Struct FastVideoInstaller 

pub struct FastVideoInstaller { /* private fields */ }

FastVideo installer.

Implementations

impl FastVideoInstaller

pub fn new(config: FastVideoConfig) -> Self

pub fn build_commands(&self) -> Vec<ShellCommand>

Build the sequence of commands to install FastVideo with ROCm support.

pub fn mkdir_build_dir(&self) -> ShellCommand

pub fn git_clone(&self) -> ShellCommand

pub fn git_checkout(&self) -> ShellCommand

pub fn git_submodule_init(&self) -> ShellCommand

Initialize git submodules (cutlass, ThunderKittens).

pub fn patch_cmake(&self) -> ShellCommand

Patch source files to remove flash_attn_rocm.cpp from the build.

The feature/rocm-gfx11-support branch includes flash_attn_rocm.cpp which depends on AMD’s composable_kernel MHA device implementations. The CK submodule at the required commit (5ff2d64) is incompatible with the current HIP compiler — it fails with missing CUDA builtins (threadIdx/blockDim), missing standard library includes, and template deduction errors in device_grouped_mha_bwd_*_v2.hpp.

Three patches are applied:

1. Remove flash_attn_rocm.cpp from CMakeLists.txt source list
2. Remove #include "attention/flash_attn_rocm.cpp" and the forward declarations for mha_varlen_fwd/mha_varlen_bwd from common_extension.cpp (lines 4-23)
3. Remove pybind11 bindings for flash_attn_varlen_fwd/_bwd from common_extension.cpp

The rest of fastvideo-kernel (gemm, norm, quant) compiles correctly.

pub fn install_build_deps(&self) -> ShellCommand

Install build dependencies (scikit-build-core, cmake, ninja).

Uses python -m pip install with --break-system-packages when needed, matching the pattern used by Triton, AITER, and other source-build installers.

pub fn pip_install_kernel(&self) -> ShellCommand

Build and install fastvideo-kernel with ROCm support.

Sets CMAKE_ARGS for ROCm:

• -DCMAKE_HIP_ARCHITECTURES=<gpu_arch>
• -DFASTVIDEO_KERNEL_BUILD_TK=OFF (ThunderKittens not supported on ROCm)
• -DGPU_BACKEND=ROCM

Uses pip install --no-build-isolation --no-deps . matching the upstream build.sh, with --no-deps added to enforce the No-CUDA hard-prime tenet (prevents pip from resolving/overriding the ROCm torch). Runtime deps are installed explicitly and filtered via the unified nvidia_blocklist.

pub fn cleanup(&self) -> ShellCommand

Trait Implementations

impl Clone for FastVideoInstaller

fn clone(&self) -> FastVideoInstaller

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FastVideoInstaller

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for FastVideoInstaller

fn default() -> Self

Returns the “default value” for a type.