Convolution configuration¶

class torch_lattice.nn.functional.conv.conv_config.ConvConfig(dataflow=Dataflow.ImplicitGEMM, ifsort=False, kmap_mode='hashmap_on_the_fly', downsample_mode='spconv', split_mask_num=1, split_mask_num_bwd=3, wgrad_split_k='auto', IGEMM_center_only=False, epsilon=0.0, mm_thresh=0, FOD_fusion=False)[source]¶

Bases: Mapping[str, Any]

Validated sparse-convolution execution policy.

Parameters:

dataflow (Dataflow)
ifsort (bool)
kmap_mode (str)
downsample_mode (str)
split_mask_num (int)
split_mask_num_bwd (int)
wgrad_split_k (int | str)
IGEMM_center_only (bool)
epsilon (float)
mm_thresh (int)
FOD_fusion (bool)

dataflow: Dataflow¶

ifsort: bool¶

kmap_mode: str¶

downsample_mode: str¶

split_mask_num: int¶

split_mask_num_bwd: int¶

wgrad_split_k: int | str¶

IGEMM_center_only: bool¶

epsilon: float¶

mm_thresh: int¶

FOD_fusion: bool¶

copy()[source]¶

Return type:: ConvConfig

class torch_lattice.nn.functional.conv.conv_config.Dataflow(*values)[source]¶

Bases: Enum

ImplicitGEMM = 0¶

GatherScatter = 1¶

FetchOnDemand = 2¶

torch_lattice.nn.functional.conv.conv_config.clear_global_conv_config()[source]¶

Return type:: None

torch_lattice.nn.functional.conv.conv_config.get_default_conv_config(conv_mode=ConvMode.mode0, training=False)[source]¶

Return type:

ConvConfig

Parameters:

conv_mode (ConvMode)
training (bool)

torch_lattice.nn.functional.conv.conv_config.get_global_conv_config()[source]¶

Return type:: ConvConfig | None

torch_lattice.nn.functional.conv.conv_config.set_global_conv_config(conv_config)[source]¶

Return type:: None
Parameters:: conv_config (ConvConfig | Mapping[str, Any])

class torch_lattice.nn.functional.conv.conv_mode.ConvMode(*values)[source]¶

Bases: Enum

mode0 = 0¶

mode1 = 1¶

mode2 = 2¶

torch_lattice.nn.functional.conv.conv_mode.get_conv_mode()[source]¶

Return type:: ConvMode

torch_lattice.nn.functional.conv.conv_mode.set_conv_mode(conv_mode)[source]¶

Return type:: None
Parameters:: conv_mode (int | ConvMode)