aitemplate.backend.cuda

target_def

CUDA target specialization

Classes:

`CUDA`([template_path, ait_static_files_path, ...])	CUDA target.
`FBCUDA`([arch, remote_cache_bytes])	FBCUDA target.

class aitemplate.backend.cuda.target_def.CUDA(template_path='/home/runner/work/AITemplate/AITemplate/3rdparty/cutlass', ait_static_files_path='/home/runner/work/AITemplate/AITemplate/3rdparty/../static', arch='80', cuda_version=None, **kwargs)[source]

CUDA target.

Methods:

`cc`()	Compiler for this target.
`compile_cmd`([executable])	Compile command string template for this target.
`dev_select_flag`()	Environment variable to select the device.
`get_device_compiler_options`()	Returns a list of options for the device compiler.
`get_host_compiler_options`()	Returns a list of options for the host compiler.
`get_include_directories`()	Returns a list of include directories for a compiler.
`select_minimal_algo`(algo_names)	Select the minimal algorithm from the list of algorithms.
`src_extension`()	Source file extension for this target.

cc()[source]

Compiler for this target.

Raises:: NotImplementedError – Need to be implemented by subclass.

compile_cmd(executable=False)[source]

Compile command string template for this target.

Parameters:: executable (bool, optional) – Whether the command with compile an executable object by default False
Raises:: NotImplementedError – Need to be implemented by subclass.

dev_select_flag()[source]

Environment variable to select the device.

Returns:: Environment variable to select the device.
Return type:: str

get_device_compiler_options() → List[str][source]

Returns a list of options for the device compiler.

Raises:: NotImplementedError – Need to be implemented by subclass.

get_host_compiler_options() → List[str][source]

Returns a list of options for the host compiler.

Raises:: NotImplementedError – Need to be implemented by subclass.

get_include_directories() → List[str][source]

Returns a list of include directories for a compiler.

Raises:: NotImplementedError – Need to be implemented by subclass.

select_minimal_algo(algo_names: List[str])[source]

Select the minimal algorithm from the list of algorithms.

This is used in CI to speed up the test without running actually profiling.

Parameters:: algo_names (List[str]) – All the available algorithm names for selection.

src_extension()[source]

Source file extension for this target.

Returns:: Source file extension for this target.
Return type:: str

class aitemplate.backend.cuda.target_def.FBCUDA(arch='80', remote_cache_bytes=None, **kwargs)[source]

FBCUDA target. Used in Meta internal env only.

Methods:

`binary_compile_cmd`()	There is no ld by default in the prod env.
`cc`()	Compiler for this target.
`compile_options`()	Options for compiling the target.
`get_device_compiler_options`()	Returns a list of options for the device compiler.
`get_host_compiler_options`()	Returns a list of options for the host compiler.
`get_include_directories`()	Returns a list of include directories for a compiler.
`in_ci_env`()	Check if the current environment is CI.
`postprocess_build_dir`(build_dir)	Postprocess a build directory, allows final modification of the build directory before building.
`remote_logger`(record)	Upload the record to Scuba table

binary_compile_cmd()[source]: There is no ld by default in the prod env. Instead, we use ld from the gvfs path.

cc()[source]

Compiler for this target.

Raises:: NotImplementedError – Need to be implemented by subclass.

compile_options()[source]

Options for compiling the target.

Return type:: str