mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
Implements gather_kv_cache on the ROCm backend (was a day-one stub that bailed -> normal 'hanzo run -i' CRASHED on ROCm with SIGSEGV once prefix-cache became default). Tensor-ops un-paging (index_select over a source-index map; identical math to the CUDA/Metal kernels, off the hot path = once/prefill). Bit-exact test 5/5 (f16/f32/cast/empty); run -i default flags now coherent + clean exit + PagedAttention/prefix-cache active. No ml change.
23 lines
594 B
TOML
23 lines
594 B
TOML
[build-system]
|
|
requires = ["maturin==1.7"]
|
|
build-backend = "maturin"
|
|
|
|
[project]
|
|
name = "hanzo"
|
|
version = "1.3.15"
|
|
requires-python = ">=3.10"
|
|
classifiers = [
|
|
"Programming Language :: Rust",
|
|
"Programming Language :: Python :: Implementation :: CPython",
|
|
"Programming Language :: Python :: Implementation :: PyPy",
|
|
"License :: OSI Approved :: MIT License",
|
|
"Programming Language :: Python :: 3.10"
|
|
]
|
|
dynamic = ["description"]
|
|
|
|
[tool.maturin]
|
|
features = ["pyo3/extension-module"]
|
|
profile = "release"
|
|
include = ["hanzo.pyi", "py.typed"]
|
|
sdist-include = ["hanzo.pyi", "py.typed"]
|