bundled or mandatory native
imports. CPU/CUDA/Vulkan use does not require ROCm.

The `taichi_forge_runtime.aot` module loads and executes exported kernels and
ordinary CGraphs and exported execution plans without the Python DSL, NumPy or
a CPython shim. Its Python implementation uses only the standard library and
the public C API in the same native core used by JIT. Importing the module does
not load native libraries or
probe devices; `aot.Runtime("cpu" | "cuda" | "vulkan")` selects the backend.
Existing wheels without this entrypoint need a matching newer runtime build.
`entry.prepare(...)` provides explicit fixed native execution: a CPU batch,
CUDA Graph or Vulkan fixed commands. Prepare once and replay with unchanged
bindings; array contents can change. Preparation executes no user kernels.

Exported plans can reconstruct supported native recordings, provider resources
and structure bindings. Runtime APIs also expose supported external-memory
imports. Availability depends on the selected backend, operation schema,
provider and resource contract; an installed provider does not expose every
function in its vendor library. Independent CUDA and fixed buffer-only Vulkan
execution domains additionally support declared per-submission scalar updates,
optional GPU timing and individual instance retirement. Vulkan requires enabled
timelines and available additional compute queues on the selected device.

This is not a compiler-free runtime: CPU/CUDA bitcode still needs bundled LLVM,
and GPU drivers may compile device code during loading. See
`docs/forge/aot_sdk.en.md` and `docs/forge/cuda_execution_domains.en.md` in the
source repository for deployment and submission contracts, and
`examples/aot_sdk/consume.py` for a source-free forward/explicit-gradient example.

The JIT `taichi_forge` API remains in the separate frontend distribution. Mixed
installations select one native binary and bitcode resource directory per
process. Independently created AOT and JIT runtimes own their own mutable state;
sharing a library does not make their array handles interchangeable.

The Windows wheel also includes the `taichi_runtime.lib` import library so the
CPython shim wheels can link against the already-published runtime package
during release builds.

## Release information / 发行信息

The 0.7 series is distributed through coordinated GitHub and PyPI releases.
[GitHub Releases](https://github.com/fancifulland2718/taichi-forge/releases) include
the same platform wheels, bilingual release notes, complete sources, public docs
and checksums. External vendor libraries remain separately installed dependencies.

0.7 系列通过 GitHub 与 PyPI 协调发布。GitHub Release 提供相同的平台 wheel、
中英双语发布说明、完整源码、公开文档和校验和。本包承载原生 runtime，通常由
`taichi-forge` 自动安装；外部厂商库仍由用户独立安装，不重复打包进 runtime。
发布状态以实际 Release 与 PyPI 文件为准。
