[shardformer] supported fused qkv checkpoint (#4073)

colossalai/shardformer/layer/_operation.py

@@ -1,5 +1,6 @@
 import torch
 import torch.distributed as dist
+import torch.nn.functional as F
 
 try:
     import fused_mix_prec_layer_norm_cuda
@@ -46,6 +47,53 @@ class FusedLayerNormAffineFunction1D(torch.autograd.Function):
         return grad_input, grad_weight, grad_bias, None, None
 
 
+class MatmulWithAsyncCommunication(torch.autograd.Function):
+    """
+    Linear layer execution with asynchronous communication in backprop.
+    """
+
+    @staticmethod
+    def forward(ctx, input_, weight, bias, process_group, async_grad_allreduce):
+        ctx.save_for_backward(input_, weight)
+        ctx.use_bias = bias is not None
+        ctx.process_group = process_group
+        ctx.async_grad_allreduce = async_grad_allreduce
+
+        output = torch.matmul(input_, weight)
+
+        if bias is not None:
+            output = output + bias
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, weight = ctx.saved_tensors
+        use_bias = ctx.use_bias
+
+        total_input = input
+        grad_input = grad_output.matmul(weight.T)
+        grad_output = grad_output.contiguous()
+        # Convert the tensor shapes to 2D for execution compatibility
+        if len(grad_output.shape) > 2:
+            grad_output = grad_output.view(-1, grad_output.shape[-1])
+            total_input = total_input.view(-1, total_input.shape[-1])
+
+        if ctx.async_grad_allreduce:
+            # Asynchronous all-reduce
+            handle = dist.all_reduce(grad_input, group=ctx.process_group, async_op=True)
+            # Delay the start of weight gradient computation shortly (3us) to have
+            # all-reduce scheduled first and have GPU resources allocated
+            _ = torch.empty(1, device=grad_output.device) + 1
+
+        grad_weight = total_input.t().matmul(grad_output)
+        grad_bias = grad_output.sum(dim=0) if use_bias else None
+
+        if ctx.async_grad_allreduce:
+            handle.wait()
+
+        return grad_input, grad_weight, grad_bias, None, None, None
+
+
 class LinearWithAsyncCommunication(torch.autograd.Function):
     """
     Linear layer execution with asynchronous communication in backprop.
@@ -58,9 +106,10 @@ class LinearWithAsyncCommunication(torch.autograd.Function):
         ctx.process_group = process_group
         ctx.async_grad_allreduce = async_grad_allreduce
 
-        output = torch.matmul(input_, weight.t())
         if bias is not None:
-            output = output + bias
+            output = F.linear(input_, weight, bias)
+        else:
+            output = F.linear(input_, weight)
         return output
 
     @staticmethod
@@ -114,7 +163,7 @@ class _SplitForwardGatherBackward(torch.autograd.Function):
         return _gather(grad_output, ctx.dim, ctx.process_group), None, None
 
 
-class _ReduceInput(torch.autograd.Function):
+class _ReduceForward(torch.autograd.Function):
     """
     All-reduce the input from the model parallel region.
 
@@ -132,6 +181,25 @@ class _ReduceInput(torch.autograd.Function):
         return grad_output, None
 
 
+class _ReduceBackward(torch.autograd.Function):
+    """
+    All-reduce the input from the model parallel region.
+
+    Args:
+        input_: input matrix.
+        parallel_mode: parallel mode.
+    """
+
+    @staticmethod
+    def forward(ctx, input_, process_group):
+        ctx.process_group = process_group
+        return input_
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _reduce(grad_output, ctx.process_group), None
+
+
 def _reduce(input_, process_group):
     # skip if only one rank involved
     if dist.get_world_size(process_group) == 1:
@@ -198,6 +266,10 @@ class _GatherForwardSplitBackward(torch.autograd.Function):
         return _split(grad_output, ctx.dim, ctx.process_group), None, None
 
 
+def matmul_with_async_comm(input_, weight, bias, process_group, async_grad_allreduce):
+    return MatmulWithAsyncCommunication.apply(input_, weight, bias, process_group, async_grad_allreduce)
+
+
 def linear_with_async_comm(input_, weight, bias, process_group, async_grad_allreduce):
     return LinearWithAsyncCommunication.apply(input_, weight, bias, process_group, async_grad_allreduce)
 
@@ -210,5 +282,9 @@ def split_forward_gather_backward(input_, dim, process_group):
     return _SplitForwardGatherBackward.apply(input_, dim, process_group)
 
 
-def reduce_input(input_, process_group):
-    return _ReduceInput.apply(input_, process_group)
+def reduce_forward(input_, process_group):
+    return _ReduceForward.apply(input_, process_group)
+
+
+def reduce_backward(input_, process_group):
+    return _ReduceBackward.apply(input_, process_group)

colossalai/shardformer/layer/embedding.py

@@ -15,7 +15,7 @@ from colossalai.nn import init as init
 from colossalai.nn.layer.utils import divide
 from colossalai.tensor.d_tensor.api import shard_colwise, shard_rowwise, sharded_tensor_to_param
 
-from ._operation import gather_forward_split_backward, reduce_input
+from ._operation import gather_forward_split_backward, reduce_forward
 from .parallel_module import ParallelModule
 from .utils import create_randomizer_with_offset
 
@@ -276,5 +276,5 @@ class VocabParallelEmbedding1D(ParallelModule):
         # Mask the output embedding.
         output_parallel[input_mask, :] = 0.
         # Reduce across all the model parallel GPUs.
-        output = reduce_input(output_parallel, self.process_group)
+        output = reduce_forward(output_parallel, self.process_group)
         return output

colossalai/shardformer/layer/linear.py

@@ -15,12 +15,11 @@ from torch.nn.parameter import Parameter
 from colossalai.nn import init as init
 from colossalai.nn.layer.utils import divide
 from colossalai.tensor.d_tensor import shard_colwise, shard_rowwise, sharded_tensor_to_param
-from colossalai.utils.cuda import get_current_device
 
 from ._operation import (
     gather_forward_split_backward,
     linear_with_async_comm,
-    reduce_input,
+    reduce_forward,
     split_forward_gather_backward,
 )
 from .parallel_module import ParallelModule
@@ -148,9 +147,10 @@ class Linear1D_Col(ParallelModule):
         assert input_.shape[-1] == self.weight.shape[-1], \
             'Invalid shapes in Linear1D_Col forward: input={}, weight={}. Expected last dim of input {}.'.format(
                 input_.shape, self.weight.shape, self.weight.shape[-1])
+
         # Set up backprop all-reduce.
-        # input_parallel = reduce_grad(input_, ParallelMode.PARALLEL_1D)
         input_parallel = input_
+
         # Matrix multiply.
         bias = self.bias if not self.skip_bias_add else None
         output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True)
@@ -209,17 +209,14 @@ class Linear1D_Row(ParallelModule):
         self.parallel_input = parallel_input
         self.skip_bias_add = skip_bias_add
         self.process_group = process_group
+        self.num_partitions = dist.get_world_size(self.process_group)
 
         if skip_bias_add and not bias:
             raise ValueError('cannot skip bias addition if bias is None')
 
         # Parameters.
         # Initialize weight.
-        if device is None:
-            device = get_current_device()
-
         factory_kwargs = {'device': device, 'dtype': dtype}
-
         weight = torch.empty(self.out_features, self.in_features, **factory_kwargs)
         sharded_weight = shard_colwise(weight, self.process_group)
         self.weight = sharded_tensor_to_param(sharded_weight)
@@ -327,8 +324,7 @@ class Linear1D_Row(ParallelModule):
                 output = torch.cat(output_parallel_list, dim=-1)
         else:
             output_parallel = F.linear(input_, self.weight)
-            # output_parallel = linear_with_async_comm(input_, self.weight, None, ParallelMode.PARALLEL_1D, False)
-            output = reduce_input(output_parallel, self.process_group)
+            output = reduce_forward(output_parallel, self.process_group)
 
         if not self.skip_bias_add:
             if self.bias is not None:
@@ -336,5 +332,3 @@ class Linear1D_Row(ParallelModule):
             return output
         else:
             return output, self.bias
-            return output, self.bias
-            return output, self.bias

colossalai/shardformer/layer/linear_conv.py

@@ -14,13 +14,18 @@ from torch.nn.parameter import Parameter
 
 from colossalai.nn import init as init
 from colossalai.nn.layer.utils import divide
-from colossalai.tensor.d_tensor.api import shard_colwise, shard_rowwise
-from colossalai.utils.cuda import get_current_device
+from colossalai.tensor.d_tensor.api import (
+    customized_distributed_tensor_to_param,
+    distribute_tensor_with_customization,
+    shard_rowwise,
+    sharded_tensor_to_param,
+)
 
 from ._operation import (
     gather_forward_split_backward,
-    linear_with_async_comm,
-    reduce_input,
+    matmul_with_async_comm,
+    reduce_backward,
+    reduce_forward,
     split_forward_gather_backward,
 )
 from .parallel_module import ParallelModule
@@ -29,11 +34,69 @@ from .utils import create_randomizer_with_offset
 __all__ = ['LinearConv1D_Col', 'LinearConv1D_Row']
 
 
+def split_fused_qkv(qkv: torch.Tensor, n_fused: int, process_group: ProcessGroup):
+    """
+    The fused qkv tensor looks like [Q1, Q2, K1, K2, V1, V2], this function will split them into [Q1, K1, V1] and [Q2, K2, V2].
+    """
+    # get the number of slice for the fused qkv
+    rank = dist.get_rank(group=process_group)
+    world_size = dist.get_world_size(group=process_group)
+    order = torch.arange(world_size * n_fused)
+
+    # split the fused qkv
+    # from
+    # [Q, K, V]
+    # to
+    # [Q1, Q2, K1, K2, V1, V2]
+    weight_chunks = torch.chunk(qkv, world_size * n_fused, dim=-1)
+
+    # rearrange the slice into the final order
+    # from
+    # [Q1, Q2, K1, K2, V1, V2]
+    # to
+    # [Q1, K1, V1], [Q2, K2, V2]
+    weight_chunks_of_current_rank = [weight_chunks[i] for i in order[rank::world_size]]
+    weight_of_current_rank = torch.cat(weight_chunks_of_current_rank, dim=-1)
+    return weight_of_current_rank
+
+
+def gather_fused_qkv(qkv: torch.Tensor, n_fused: int, process_group: ProcessGroup):
+    """
+    The splitted qkv tensor looks like [Q1, K1, V1] and [Q2, K2, V2], this function will gather them into [Q1, Q2, K1, K2, V1, V2].
+    """
+    world_size = dist.get_world_size(group=process_group)
+
+    # gather the tensors
+    # from
+    # [Q1, K1, V1], [Q2, K2, V2]
+    # to
+    # [Q1, K1, V1, Q2, K2, V2]
+    origin_device = qkv.device
+    qkv = qkv.cuda()
+    gather_list = [torch.zeros_like(qkv) for _ in range(world_size)]
+    dist.all_gather(gather_list, qkv, group=process_group)
+    gather_weight = torch.cat(gather_list, dim=-1)
+    gather_weight = gather_weight.to(origin_device)
+    qkv = qkv.to(origin_device)
+
+    # rearrange the tensor slices
+    # from
+    # [Q1, K1, V1, Q2, K2, V2]
+    # to
+    # [Q1, Q2, K1, K2, V1, V2]
+    weight_chunks = torch.chunk(gather_weight, world_size * n_fused, dim=-1)
+    reordered_chunk_list = []
+    for i in range(n_fused):
+        reordered_chunk_list.extend(weight_chunks[i::n_fused])
+    reordered_gather_weight = torch.cat(reordered_chunk_list, dim=-1)
+    return reordered_gather_weight
+
+
 class LinearConv1D_Col(ParallelModule):
     r"""Linear layer with column parallelism.
 
     The linear layer is defined as :math:`Y = XA + b`. A is parallelized along
-    its second dimension as :math:`A = [A_1, ..., A_p]`. This layer is used to fit `Conv1D` layer in gpt2 of huggingface.
+    its second dimension as :math:`A = [A_1, ..., A_p]`. This layer is used to fit `Conv1D` layer (Fused QKV) in gpt2 of huggingface.
 
     Args:
         in_features (int): size of each input sample.
@@ -41,6 +104,7 @@ class LinearConv1D_Col(ParallelModule):
         bias (bool, optional): If set to ``False``, the layer will not learn an additive bias, defaults to ``True``.
         dtype (`torch.dtype`): The dtype of parameters, defaults to None.
         device (`torch.device`): The device of parameters, defaults to None.
+        n_fused (int): The number items fused, defaults to 3 (QKV).
         process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None.
         gather_output (bool, optional): If true, call all-gather on output and make Y available
                     to all GPUs, otherwise, every GPU will have its output
@@ -63,8 +127,10 @@ class LinearConv1D_Col(ParallelModule):
                  dtype: torch.dtype = None,
                  device: torch.device = None,
                  process_group: ProcessGroup = None,
+                 async_communication: bool = False,
                  gather_output: bool = False,
                  skip_bias_add: bool = False,
+                 n_fused: int = 3,
                  weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
                  bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1)):
         super().__init__()
@@ -75,23 +141,34 @@ class LinearConv1D_Col(ParallelModule):
         self.gather_output = gather_output
         self.skip_bias_add = skip_bias_add
         self.device = device
+        self.n_fused = n_fused
         self.process_group = process_group
-        self.num_partitions = dist.get_world_size(self.process_group)
+        self.async_communication = async_communication
 
         if skip_bias_add and not bias:
             raise ValueError('cannot skip bias addition if bias is None')
 
-        self.out_features_per_partition = divide(out_features, self.num_partitions)
-
         # Parameters.
         # Initialize weight.
-        if device is None:
-            device = get_current_device()
         factory_kwargs = {'device': device, 'dtype': dtype}
-        self.weight = Parameter(torch.empty(self.out_features_per_partition, self.in_features, **factory_kwargs))
+        weight = torch.empty(self.in_features, self.out_features, **factory_kwargs)
+
+        def shard_fn(tensor):
+            return split_fused_qkv(tensor, self.n_fused, self.process_group)
+
+        def gather_fn(tensor):
+            return gather_fused_qkv(tensor, 3, self.process_group)
+
+        with torch.no_grad():
+            sharded_weight = distribute_tensor_with_customization(weight, shard_fn, gather_fn)
+        self.weight = customized_distributed_tensor_to_param(sharded_weight)
 
         if bias:
-            self.bias = Parameter(torch.empty(self.out_features_per_partition, **factory_kwargs))
+            bias = torch.empty(self.out_features, **factory_kwargs)
+
+            with torch.no_grad():
+                sharded_bias = distribute_tensor_with_customization(bias, shard_fn, gather_fn)
+            self.bias = customized_distributed_tensor_to_param(sharded_bias)
         else:
             self.bias = None
 
@@ -103,7 +180,7 @@ class LinearConv1D_Col(ParallelModule):
         self.reset_parameters(weight_initializer, bias_initializer)
 
     @staticmethod
-    def from_native_module(module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], n_fused: int,
+    def from_native_module(module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]], n_fused: int,
                            *args, **kwargs) -> ParallelModule:
         r"""
         Convert a huggingface layer `Conv1D` in gpt2 to a parallelized linear layer.
@@ -135,29 +212,12 @@ class LinearConv1D_Col(ParallelModule):
 
         # TODO: copy the sharded weights
         with torch.no_grad():
-            # the weigh to the linear layer is a transpose
-            # thus shard on row is equal to shard on column
-
-            # first rearange the order of weight and bias
-            world_size = dist.get_world_size(group=process_group)
-            order = torch.arange(world_size * n_fused)
-            new_order = []
-            for i in range(world_size):
-                new_order.append(order[i::world_size])
-            new_order = torch.cat(new_order)
-
-            weight_chunks = torch.chunk(module.weight.data, world_size * n_fused, dim=1)
-            rearanged_weight_chunks = [weight_chunks[i] for i in new_order]
-            rearanged_weight = torch.cat(rearanged_weight_chunks, dim=1)
-            sharded_weight = shard_colwise(rearanged_weight, process_group)
-            linear_1d.weight.data.copy_(sharded_weight.T.contiguous())
+            sharded_weight = split_fused_qkv(module.weight.data, n_fused=n_fused, process_group=process_group)
+            linear_1d.weight.data.copy_(sharded_weight.data)
 
             if bias:
-                bias_chunks = torch.chunk(module.bias.data, world_size * n_fused, dim=0)
-                rearanged_bias_chunks = [bias_chunks[i] for i in new_order]
-                rearanged_bias = torch.cat(rearanged_bias_chunks, dim=0)
-                sharded_bias = shard_colwise(rearanged_bias, process_group)
-                linear_1d.bias.copy_(sharded_bias.contiguous())
+                sharded_bias = split_fused_qkv(module.bias.data, n_fused=n_fused, process_group=process_group)
+                linear_1d.bias.data.copy_(sharded_bias.data)
 
         return linear_1d
 
@@ -169,15 +229,18 @@ class LinearConv1D_Col(ParallelModule):
                 bias_initializer(self.bias, fan_in=fan_in)
 
     def forward(self, input_: Tensor) -> Tuple[Tensor, Tensor]:
-        assert input_.shape[-1] == self.weight.shape[-1], \
+        assert input_.shape[-1] == self.weight.shape[0], \
             'Invalid shapes in Linear1D_Col forward: input={}, weight={}. Expected last dim of input {}.'.format(
                 input_.shape, self.weight.shape, self.weight.shape[-1])
         # Set up backprop all-reduce.
-        # input_parallel = reduce_grad(input_, ParallelMode.PARALLEL_1D)
-        input_parallel = input_
+        input_parallel = reduce_backward(input_, self.process_group)
+        # input_parallel = input_
+
         # Matrix multiply.
         bias = self.bias if not self.skip_bias_add else None
-        output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True)
+
+        output_parallel = matmul_with_async_comm(input_parallel, self.weight, bias, self.process_group,
+                                                 self.async_communication)
 
         if self.gather_output:
             # All-gather across the partitions.
@@ -192,7 +255,8 @@ class LinearConv1D_Col(ParallelModule):
 
 
 class LinearConv1D_Row(ParallelModule):
-    r""" Linear layer with row parallelism
+    r""" Linear layer with row parallelism.
+    This layer is used to fit `Conv1D` layer (Fused QKV) in gpt2 of huggingface.
 
     Args:
         in_features (int): size of each input sample.
@@ -243,11 +307,10 @@ class LinearConv1D_Row(ParallelModule):
 
         # Parameters.
         # Initialize weight.
-        if device is None:
-            device = get_current_device()
-
         factory_kwargs = {'device': device, 'dtype': dtype}
-        self.weight = Parameter(torch.empty(self.out_features, self.input_size_per_partition, **factory_kwargs))
+        weight = torch.empty(self.in_features, self.out_features, **factory_kwargs)
+        sharded_weight = shard_rowwise(weight, self.process_group)
+        self.weight = sharded_tensor_to_param(sharded_weight)
 
         if self.stream_chunk_num > 1:
             # TODO() work for inference only
@@ -295,7 +358,7 @@ class LinearConv1D_Row(ParallelModule):
             # the weigh to the linear layer is a transpose
             # thus shard on col is equal to shard on row
             sharded_weight = shard_rowwise(module.weight.data, process_group)
-            linear_1d.weight.data.copy_(sharded_weight.T.contiguous())
+            linear_1d.weight.data.copy_(sharded_weight.data)
 
             if bias:
                 linear_1d.bias.copy_(module.bias.data)
@@ -325,12 +388,12 @@ class LinearConv1D_Row(ParallelModule):
     def forward(self, input_: Tensor) -> Tensor:
         # Set up backprop all-reduce.
         if self.parallel_input:
-            assert input_.shape[-1] == self.weight.shape[-1], \
+            assert input_.shape[-1] == self.weight.shape[0], \
                 'Invalid shapes in Linear1D_Row forward: input={}, weight={}. Expected last dim of input {}.'.format(
                 input_.shape, self.weight.shape, self.weight.shape[-1])
             input_ = input_
         else:
-            assert divide(input_.shape[-1], self.num_partitions) == self.weight.shape[-1], \
+            assert divide(input_.shape[-1], self.num_partitions) == self.weight.shape[0], \
                 'Invalid shapes in Linear1D_Row forward: input={}, weight={}. Expected last dim of input {}.'.format(
                 input_.shape, self.weight.shape, self.weight.shape[-1] * self.num_partitions)
             input_ = split_forward_gather_backward(input_, dim=-1, process_group=self.process_group)
@@ -342,7 +405,7 @@ class LinearConv1D_Row(ParallelModule):
                 output_parallel_list = [None for i in range(self.stream_chunk_num)]
                 handle_list = []
                 for i in range(self.stream_chunk_num):
-                    output_parallel_list[i] = F.linear(input_, self.weight_list[i])
+                    output_parallel_list[i] = torch.matmul(input_, self.weight_list[i])
                     handle = torch.distributed.all_reduce(output_parallel_list[i],
                                                           group=self.process_group,
                                                           async_op=True)
@@ -352,9 +415,8 @@ class LinearConv1D_Row(ParallelModule):
                     handle.wait()
                 output = torch.cat(output_parallel_list, dim=-1)
         else:
-            output_parallel = F.linear(input_, self.weight)
-            # output_parallel = linear_with_async_comm(input_, self.weight, None, ParallelMode.PARALLEL_1D, False)
-            output = reduce_input(output_parallel, self.process_group)
+            output_parallel = torch.matmul(input_, self.weight)
+            output = reduce_forward(output_parallel, self.process_group)
 
         if not self.skip_bias_add:
             if self.bias is not None:

colossalai/shardformer/layer/parallel_module.py

@@ -12,11 +12,14 @@ from torch.nn.modules.module import _EXTRA_STATE_KEY_SUFFIX, Module
 
 from colossalai.tensor.d_tensor import (
     distribute_tensor,
+    distribute_tensor_with_customization,
     get_device_mesh,
     get_sharding_spec,
+    is_customized_distributed_tensor,
     is_distributed_tensor,
     sharded_tensor_to_param,
     to_global,
+    to_global_for_customized_distributed_tensor,
 )
 
 __all__ = ['ParallelModule']
@@ -54,9 +57,10 @@ class ParallelModule(nn.Module, ABC):
         for name, param in self._parameters.items():
             if param is not None:
                 param_ = param if keep_vars else param.detach()
-
                 if is_distributed_tensor(param_):
                     destination[prefix + name] = to_global(param_)
+                elif is_customized_distributed_tensor(param_):
+                    destination[prefix + name] = to_global_for_customized_distributed_tensor(param_)
                 else:
                     destination[prefix + name] = param_
 
@@ -124,6 +128,8 @@ class ParallelModule(nn.Module, ABC):
                     sharding_spec = get_sharding_spec(param)
                     sharded_tensor = distribute_tensor(input_param, device_mesh, sharding_spec)
                     input_param = sharded_tensor_to_param(sharded_tensor)
+                elif is_customized_distributed_tensor(param):
+                    input_param = distribute_tensor_with_customization(input_param, param.shard_fn, param.gather_fn)
 
                 # This is used to avoid copying uninitialized parameters into
                 # non-lazy modules, since they dont have the hook to do the checks

colossalai/tensor/d_tensor/__init__.py

@@ -1,10 +1,13 @@
 from .api import (
     compute_global_numel,
+    customized_distributed_tensor_to_param,
     distribute_tensor,
+    distribute_tensor_with_customization,
     get_device_mesh,
     get_global_shape,
     get_layout,
     get_sharding_spec,
+    is_customized_distributed_tensor,
     is_distributed_tensor,
     is_sharded,
     redistribute,
@@ -12,6 +15,7 @@ from .api import (
     shard_rowwise,
     sharded_tensor_to_param,
     to_global,
+    to_global_for_customized_distributed_tensor,
 )
 from .layout import Layout
 from .sharding_spec import ShardingSpec
@@ -19,6 +23,6 @@ from .sharding_spec import ShardingSpec
 __all__ = [
     'is_distributed_tensor', 'distribute_tensor', 'to_global', 'is_sharded', 'shard_rowwise', 'shard_colwise',
     'sharded_tensor_to_param', 'compute_global_numel', 'get_sharding_spec', 'get_global_shape', 'get_device_mesh',
-    'redistribute', 'get_layout'
-    'Layout', 'ShardingSpec'
+    'redistribute', 'get_layout', 'is_customized_distributed_tensor', 'distribute_tensor_with_customization',
+    'to_global_for_customized_distributed_tensor', 'customized_distributed_tensor_to_param', 'Layout', 'ShardingSpec'
 ]

colossalai/tensor/d_tensor/api.py

@@ -305,3 +305,130 @@ def get_sharding_spec(dtensor: torch.Tensor) -> ShardingSpec:
     """
     assert is_distributed_tensor(dtensor), 'The input tensor is not a distributed tensor.'
     return dtensor.dist_layout.sharding_spec
+
+
+# ======================================================
+# Some sharding does not obey the SPMD style
+# e.g. Fused QKV layer in GPT2
+# we support customize sharding with the following APIs
+# ======================================================
+def is_customized_distributed_tensor(tensor: torch.Tensor):
+    """
+    Check whether the given tensor is a customized distributed tensor.
+
+    Args:
+        tensor (torch.Tensor): The tensor to be checked.
+
+    Returns:
+        bool: Whether the given tensor is a customized distributed tensor.
+    """
+    return hasattr(tensor, 'shard_fn') and hasattr(tensor, 'gather_fn')
+
+
+def _hijack_detach_and_clone_for_customized_distributed_tensor(dtensor: torch.Tensor) -> torch.Tensor:
+    """
+    Hijack the detach and clone methods of the tensor to make sure the dist_layout is copied.
+
+    Args:
+        tensor (torch.Tensor): The tensor to be hijacked.
+
+    Returns:
+        torch.Tensor: The hijacked tensor.
+    """
+    dtensor._old_detach = dtensor.detach
+    dtensor._old_clone = dtensor.clone
+
+    def new_detach(self):
+        t_ = self._old_detach()
+        t_.shard_fn = self.shard_fn
+        t_.gather_fn = self.gather_fn
+        return t_
+
+    def new_clone(self, *args, **kwargs):
+        t_ = self._old_clone(*args, **kwargs)
+        t_.shard_fn = self.shard_fn
+        t_.gather_fn = self.gather_fn
+        return t_
+
+    # bind the new methods to the tensor
+    dtensor.detach = new_detach.__get__(dtensor)
+    dtensor.clone = new_clone.__get__(dtensor)
+    return dtensor
+
+
+def distribute_tensor_with_customization(tensor: torch.Tensor, shard_fn, gather_fn: callable):
+    """
+    Distribute the given tensor with the given shard_fn and gather_fn.
+
+    Example:
+
+    ```python
+    # define shard and gather functions
+    def shard_fn(tensor):
+        rank = torch.distributed.get_rank()
+        world_size = torch.distributed.get_world_size()
+        return tensor.chunk(world_size, dim=0)[rank]
+
+    def gather_fn(tensor):
+        rank = torch.distributed.get_rank()
+        world_size = torch.distributed.get_world_size()
+        shard_list = [torch.zeros_like(tensor) for _ in range(world_size)]
+        torch.distributed.all_gather(shard_list, tensor)
+        return torch.cat(shard_list, dim=0)
+
+    # create a distributed tensor
+    tensor = torch.rand(4, 4)
+    dtensor = distribute_tensor_with_customization(tensor, shard_fn, gather_fn)
+    ```
+
+    Args:
+        tensor (torch.Tensor): The tensor to be distributed.
+        shard_fn (callable): The function to shard the tensor.
+        gather_fn (callable): The function to gather the tensor.
+
+    Returns:
+        torch.Tensor: The distributed tensor.
+    """
+    assert callable(shard_fn), 'The shard_fn must be callable.'
+    assert callable(gather_fn), 'The gather_fn must be callable.'
+    assert not is_distributed_tensor(tensor), 'The input tensor is already a distributed tensor.'
+
+    sharded_tensor = shard_fn(tensor)
+
+    # set the shard_fn and gather_fn as attributes of the distributed tensor
+    sharded_tensor.shard_fn = shard_fn
+    sharded_tensor.gather_fn = gather_fn
+
+    # set the shard_fn and gather_fn as attributes of the distributed tensor
+    _hijack_detach_and_clone_for_customized_distributed_tensor(sharded_tensor)
+
+    return sharded_tensor
+
+
+def to_global_for_customized_distributed_tensor(dtensor: torch.Tensor) -> torch.Tensor:
+    """
+    Gather the given tensor to the global tensor.
+
+    Args:
+        dtensor (torch.Tensor): The distributed tensor.
+
+    Returns:
+        torch.Tensor: The global tensor.
+    """
+    assert is_customized_distributed_tensor(dtensor), 'The input tensor is not a customized distributed tensor.'
+    return dtensor.gather_fn(dtensor)
+
+
+def customized_distributed_tensor_to_param(dtensor: torch.Tensor, requires_grad: bool = True):
+    """
+    Convert the given customized distributed tensor to a parameter.
+    """
+    assert is_customized_distributed_tensor(dtensor), 'The input tensor is not a customized distributed tensor.'
+
+    param = torch.nn.Parameter(dtensor, requires_grad=requires_grad)
+
+    # make it distributed as well
+    param.shard_fn = dtensor.shard_fn
+    param.gather_fn = dtensor.gather_fn
+    _hijack_detach_and_clone_for_customized_distributed_tensor(param)
+    return param

tests/test_shardformer/test_layer/test_linear_1d.py

@@ -27,8 +27,13 @@ def check_linear_1d_col():
 
     # check computation correctness
     x = torch.rand(4, 32).cuda()
-    out = linear(x)
-    gather_out = linear_col(x)
+    x_for_unshard = x.expand_as(x.clone())
+    x_for_unshard.requires_grad_(True)
+    x_for_shard = x.expand_as(x.clone())
+    x_for_shard.requires_grad_(True)
+
+    out = linear(x_for_unshard)
+    gather_out = linear_col(x_for_shard)
     assert_close(out, gather_out)
 
     # check backward correctness
@@ -39,6 +44,11 @@ def check_linear_1d_col():
     target_grad = torch.chunk(linear.weight.grad, 2, dim=0)[rank]
     assert_close(target_grad, linear_col.weight.grad)
 
+    # check the input gradients
+    assert x_for_shard.grad is not None
+    assert x_for_unshard.grad is not None
+    assert_close(x_for_unshard.grad, x_for_shard.grad)
+
 
 def check_linear_1d_row():
     linear = nn.Linear(32, 128).cuda()
@@ -49,8 +59,14 @@ def check_linear_1d_row():
 
     # check computation correctness
     x = torch.rand(4, 32).cuda()
-    out = linear(x)
-    gather_out = linear_row(x)
+    x_for_unshard = x.expand_as(x.clone())
+    x_for_unshard.requires_grad_(True)
+    x_for_shard = x.expand_as(x.clone())
+    x_for_shard.requires_grad_(True)
+
+    # run forward
+    out = linear(x_for_unshard)
+    gather_out = linear_row(x_for_shard)
     assert_close(out, gather_out)
 
     # check backward correctness
@@ -61,11 +77,49 @@ def check_linear_1d_row():
     target_grad = torch.chunk(linear.weight.grad, 2, dim=1)[rank]
     assert_close(target_grad, linear_row.weight.grad)
 
+    # check the input gradients
+    assert x_for_shard.grad is not None
+    assert x_for_unshard.grad is not None
+    assert_close(x_for_unshard.grad, x_for_shard.grad)
+
+
+def check_linear_col_plus_row():
+    linear_1 = nn.Linear(32, 128).cuda()
+    linear_2 = nn.Linear(128, 32).cuda()
+    linear_col = Linear1D_Col.from_native_module(linear_1, process_group=None, gather_output=False)
+    linear_row = Linear1D_Row.from_native_module(linear_2, process_group=None, parallel_input=True)
+
+    # check computation correctness
+    x = torch.rand(4, 32).cuda()
+    x_for_unshard = x.expand_as(x.clone())
+    x_for_unshard.requires_grad_(True)
+    x_for_shard = x.expand_as(x.clone())
+    x_for_shard.requires_grad_(True)
+
+    # run forward
+    unshard_out = linear_2(linear_1(x_for_unshard))
+    shard_out = linear_row(linear_col(x_for_shard))
+    assert_close(unshard_out, shard_out)
+
+    # check backward correctness
+    unshard_out.sum().backward()
+    shard_out.sum().backward()
+
+    rank = dist.get_rank()
+    target_1_grad = torch.chunk(linear_1.weight.grad, 2, dim=0)[rank]
+    assert_close(target_1_grad, linear_col.weight.grad)
+
+    # check the input gradients
+    assert x_for_shard.grad is not None
+    assert x_for_unshard.grad is not None
+    assert_close(x_for_unshard.grad, x_for_shard.grad)
+
 
 def run_dist(rank, world_size, port):
     colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
     check_linear_1d_col()
-    # check_linear_1d_row()
+    check_linear_1d_row()
+    check_linear_col_plus_row()
 
 
 @rerun_if_address_is_in_use()

tests/test_shardformer/test_layer/test_linearconv_1d.py

@@ -5,6 +5,7 @@ from torch.testing import assert_close
 
 import colossalai
 from colossalai.shardformer.layer import LinearConv1D_Col, LinearConv1D_Row
+from colossalai.shardformer.layer.linear_conv import split_fused_qkv
 from colossalai.testing import rerun_if_address_is_in_use, spawn
 
 
@@ -53,9 +54,15 @@ def check_linear_conv_1d_col():
     linear = Conv1D(192, 48).cuda()
     linear_conv_col = LinearConv1D_Col.from_native_module(linear, process_group=None, gather_output=True, n_fused=3)
 
-    assert linear_conv_col.weight.shape == torch.Size([96, 48])
+    assert linear.weight.shape == torch.Size([48, 192])
+    assert linear.bias.shape == torch.Size([192])
+    assert linear_conv_col.weight.shape == torch.Size([48, 96])
     assert linear_conv_col.bias.shape == torch.Size([96])
 
+    # ensure weights are reversibly loadable
+    linear_conv_col.load_state_dict(linear.state_dict())
+    linear.load_state_dict(linear_conv_col.state_dict())
+
     # check computation correctness
     x = torch.rand(4, 48).cuda()
     out = linear(x)
@@ -66,16 +73,16 @@ def check_linear_conv_1d_col():
     out.sum().backward()
     gather_out.sum().backward()
 
-    rank = dist.get_rank()
-    target_grad = torch.chunk(linear.weight.grad, 2, dim=1)[rank]
-    assert_close(target_grad.transpose(0, 1).contiguous(), linear_conv_col.weight.grad)
+    target_grad = split_fused_qkv(linear.weight.grad, 3, None)
+    assert_close(target_grad, linear_conv_col.weight.grad)
 
 
 def check_linear_1d_row():
     linear = Conv1D(192, 48).cuda()
     linear_row = LinearConv1D_Row.from_native_module(linear, process_group=None, parallel_input=False)
 
-    assert linear_row.weight.shape == torch.Size([192, 24])
+    assert linear.weight.shape == torch.Size([48, 192])
+    assert linear_row.weight.shape == torch.Size([24, 192])
     assert linear_row.bias.shape == torch.Size([192])
 
     # check computation correctness
@@ -89,13 +96,14 @@ def check_linear_1d_row():
     gather_out.sum().backward()
 
     rank = dist.get_rank()
-    target_grad = torch.chunk(linear.weight.grad, 2, dim=1)[rank]
+    target_grad = torch.chunk(linear.weight.grad, 2, dim=0)[rank]
     assert_close(target_grad, linear_row.weight.grad)
 
 
 def run_dist(rank, world_size, port):
     colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
     check_linear_conv_1d_col()
+    check_linear_1d_row()
 
 
 @rerun_if_address_is_in_use()

tests/test_shardformer/test_model/test_shard_gpt2.py

@@ -20,20 +20,21 @@ def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transfo
 
     # check grad equality
     if org_model.__class__.__name__ == 'GPT2Model':
-        org_grad = org_model.h[0].attn.c_attn.weight.grad
-        shard_grad = sharded_model.h[0].attn.c_attn.weight.grad.transpose(0, 1).contiguous()
+        org_grad = org_model.h[0].mlp.c_fc.weight.grad
+        shard_grad = sharded_model.h[0].mlp.c_fc.weight.grad
     else:
         org_grad = org_model.transformer.h[0].mlp.c_fc.weight.grad
-        shard_grad = sharded_model.transformer.h[0].mlp.c_fc.weight.grad.transpose(0, 1).contiguous()
+        shard_grad = sharded_model.transformer.h[0].mlp.c_fc.weight.grad
 
     shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
     shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
     all_shard_grad = torch.cat(shard_grad_list, dim=1)
 
     assert torch.allclose(org_loss, shard_loss,
-                          atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
-    assert torch.allclose(org_grad, all_shard_grad,
-                          atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{all_shard_grad}"
+                          atol=1e-5), f"shard model loss is not equal to origin model loss\n{org_loss}\n{shard_loss}"
+    assert torch.allclose(
+        org_grad, all_shard_grad,
+        atol=1e-5), f"shard model grad is not equal to origin model grad\n{org_grad}\n{all_shard_grad}"
 
 
 def check_gpt2(rank, world_size, port):