Speed up docs build (#44635)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-06-05 14:51:44 +00:00
committed by GitHub
parent c66b19800b
commit a80af24356
32 changed files with 234 additions and 159 deletions
@@ -123,7 +123,8 @@ def initialize_online_processing(layer: torch.nn.Module):
Called by either `initialize_layerwise_reload` or an online quantization scheme,
prevents double wrapping in the case of online quantization + reloading
:param layer: layer whose parameter weight loaders will be wrapped
Args:
layer: layer whose parameter weight loaders will be wrapped
"""
info = get_layerwise_info(layer)
@@ -222,8 +223,9 @@ def finalize_layerwise_processing(model: torch.nn.Module, model_config: ModelCon
This function should be applied after `initialize_layerwise_reload` is applied
unwrap the layerwise weight loaders.
:param model: model to finalize processing for
:param model_config: config needed for applying processing to attention layers
Args:
model: model to finalize processing for
model_config: config needed for applying processing to attention layers
"""
if hasattr(model, "_original_do_torchao_reload"):
model._do_torchao_reload = model._original_do_torchao_reload
@@ -175,9 +175,12 @@ def get_numel_loaded(
"""
Determine how many elements would be loaded by a weight loader call.
:param weight loader: used to load weights
:param args: bound arguments to weight loader
:return: number of elements loaded by the weight loader, the return value of the
Args:
weight_loader: used to load weights
args: bound arguments to weight loader
Returns:
number of elements loaded by the weight loader, the return value of the
weight loader
"""
with CopyCounter() as counter:
@@ -20,9 +20,12 @@ def sanitize_layer_refs(tensor: torch.Tensor, layer: torch.nn.Module) -> torch.T
tensors will reference layers, and the WeakKeyDictionary will never evict entries,
even when the model is deleted.
:param tensor: tensor to be sanitized
:param layer: layer whose references should be removed
:return: sanitized tensor
Args:
tensor: tensor to be sanitized
layer: layer whose references should be removed
Returns:
sanitized tensor
"""
for key, value in tensor.__dict__.items():
if isinstance(value, MethodType) and value.__self__ is layer:
@@ -38,10 +41,12 @@ def restore_layer_refs(tensor: torch.Tensor, layer: torch.nn.Module) -> torch.Te
Used by `restore_layer_on_meta` to add back layer references, allowing for proper
weight loading.
:param tensor: tensor to be sanitized
:param layer: layer whose references should be removed
:return: sanitized tensor
Args:
tensor: tensor to be sanitized
layer: layer whose references should be removed
Returns:
sanitized tensor
"""
for key, value in tensor.__dict__.items():
if isinstance(value, MethodType) and value.__self__ is layer_ref_sentinel:
@@ -49,8 +49,11 @@ def has_device_tensors(bound_args: BoundArguments) -> bool:
"""
Return True if the loaded weights exist on an accelerator device
:param bound_args: args to load weights
:return: True if weights are on accelerator device
Args:
bound_args: args to load weights
Returns:
True if weights are on accelerator device
"""
return any(
isinstance(value, torch.Tensor) and value.device.type not in ("meta", "cpu")
@@ -62,8 +65,11 @@ def get_info_size(info: LayerReloadingInfo) -> int:
"""
Calculate the number of bytes used by loaded weights for a given layer
:param info: layerwise info to get size of
:return: number of bytes used by loaded weights
Args:
info: layerwise info to get size of
Returns:
number of bytes used by loaded weights
"""
return sum(
value.nbytes