Skip to content

vllm.model_executor.model_loader.reload.meta

Functions:

CopyCounter

Bases: TorchDispatchMode

Tracks total number of elements modified with copy_.

Useful for keeping track of weight loading where underlying weights can be arbitrarily transformed (such as with narrow) before calling copy.

Note: Assumes that copy kwargs are not used.

Source code in vllm/model_executor/model_loader/reload/meta.py 
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
class CopyCounter(TorchDispatchMode):
    """
    Tracks total number of elements modified with `copy_`.

    Useful for keeping track of weight loading where underlying weights can be
    arbitrarily transformed (such as with `narrow`) before calling copy.

    Note: Assumes that copy kwargs are not used.
    """

    def __init__(self):
        super().__init__()
        self.copied_numel = 0

    def __torch_dispatch__(self, func, types, args=(), kwargs=None):
        if kwargs is None:
            kwargs = {}

        if func is torch.ops.aten.copy_.default:
            assert args[0].numel() == args[1].numel()
            self.copied_numel += args[0].numel()

        return func(*args, **kwargs)

get_numel_loaded(weight_loader, args)

Determine how many elements would be loaded by a weight loader call.

Parameters:

Returns:

  • int –

    number of elements loaded by the weight loader, the return value of the

  • object –

    weight loader

Source code in vllm/model_executor/model_loader/reload/meta.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
def get_numel_loaded(
    weight_loader: Callable, args: inspect.BoundArguments
) -> tuple[int, object]:
    """
    Determine how many elements would be loaded by a weight loader call.

    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:
        return_value = weight_loader(*args.args, **args.kwargs)
    return counter.copied_numel, return_value

materialize_layer(layer, info)

Materialize all meta tensors in a layer to actual tensors.

Source code in vllm/model_executor/model_loader/reload/meta.py 
136
137
138
139
140
141
142
143
144
def materialize_layer(layer: torch.nn.Module, info: LayerReloadingInfo):
    """Materialize all meta tensors in a layer to actual tensors."""
    if layer.__class__.__name__ in SKIP_MODULES:
        return

    with info.restore_device:
        for name, tensor in get_layer_tensors(layer).items():
            if name not in SKIP_TENSORS and tensor.is_meta:
                setattr(layer, name, materialize_meta_tensor(tensor))

materialize_meta_tensor(meta_tensor)

Materialize a meta tensor into an actual tensor on the current device. Should be called within the torch device context for the given rank.

Source code in vllm/model_executor/model_loader/reload/meta.py 
43
44
45
46
47
48
49
50
51
52
53
54
55
56
def materialize_meta_tensor(meta_tensor: torch.Tensor) -> torch.Tensor:
    """
    Materialize a meta tensor into an actual tensor on the current device.
    Should be called within the torch device context for the given rank.
    """
    tensor = torch.empty_strided(
        size=tuple(meta_tensor.size()),
        stride=tuple(meta_tensor.stride()),
        dtype=meta_tensor.dtype,
        requires_grad=False,
    )
    tensor.__class__ = meta_tensor.__class__
    tensor.__dict__ = meta_tensor.__dict__.copy()
    return tensor

restore_layer_on_meta(layer, info)

Restore a layer to model format with tensors on the meta device

Source code in vllm/model_executor/model_loader/reload/meta.py 
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
def restore_layer_on_meta(layer: torch.nn.Module, info: LayerReloadingInfo):
    """Restore a layer to model format with tensors on the meta device"""
    if layer.__class__.__name__ in SKIP_MODULES:
        return

    for name in get_layer_tensors(layer):
        if name not in SKIP_TENSORS:
            delattr(layer, name)

    restore_params, restore_buffers = info.restore_metadata
    for name, param in restore_params.items():
        if name not in SKIP_TENSORS:
            param = restore_layer_refs(param, layer)
            layer.register_parameter(name, param)

    for name, buffer in restore_buffers.items():
        if name not in SKIP_TENSORS:
            buffer = restore_layer_refs(buffer, layer)
            layer.register_buffer(name, buffer)

to_meta_tensor(tensor)

Convert a tensor to a meta tensor while preserving class and attributes.

Source code in vllm/model_executor/model_loader/reload/meta.py 
35
36
37
38
39
40
def to_meta_tensor(tensor: torch.Tensor) -> torch.Tensor:
    """Convert a tensor to a meta tensor while preserving class and attributes."""
    meta_tensor = tensor.data.to("meta")
    meta_tensor.__class__ = tensor.__class__
    meta_tensor.__dict__ = tensor.__dict__.copy()
    return meta_tensor