bridge.utils.cuda_graph

Module Contents

Functions

_as_list
_member_name
_member_name_list
_member_names
_supports_cuda_graph_modules
_module_value
cuda_graph_module_names	Return configured per-layer CUDA graph module names.
set_cuda_graph_modules	Set per-layer CUDA graph modules using the current MCore API when available.
clear_cuda_graph_modules	Clear per-layer CUDA graph modules using the active MCore API.
set_full_iteration_cuda_graph	Enable full-iteration CUDA graph capture using the current MCore API.
has_cuda_graph_module	Return whether a per-layer CUDA graph module is enabled.
is_full_iteration_cuda_graph	Return whether config enables full-iteration CUDA graph capture.
uses_local_cuda_graph_manager	Return whether Bridge should create a local MCore CudaGraphManager.

API

bridge.utils.cuda_graph._as_list(value: Any) → list[Any]

bridge.utils.cuda_graph._member_name(value: Any) → str

bridge.utils.cuda_graph._member_name_list(value: Any) → list[str]

bridge.utils.cuda_graph._member_names(value: Any) → set[str]

bridge.utils.cuda_graph._supports_cuda_graph_modules(config: Any) → bool

bridge.utils.cuda_graph._module_value(name: str)

bridge.utils.cuda_graph.cuda_graph_module_names(config: Any) → list[str]

Return configured per-layer CUDA graph module names.

bridge.utils.cuda_graph.set_cuda_graph_modules(

config: Any,

modules: Any,

) → None

Set per-layer CUDA graph modules using the current MCore API when available.

bridge.utils.cuda_graph.clear_cuda_graph_modules(config: Any) → None

Clear per-layer CUDA graph modules using the active MCore API.

bridge.utils.cuda_graph.set_full_iteration_cuda_graph(config: Any) → None

Enable full-iteration CUDA graph capture using the current MCore API.

bridge.utils.cuda_graph.has_cuda_graph_module(config: Any, module: Any) → bool

Return whether a per-layer CUDA graph module is enabled.

Supports both the current MCore cuda_graph_modules API and the deprecated cuda_graph_scope values still present in older Bridge configs.

bridge.utils.cuda_graph.is_full_iteration_cuda_graph(config: Any) → bool

Return whether config enables full-iteration CUDA graph capture.

bridge.utils.cuda_graph.uses_local_cuda_graph_manager(config: Any) → bool

Return whether Bridge should create a local MCore CudaGraphManager.