[tensor] support runtime ShardingSpec apply (#1453)

* [tensor] support runtime ShardingSpec apply

* polish code

* polish code

tests/test_device/test_init_logical_pg.py

@@ -0,0 +1,49 @@
+import torch
+from functools import partial
+import pytest
+import torch.distributed as dist
+import torch.multiprocessing as mp
+from torch.distributed import ReduceOp
+
+from colossalai.core import global_context as gpc
+from colossalai.initialize import launch
+from colossalai.utils import free_port
+from colossalai.testing import rerun_if_address_is_in_use
+from colossalai.device.device_mesh import DeviceMesh
+
+
+def check_layer(rank, world_size, port):
+    launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+
+    physical_mesh_id = torch.arange(0, 4)
+    assert rank == gpc.get_global_rank()
+
+    tensor_to_check = torch.tensor([2, 2, 2, 2]).cuda()
+    mesh_shape = (2, 2)
+    # [[0, 1,
+    #  [2, 3]]
+    device_mesh = DeviceMesh(physical_mesh_id, mesh_shape, init_process_group=True)
+    logical_pg_dict = {0: [[0, 2], [1, 3]], 1: [[0, 1], [2, 3]]}
+    logical_process_groups = device_mesh.process_groups_dict
+
+    for mesh_dim, pgs in logical_pg_dict.items():
+        for index, pg in enumerate(pgs):
+            if rank in pg:
+                tensor = torch.ones(4).cuda()
+                group = logical_process_groups[mesh_dim][index][1]
+                dist.all_reduce(tensor, op=ReduceOp.SUM, group=group)
+                assert tensor.equal(tensor_to_check)
+
+    gpc.destroy()
+
+
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+def test_logical_pg():
+    world_size = 4
+    run_func = partial(check_layer, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_logical_pg()

tests/test_tensor/test_comm_spec_apply.py

@@ -0,0 +1,177 @@
+import torch
+from functools import partial
+import pytest
+import torch.distributed as dist
+import torch.multiprocessing as mp
+from torch.distributed import ReduceOp
+from colossalai.core import global_context as gpc
+from colossalai.initialize import launch
+from colossalai.utils import free_port
+from colossalai.testing import rerun_if_address_is_in_use
+from colossalai.device.device_mesh import DeviceMesh
+from colossalai.tensor.shape_consistency import CommSpec, CollectiveCommPattern
+from colossalai.logging import disable_existing_loggers
+from colossalai.tensor.sharding_spec import ShardingSpec
+
+
+def check_all_gather(device_mesh, rank):
+    # tensor to comm
+    if rank in (0, 2):
+        sharded_tensor_to_comm = torch.ones(2, 2).cuda()
+    else:
+        sharded_tensor_to_comm = torch.zeros(2, 2).cuda()
+
+    # tensor to check
+    tensor_to_check = torch.cat((torch.ones(2, 2), torch.zeros(2, 2)), 1).cuda()
+
+    # test all gather
+    dim_partition_dict = {1: [1]}
+
+    # DistSpec:
+    #     shard_sequence: R,S1
+    #     device_mesh_shape: (2, 2)
+    sharding_spec = ShardingSpec(device_mesh, tensor_to_check.shape, dim_partition_dict=dim_partition_dict)
+
+    # CommSpec:(comm_pattern:allgather, gather_dim:1, logical_process_axis:1)
+    comm_spec = CommSpec(CollectiveCommPattern.ALLGATHER, sharding_spec, gather_dim=1, logical_process_axis=1)
+    comm_spec.covert_spec_to_action(sharded_tensor_to_comm)
+
+    assert sharded_tensor_to_comm.equal(tensor_to_check)
+
+
+def check_shard(device_mesh, rank):
+    # tensor to comm
+    sharded_tensor_to_comm_0 = torch.zeros(2, 2).cuda()
+    sharded_tensor_to_comm_1 = torch.ones(2, 2).cuda()
+    # tensor([[0., 0., 1., 1.],
+    #         [0., 0., 1., 1.]])
+    tensor_to_shard = torch.cat((sharded_tensor_to_comm_0, sharded_tensor_to_comm_1), 1)
+
+    # test shard
+    dim_partition_dict = {}
+
+    # DistSpec:
+    #     shard_sequence: R,R
+    #     device_mesh_shape: (2, 2)
+    sharding_spec = ShardingSpec(device_mesh, tensor_to_shard.shape, dim_partition_dict=dim_partition_dict)
+
+    # CommSpec:(comm_pattern:shard, shard_dim:1, logical_process_axis:1)
+    comm_spec = CommSpec(CollectiveCommPattern.SHARD, sharding_spec, shard_dim=1, logical_process_axis=1)
+    comm_spec.covert_spec_to_action(tensor_to_shard)
+
+    if rank in (0, 2):
+        assert tensor_to_shard.equal(sharded_tensor_to_comm_0)
+    if rank in (1, 3):
+        assert tensor_to_shard.equal(sharded_tensor_to_comm_1)
+
+
+def check_all_to_all(device_mesh, rank):
+    # tensor to comm
+    if rank in (0, 1):
+        sharded_tensor_0 = torch.zeros(2, 1)
+        sharded_tensor_1 = torch.ones(2, 1)
+        # tensor([[0., 1.],
+        #         [0., 1.]])
+        tensor_to_comm = torch.cat((sharded_tensor_0, sharded_tensor_1), 1).cuda()
+    if rank in (2, 3):
+        sharded_tensor_0 = torch.ones(2, 1) * 2
+        sharded_tensor_1 = torch.ones(2, 1) * 3
+        # tensor([[2., 3.],
+        #         [2., 3.]])
+        tensor_to_comm = torch.cat((sharded_tensor_0, sharded_tensor_1), 1).cuda()
+
+    if rank in (0, 1):
+        # tensor([[0.],
+        #         [0.],
+        #         [2.],
+        #         [2.]])
+        tensor_to_check = torch.tensor([[0], [0], [2], [2]], dtype=tensor_to_comm.dtype).cuda()
+    if rank in (2, 3):
+        # tensor([[1.],
+        #         [1.],
+        #         [3.],
+        #         [3.]])
+        tensor_to_check = torch.tensor([[1], [1], [3], [3]], dtype=tensor_to_comm.dtype).cuda()
+
+    # test shard
+    dim_partition_dict = {0: [0]}
+
+    # DistSpec:
+    #     shard_sequence: S0,R
+    #     device_mesh_shape: (2, 2)
+    sharding_spec = ShardingSpec(device_mesh, torch.Size((4, 2)), dim_partition_dict=dim_partition_dict)
+
+    # CommSpec:(comm_pattern:shard, shard_dim:1, logical_process_axis:1)
+    comm_spec = CommSpec(CollectiveCommPattern.ALLTOALL,
+                         sharding_spec,
+                         gather_dim=0,
+                         shard_dim=1,
+                         logical_process_axis=0)
+    comm_spec.covert_spec_to_action(tensor_to_comm)
+
+    assert tensor_to_comm.equal(tensor_to_check)
+
+
+def check_all_reduce(device_mesh, rank):
+    # tensor to comm
+    tensor_to_comm = torch.ones(2, 2).cuda() * rank
+
+    # reduce through logical process axis 0
+    # tensor to check
+    if rank in (0, 2):
+        # tensor([[2., 2.],
+        #         [2., 2.]])
+        tensor_to_check = torch.tensor([[2, 2], [2, 2]], dtype=tensor_to_comm.dtype).cuda()
+    if rank in (1, 3):
+        # tensor([[4., 4.],
+        #         [4., 4.]])
+        tensor_to_check = torch.tensor([[4, 4], [4, 4]], dtype=tensor_to_comm.dtype).cuda()
+
+    dim_partition_dict = {}
+    # DistSpec:
+    #     shard_sequence: R,R
+    #     device_mesh_shape: (2, 2)
+    sharding_spec = ShardingSpec(device_mesh, tensor_to_comm.shape, dim_partition_dict=dim_partition_dict)
+
+    # CommSpec:CommSpec:(comm_pattern:all_reduce, logical_process_axis:0)
+    comm_spec = CommSpec(CollectiveCommPattern.ALLREDUCE, sharding_spec, logical_process_axis=0)
+    comm_spec.covert_spec_to_action(tensor_to_comm)
+
+    assert tensor_to_comm.equal(tensor_to_check)
+
+
+def check_comm(rank, world_size, port):
+    disable_existing_loggers()
+    launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+
+    physical_mesh_id = torch.arange(0, 4)
+    assert rank == gpc.get_global_rank()
+
+    mesh_shape = (2, 2)
+    # [[0, 1,
+    #  [2, 3]]
+    device_mesh = DeviceMesh(physical_mesh_id, mesh_shape, init_process_group=True)
+    # test all gather
+    check_all_gather(device_mesh, rank)
+
+    # test shard
+    check_shard(device_mesh, rank)
+
+    # test all to all
+    check_all_to_all(device_mesh, rank)
+
+    # test all reduce
+    check_all_reduce(device_mesh, rank)
+    gpc.destroy()
+
+
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+def test_comm_spec():
+    world_size = 4
+    run_func = partial(check_comm, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_comm_spec()

tests/test_tensor/test_shape_consistency_apply.py

@@ -0,0 +1,81 @@
+from functools import partial
+import pytest
+import torch
+import torch.multiprocessing as mp
+
+from colossalai.tensor.sharding_spec import ShardingSpec
+from colossalai.device.device_mesh import DeviceMesh
+from colossalai.initialize import launch
+from colossalai.utils import free_port
+from colossalai.testing import rerun_if_address_is_in_use
+from colossalai.logging import disable_existing_loggers
+from colossalai.tensor.shape_consistency import ShapeConsistencyManager, CollectiveCommPattern
+
+
+def check_apply(rank, world_size, port):
+    disable_existing_loggers()
+    launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+
+    physical_mesh_id = torch.arange(0, 4)
+    mesh_shape = (2, 2)
+    # [[0, 1,
+    #  [2, 3]]
+    device_mesh = DeviceMesh(physical_mesh_id, mesh_shape, init_process_group=True)
+    entire_shape = torch.Size((4, 2))
+    shape_consistency_manager = ShapeConsistencyManager()
+    dim_partition_source = {0: [0]}
+    dim_partition_target = {1: [0]}
+
+    # DistSpec:
+    #     shard_sequence: S0,R
+    #     device_mesh_shape: (2, 2)
+    sharding_spec_source = ShardingSpec(device_mesh, entire_shape, dim_partition_source)
+
+    # DistSpec:
+    #     shard_sequence: R,S0
+    #     device_mesh_shape: (2, 2)
+    sharding_spec_target = ShardingSpec(device_mesh, entire_shape, dim_partition_target)
+
+    if rank in (0, 1):
+        sharded_tensor_0 = torch.zeros(2, 1)
+        sharded_tensor_1 = torch.ones(2, 1)
+        # tensor([[0., 1.],
+        #         [0., 1.]])
+        tensor_to_comm = torch.cat((sharded_tensor_0, sharded_tensor_1), 1).cuda()
+    if rank in (2, 3):
+        sharded_tensor_0 = torch.ones(2, 1) * 2
+        sharded_tensor_1 = torch.ones(2, 1) * 3
+        # tensor([[2., 3.],
+        #         [2., 3.]])
+        tensor_to_comm = torch.cat((sharded_tensor_0, sharded_tensor_1), 1).cuda()
+
+    if rank in (0, 1):
+        # tensor([[0.],
+        #         [0.],
+        #         [2.],
+        #         [2.]])
+        tensor_to_check = torch.tensor([[0], [0], [2], [2]], dtype=tensor_to_comm.dtype).cuda()
+    if rank in (2, 3):
+        # tensor([[1.],
+        #         [1.],
+        #         [3.],
+        #         [3.]])
+        tensor_to_check = torch.tensor([[1], [1], [3], [3]], dtype=tensor_to_comm.dtype).cuda()
+
+    tensor_to_comm.sharding_spec = sharding_spec_source
+    shape_consistency_manager.apply(tensor_to_comm, sharding_spec_target)
+    print(tensor_to_comm)
+    assert tensor_to_comm.equal(tensor_to_check)
+    assert str(tensor_to_comm.sharding_spec.sharding_sequence) == str(sharding_spec_target.sharding_sequence)
+
+
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+def test_apply():
+    world_size = 4
+    run_func = partial(check_apply, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_apply()