[lazyinit] combine lazy tensor with dtensor (#3204)

* [lazyinit] lazy tensor add distribute

* [lazyinit] refactor distribute

* [lazyinit] add test dist lazy init

* [lazyinit] add verbose info for dist lazy init

* [lazyinit] fix rnn flatten weight op

* [lazyinit] polish test

* [lazyinit] polish test

* [lazyinit] fix lazy tensor data setter

* [lazyinit] polish test

* [lazyinit] fix clean

* [lazyinit] make materialize inplace

* [lazyinit] refactor materialize

* [lazyinit] refactor test distribute

* [lazyinit] fix requires_grad

* [lazyinit] fix tolist after materialization

* [lazyinit] refactor distribute module

* [lazyinit] polish docstr

* [lazyinit] polish lazy init context

* [lazyinit] temporarily skip test

* [lazyinit] polish test

* [lazyinit] add docstr

colossalai/utils/model/experimental.py

@@ -1,11 +1,15 @@
-from typing import Callable, List, Optional, Union
+from types import MethodType
+from typing import Callable, Optional, Union
 
 import torch
+import torch.distributed as dist
 import torch.nn as nn
 from torch import Tensor
 from torch.utils._pytree import tree_map
 
 from colossalai.fx.profiler.tensor import MetaTensor
+from colossalai.tensor.d_tensor.d_tensor import DTensor
+from colossalai.tensor.d_tensor.layout import Layout
 
 # reference: https://pytorch.org/cppdocs/notes/tensor_creation.html
 _NORMAL_FACTORY = [
@@ -30,6 +34,11 @@ _NO_META_FACTORY = [
 
 _EARLY_MATERIALIZED_OPS = ['__getitem__', 'split']
 
+# If your intent is to change the metadata of a Tensor (such as sizes / strides / storage / storage_offset)
+# without autograd tracking the change, remove the .data / .detach() call and wrap the change in a `with torch.no_grad():` block.
+# These ops cannot be unwrapped using .data
+_CHANGE_META_OPS = ['_cudnn_rnn_flatten_weight', 'requires_grad_', '__get__']
+
 _LEGACY_TENSOR_CONSTRUCTOR = {
     'FloatTensor': torch.float,
     'DoubleTensor': torch.double,
@@ -43,6 +52,8 @@ _LEGACY_TENSOR_CONSTRUCTOR = {
     'BoolTensor': torch.bool,
 }
 
+_EMPTY_DATA = torch.empty(0)
+
 
 class _MyTensor(Tensor):
     """This class is only for correctness verification.
@@ -64,6 +75,29 @@ class _MyTensor(Tensor):
         return super().__torch_function__(func, types, args, kwargs)
 
 
+def _convert_cls(tensor: 'LazyTensor', target: torch.Tensor) -> torch.Tensor:
+    """Convert a lazy tensor's class to target's class, with target's data.
+
+    The reason why we change the class of a lazy tensor in-place is that this can easily handle shared modules/parameters, which is common in huggingface models.
+    If we create a new tensor and update the module by ``setattr(module, name, param)``, the shared parameters will not be updated. And we have to track all shared parameters and update them manually.
+
+    Args:
+        tensor (LazyTensor): the LazyTensor to be converted
+        target (torch.Tensor): target tensor
+
+    Returns:
+        torch.Tensor: the converted tensor
+    """
+    cls_to_become = nn.Parameter if isinstance(tensor, nn.Parameter) else torch.Tensor
+    tensor.__class__ = cls_to_become
+    tensor.data = target
+    tensor.requires_grad = target.requires_grad
+    # subclass of torch.Tensor does not have tolist() method
+    # overwrite this method after materialization or distribution
+    tensor.tolist = MethodType(torch.Tensor.tolist, target)
+    return tensor
+
+
 class LazyTensor(torch.Tensor):
     """A naive implementation of LazyTensor (https://arxiv.org/pdf/2102.13267.pdf).
 
@@ -112,14 +146,8 @@ class LazyTensor(torch.Tensor):
                 elem = func(*args, **{**kwargs, 'device': 'meta'})
                 meta_data = MetaTensor(elem, fake_device=device)
             elem = meta_data._tensor
-        r = torch.Tensor._make_wrapper_subclass(cls,
-                                                elem.size(),
-                                                strides=elem.stride(),
-                                                storage_offset=elem.storage_offset(),
-                                                dtype=elem.dtype,
-                                                layout=elem.layout,
-                                                device=elem.device,
-                                                requires_grad=elem.requires_grad)
+        # As a meta tensor cannot be modified __class__ to torch.Tensor, we should use an empty real tensor here
+        r = torch.Tensor._make_subclass(cls, _EMPTY_DATA, require_grad=elem.requires_grad)
         r._meta_data = meta_data
         return r
 
@@ -129,15 +157,28 @@ class LazyTensor(torch.Tensor):
         self._materialized_data: Optional[torch.Tensor] = concrete_data    # materialized data
 
     def materialize(self) -> torch.Tensor:
-        """Materialize the ``LazyTensor`` to ``torch.Tensor``.
+        """Materialize the ``LazyTensor`` to ``torch.Tensor`` by modifying __class__ (inplace).
 
         Returns:
-            torch.Tensor: The materialized tensor.
+            torch.Tensor: The materialized tensor (self).
         """
         target = self._materialize_data()
-        if isinstance(self, nn.Parameter):
-            target = nn.Parameter(target, requires_grad=self.requires_grad)
-        return target
+        self.clean()
+        return _convert_cls(self, target)
+
+    def distribute(self, layout: Layout) -> torch.Tensor:
+        """Distribute the ``LazyTensor`` to ``torch.Tensor`` by modifying __class__ (inplace), according to the layout.
+
+        Args:
+            layout (Layout): Distribution layout.
+
+        Returns:
+            torch.Tensor: The distributed tensor (self).
+        """
+        target = self._materialize_data()
+        self.clean()
+        local_tensor = DTensor(target, layout).local_tensor
+        return _convert_cls(self, local_tensor)
 
     def clean(self) -> None:
         """Clean all stored operations, meta data and materialized data, which prevents memory leaking. This should be called after all tensors are materialized.
@@ -216,6 +257,8 @@ class LazyTensor(torch.Tensor):
         is_inplace: bool = (func.__name__.endswith('_') and not (func.__name__.endswith('__'))
                             or func.__name__ == "__setitem__")
 
+        is_change_meta_op: bool = func.__name__ in _CHANGE_META_OPS
+
         if isinstance(func, torch._C.ScriptMethod):
             # FIXME(ver217): torch script functions are not verified
 
@@ -239,10 +282,10 @@ class LazyTensor(torch.Tensor):
                 if isinstance(x, LazyTensor):
                     if x._materialized_data is not None:
                         # for early materialized tensor, use its materialized data directly
-                        return x._materialized_data.data
+                        return x._materialized_data if is_change_meta_op else x._materialized_data.data
                     t = x if is_inplace else x.clone()
                     t._op_buffer.append((func, args, kwargs))
-                    meta = x._meta_data.data
+                    meta = x._meta_data if is_change_meta_op else x._meta_data.data
                     meta_to_lazy[meta] = t
                     return meta
                 return x
@@ -290,13 +333,36 @@ class LazyTensor(torch.Tensor):
 
     @data.setter
     def data(self, other: 'LazyTensor'):
+        """This is sightly different from oringinal `data` setter.
+
+        E.g.:
+            >>> a = torch.randn(3, 3) # a is a Tensor
+            >>> b = torch.rand(2, 2)
+            >>> a.data = b
+            >>> b.add_(1)   # this will affect a
+            >>> x = torch.randn(3, 3) # x is a LazyTensor
+            >>> y = torch.rand(2, 2) # y is a LazyTensor
+            >>> x.data = y
+            >>> y.add_(1)   # this will not affect x
+
+        """
         if other is self:
             return
-        # TODO(ver217): to avoid infinity recursion, do early materialization
-        self._materialized_data = other._materialize_data()
+
+        self._op_buffer.append(other._factory_method)
+
+        def replace(x):
+            if x is other:
+                return self
+            return x
+
+        for func, args, kwargs in other._op_buffer:
+            self._op_buffer.append((func, tree_map(replace, args), tree_map(replace, kwargs)))
 
     def tolist(self) -> list:
-        t = self.materialize()
+        # Though self.__class__ is modified to torch.Tensor, in C++ side, it is still a subclass of torch.Tensor
+        # And subclass of torch.Tensor does not have tolist() method
+        t = self._materialize_data()
         return t.tolist()
 
     def __hash__(self):
@@ -421,71 +487,84 @@ class LazyInitContext:
             setattr(torch, name, orig)
 
     @staticmethod
-    def materialize(module: torch.nn.Module, verbose: bool = False):
-        """Initialize all ``nn.Parameter`` from ``LazyTensor``.
+    def materialize(module: nn.Module, verbose: bool = False) -> nn.Module:
+        """Initialize all ``nn.Parameter`` from ``LazyTensor``. This function will modify the module in-place.
 
         Args:
-            module (torch.nn.Module): Target ``nn.Module``
+            module (nn.Module): Target ``nn.Module``
             verbose (bool): Whether to print lazy initialization rate. Defaults to False.
         """
-        if verbose:
-            param_cnt = 0
-            param_lazy_cnt = 0
-            buf_cnt = 0
-            buf_lazy_cnt = 0
-            non_lazy_numel = 0
-
-        # do post cleaning to handle shared parameter
-        visited_lazy_tensors: List[LazyTensor] = []
-        # handle shared module
-        visited_modules = set()
-
-        @torch.no_grad()
-        def init_recursively(module: nn.Module):
-            nonlocal param_cnt, param_lazy_cnt, buf_cnt, buf_lazy_cnt, non_lazy_numel
-            # recursively initialize the module
-            for mod in module.children():
-                if id(mod) not in visited_modules:
-                    visited_modules.add(id(mod))
-                    init_recursively(mod)
-
-            # initialize tensors directly attached to the current module
-            for name, param in module.named_parameters(recurse=False):
-                if verbose:
-                    param_cnt += 1
-                    if getattr(param, '_materialized_data', False) is None:
-                        # if no _materialized_data attr, the tensor is not lazy
-                        param_lazy_cnt += 1
-                    else:
-                        non_lazy_numel += param.numel()
-                if hasattr(param, 'materialize'):
-                    # TODO(ver217): apex layers cannot be captured
-                    visited_lazy_tensors.append(param)
-                    setattr(module, name, param.materialize())
-
-            for name, buf in module.named_buffers(recurse=False):
-                if verbose:
-                    buf_cnt += 1
-                    if getattr(buf, "_materialized_data", False) is None:
-                        # if no _materialized_data attr, the tensor is not lazy
-                        buf_lazy_cnt += 1
-                    else:
-                        non_lazy_numel += buf.numel()
-                if hasattr(buf, 'materialize'):
-                    # TODO(ver217): apex layers cannot be captured
-                    visited_lazy_tensors.append(buf)
-                    setattr(module, name, buf.materialize())
 
-        init_recursively(module)
+        def apply_fn(name: str, p: LazyTensor):
+            p.materialize()
+
+        return _apply_to_lazy_module(module, apply_fn, verbose)
+
+    @staticmethod
+    def distribute(module: nn.Module, layout_dict: dict, verbose: bool = False) -> nn.Module:
+        """Distribute all ``nn.Parameter`` from ``LazyTensor``. This function will modify the module in-place.
+
+        Args:
+            module (nn.Module): Target ``nn.Module``
+            layout_dict (dict): Dict of layout for each parameter/buffer. The key is the parameter/buffer name, and the value is the layout.
+            verbose (bool, optional): Whether to print lazy initialization rate. Defaults to False.
+        """
+
+        def apply_fn(name: str, p: LazyTensor):
+            p.distribute(layout_dict[name])
+
+        return _apply_to_lazy_module(module, apply_fn, verbose)
+
 
-        for t in visited_lazy_tensors:
-            t.clean()
+def _apply_to_lazy_module(module: nn.Module,
+                          apply_fn: Callable[[str, torch.Tensor], None],
+                          verbose: bool = False) -> nn.Module:
+    if verbose:
+        # verbose info
+        param_cnt = 0
+        param_lazy_cnt = 0
+        buf_cnt = 0
+        buf_lazy_cnt = 0
+        total_numel = 0
+        non_lazy_numel = 0
+
+    for name, p in module.named_parameters():
+        if verbose:
+            param_cnt += 1
+            total_numel += p.numel()
+            if getattr(p, '_materialized_data', False) is None:
+                # if no _materialized_data attr, the tensor is not lazy
+                param_lazy_cnt += 1
+            else:
+                non_lazy_numel += p.numel()
+        if isinstance(p, LazyTensor):
+            apply_fn(name, p)
 
+    for name, buf in module.named_buffers():
         if verbose:
-            print(f'Param lazy rate: {param_lazy_cnt}/{param_cnt}')
-            print(f'Buffer lazy rate: {buf_lazy_cnt}/{buf_cnt}')
-            print(f'Non-lazy numel: {non_lazy_numel} ({non_lazy_numel/1024**2:.3f} M)')
-        return module
+            buf_cnt += 1
+            total_numel += buf.numel()
+            if getattr(buf, "_materialized_data", False) is None:
+                # if no _materialized_data attr, the tensor is not lazy
+                buf_lazy_cnt += 1
+            else:
+                non_lazy_numel += buf.numel()
+        if isinstance(buf, LazyTensor):
+            apply_fn(name, buf)
+
+    if verbose:
+        non_lazy_numel_ratio = non_lazy_numel / total_numel * 100 if non_lazy_numel != 0 else 0
+        _print_rank_0(f'Param lazy rate: {param_lazy_cnt}/{param_cnt}')
+        _print_rank_0(f'Buffer lazy rate: {buf_lazy_cnt}/{buf_cnt}')
+        _print_rank_0(
+            f'Non lazy numel: {non_lazy_numel} ({non_lazy_numel/1024**2:.3f} M), ratio: {non_lazy_numel_ratio}%')
+
+    return module
+
+
+def _print_rank_0(*args, **kwargs):
+    if not dist.is_initialized() or dist.get_rank() == 0:
+        print(*args, **kwargs)
 
 
 def _is_int_tuple(args) -> bool:

tests/test_utils/test_lazy_init/test_distribute.py

@@ -0,0 +1,110 @@
+from functools import partial
+from typing import Optional
+
+import pytest
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+
+import colossalai
+from colossalai.device.device_mesh import DeviceMesh
+from colossalai.tensor.d_tensor.layout import Layout
+from colossalai.tensor.d_tensor.sharding_spec import ShardingSpec
+from colossalai.testing import parameterize, rerun_if_address_is_in_use
+from colossalai.utils import free_port
+from colossalai.utils.common import print_rank_0
+from colossalai.utils.model.experimental import LazyInitContext, LazyTensor, _MyTensor
+from tests.kit.model_zoo import model_zoo
+
+# from utils import assert_dist_model_equal, set_seed
+
+
+def find_shard_dim(shape: torch.Size) -> Optional[int]:
+    for dim, size in enumerate(shape):
+        if size % 2 == 0:
+            return dim
+
+
+def make_layout(device_mesh: DeviceMesh, original_tensor: torch.Tensor) -> Layout:
+    shard_dim = find_shard_dim(original_tensor.shape)
+    dim_partition_dict = {shard_dim: [0]} if shard_dim is not None else {}
+    target_sharding_spec = ShardingSpec(dim_size=original_tensor.dim(), dim_partition_dict=dim_partition_dict)
+    layout = Layout(device_mesh=device_mesh,
+                    device_type=torch.device('cuda'),
+                    sharding_spec=target_sharding_spec,
+                    entire_shape=original_tensor.shape)
+    return layout
+
+
+def _get_current_name(prefix: str, name: str) -> str:
+    return f'{prefix}.{name}'.lstrip('.')
+
+
+def generate_layout_dict(model: nn.Module, device_mesh: DeviceMesh) -> dict:
+    layout_dict = {}
+
+    @torch.no_grad()
+    def generate_recursively(module: nn.Module, prefix: str = ''):
+        # recursively initialize the module
+        for name, mod in module.named_children():
+            generate_recursively(mod, prefix=_get_current_name(prefix, name))
+
+        # initialize tensors directly attached to the current module
+        for name, param in module.named_parameters(recurse=False):
+            if isinstance(param, LazyTensor):
+                layout = make_layout(device_mesh, param)
+                layout_dict[_get_current_name(prefix, name)] = layout
+
+        for name, buf in module.named_buffers(recurse=False):
+            if isinstance(buf, LazyTensor):
+                layout = make_layout(device_mesh, buf)
+                layout_dict[_get_current_name(prefix, name)] = layout
+
+    generate_recursively(model)
+
+    return layout_dict
+
+
+@parameterize('subset', ['torchvision', 'diffusers', 'timm', 'transformers', 'torchaudio', 'deepfm', 'dlrm'])
+def run_dist_lazy_init(subset, seed: int = 42):
+    sub_model_zoo = model_zoo.get_sub_registry(subset)
+    device_mesh = DeviceMesh(torch.Tensor([0, 1, 2, 3]), (2, 2), init_process_group=True)
+    # FIXME(ver217): uncomment this line
+    # _MyTensor._pre_op_fn = lambda *args: set_seed(seed)
+    # LazyTensor._pre_op_fn = lambda *args: set_seed(seed)
+
+    for name, entry in sub_model_zoo.items():
+        # TODO(ver217): lazy init does not support weight norm, skip these models
+        if name in ('torchaudio_wav2vec2_base', 'torchaudio_hubert_base'):
+            continue
+        print_rank_0(name)
+        model_fn, data_gen_fn, output_transform_fn, model_attr = entry
+        ctx = LazyInitContext(tensor_cls=_MyTensor)
+        with ctx:
+            model = model_fn()
+        ctx = LazyInitContext()
+        with ctx:
+            deferred_model = model_fn()
+        layout_dict = generate_layout_dict(deferred_model, device_mesh)
+        ctx.distribute(deferred_model, layout_dict, verbose=True)
+        # FIXME(ver217): uncomment this line
+        # assert_dist_model_equal(model, deferred_model, layout_dict)
+
+
+def run_dist(rank, world_size, port) -> None:
+    colossalai.launch({}, rank=rank, world_size=world_size, host='localhost', port=port)
+    run_dist_lazy_init()
+
+
+# FIXME(ver217): temporarily skip this test since torch 1.11 does not fully support meta tensor
+@pytest.mark.skip
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+def test_dist_lazy_init():
+    world_size = 4
+    run_func = partial(run_dist, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_dist_lazy_init()

tests/test_utils/test_lazy_init/utils.py

@@ -4,6 +4,7 @@ from typing import Any, Callable, Optional, Tuple
 import numpy as np
 import torch
 
+from colossalai.tensor.d_tensor.layout_converter import to_global
 from colossalai.utils.model.experimental import LazyInitContext, LazyTensor, _MyTensor
 from tests.kit.model_zoo.registry import ModelAttribute
 
@@ -67,3 +68,18 @@ def check_lazy_init(entry: TestingEntry, seed: int = 42, verbose: bool = False,
         assert_forward_equal(model, deferred_model, data_gen_fn, output_transform_fn)
     if verbose:
         print(f'{model.__class__.__name__} pass')
+
+
+def assert_dist_model_equal(model: torch.nn.Module, distributed_model: torch.nn.Module, layout_dict: dict) -> None:
+    state = model.state_dict()
+    distributed_state = distributed_model.state_dict()
+
+    assert len(state) == len(distributed_state), f'len {len(state)} vs {len(distributed_state)}'
+
+    for (n1, t1), (n2, t2) in zip(state.items(), distributed_state.items()):
+        assert n1 == n2
+        t1 = t1.cuda()
+        t2 = t2.cuda()
+        if n2 in layout_dict:
+            t2 = to_global(t2, layout_dict[n2])
+        assert torch.equal(t1, t2), f'{n1} {t1} vs {t2}'