Add mp_all_reduce asynchronize overlap. (#55662)

* [WIP] Add mp_all_reduce asynchronize overlap.

* Fix some problems.

* Fix dw compute bug, and use a temporary solution to achieve overlap.

* Use fused_linear_param_grad_add to compute dw.

* Reformat ColumnParallel _overlap_linear. Use environment flags to
control following behaviors:
1. export Flags_mp_aysnc_allreduce=True to turn on mp async all_reduce
2. export Flags_skip_mp_c_identity=True to skip two c_identity operators
   in dygraph mode.
3. export Flags_fused_linear_param_grad_add to enable fused_linear_param_grad_add
   in ColumnParallel backward with mp async all_reduce.

* Polish code.

* Remove useless communication API.

* Fix some problems in mp_async_all_reduce and skip_c_identity.

* Add test cases.

* Remove environment variable Flags_fused_linear_param_grad_add in test case.

* Reset error threshold.

* Reset threshold in test case.

* Add useful log. Remove useless test cases.

python/paddle/distributed/fleet/layers/mpu/mp_layers.py

@@ -17,8 +17,10 @@ from paddle.autograd import PyLayer
 from paddle.fluid import core
 from paddle.nn import functional as F
 
+from ....communication.reduce import ReduceOp, _get_reduce_op
 from ...base import topology as tp
 from . import mp_ops
+from .mp_ops import _get_mp_env_flag
 from .random import get_rng_state_tracker
 
 __all__ = []
@@ -32,6 +34,13 @@ def is_fused_matmul_bias_supported():
     return hasattr(core.eager.ops.legacy, 'fused_gemm_epilogue')
 
 
+def is_fused_linear_param_grad_add_supported():
+    if paddle.is_compiled_with_cuda() and not paddle.is_compiled_with_rocm():
+        return hasattr(paddle._C_ops, 'fused_linear_param_grad_add')
+    else:
+        return False
+
+
 class VocabParallelEmbedding(paddle.nn.Layer):
     """Embedding mp parallelized in the vocabulary dimension.
     this class is used for splitting embedding in mp group.
@@ -295,7 +304,8 @@ class ColumnParallelLinear(paddle.nn.Layer):
 
         self.linear = F.linear
 
-        if fuse_matmul_bias:
+        self.fuse_matmul_bias = fuse_matmul_bias
+        if self.fuse_matmul_bias:
             if not is_fused_matmul_bias_supported():
                 raise NotImplementedError(
                     "You set fuse_matmul_bias=True in ColumnParallelLinear, "
@@ -309,6 +319,139 @@ class ColumnParallelLinear(paddle.nn.Layer):
 
     def forward(self, x):
         # use inner api to process identity
+
+        def _overlap_linear():
+            fuse_matmul_bias = self.fuse_matmul_bias
+
+            class InnerOverlapLinear(paddle.autograd.PyLayer):
+                @staticmethod
+                def forward(ctx, x, weight, bias):
+                    ctx.save_for_backward(x, weight, bias)
+                    if (
+                        _get_mp_env_flag("Flags_mp_aysnc_allreduce")
+                        and _get_mp_env_flag("Flags_skip_mp_c_identity")
+                    ) is False:
+                        x = paddle._legacy_C_ops.c_identity(
+                            x,
+                            'use_calc_stream',
+                            True,
+                            'ring_id',
+                            self.model_parallel_group.id,
+                            'use_model_parallel',
+                            True,
+                        )
+                    if not fuse_matmul_bias:
+                        return paddle._C_ops.linear(x, weight, bias)
+                    else:
+                        return paddle._legacy_C_ops.fused_gemm_epilogue(
+                            x, weight, bias
+                        )
+
+                @staticmethod
+                def backward(ctx, dy):
+                    x, weight, bias = ctx.saved_tensor()
+                    dx = paddle.matmul(dy, weight, transpose_y=True)
+                    op_type = _get_reduce_op(ReduceOp.SUM, "_c_identity")
+                    task = self.model_parallel_group.process_group.all_reduce(
+                        dx, op_type, sync_op=False
+                    )
+                    # TODO(GhostScreaming): remove it in future.
+                    tmp = paddle.ones([512])
+
+                    if _get_mp_env_flag("Flags_fused_linear_param_grad_add"):
+                        if not is_fused_linear_param_grad_add_supported():
+                            raise NotImplementedError(
+                                "You set environment variable Flags_fused_linear_param_grad_add=True, "
+                                "however, the paddle you are using not support this operation. "
+                                "Please unset Flags_fused_linear_param_grad_add or use paddle compiled "
+                                "with cuda 11.6 or higher."
+                            )
+
+                        if bias is None:
+                            if hasattr(weight, "main_grad"):
+                                (
+                                    weight.main_grad,
+                                    _,
+                                ) = paddle._C_ops.fused_linear_param_grad_add(
+                                    x, dy, weight.main_grad, None, True, False
+                                )
+                                task.wait()
+                                return dx, None
+                            else:
+                                if weight.grad is not None:
+                                    (
+                                        weight.grad,
+                                        _,
+                                    ) = paddle._C_ops.fused_linear_param_grad_add(
+                                        x, dy, weight.grad, None, False, False
+                                    )
+                                    task.wait()
+                                    return dx, None
+                                else:
+                                    (
+                                        dw,
+                                        _,
+                                    ) = paddle._C_ops.fused_linear_param_grad_add(
+                                        x, dy, None, None, False, False
+                                    )
+                                    task.wait()
+                                    return dx, dw
+
+                        if hasattr(weight, "main_grad") and hasattr(
+                            bias, "main_grad"
+                        ):
+                            (
+                                weight.main_grad,
+                                bias.main_grad,
+                            ) = paddle._C_ops.fused_linear_param_grad_add(
+                                input,
+                                dy,
+                                weight.main_grad,
+                                bias.main_grad,
+                                True,
+                                True,
+                            )
+                            task.wait()
+                            return dx, None, None
+                        else:
+                            if weight.grad is not None:
+                                assert bias.grad is not None
+                                (
+                                    weight.grad,
+                                    bias.grad,
+                                ) = paddle._C_ops.fused_linear_param_grad_add(
+                                    x, dy, weight.grad, bias.grad, False, True
+                                )
+                                task.wait()
+                                return dx, None, None
+                            else:
+                                (
+                                    dw,
+                                    dbias,
+                                ) = paddle._C_ops.fused_linear_param_grad_add(
+                                    x, dy, None, None, False, True
+                                )
+                                task.wait()
+                                return dx, dw, dbias
+                    else:
+                        dw = paddle.matmul(
+                            x.reshape([-1, x.shape[-1]]),
+                            dy.reshape([-1, dy.shape[-1]]),
+                            transpose_x=True,
+                        )
+                        if bias is None:
+                            task.wait()
+                            return dx, dw
+                        else:
+                            dbias = paddle.sum(dy, axis=0)
+                            task.wait()
+                            return dx, dw, dbias
+
+            return InnerOverlapLinear.apply(x, self.weight, self.bias)
+
+        if _get_mp_env_flag("Flags_mp_aysnc_allreduce"):
+            output_parallel = _overlap_linear()
+        else:
             if self.is_mp:
                 input_parallel = mp_ops._c_identity(
                     x, group=self.model_parallel_group

python/paddle/distributed/fleet/layers/mpu/mp_ops.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
+
 import paddle
 from paddle import _legacy_C_ops
 from paddle.distributed import collective
@@ -22,6 +24,38 @@ from paddle.nn.utils import dygraph_utils
 
 from ....communication.reduce import ReduceOp, _get_reduce_op
 
+_first_get_mp_env_flag = True
+
+
+def _get_mp_env_flag(flag):
+    global _first_get_mp_env_flag
+    if _first_get_mp_env_flag:
+        print(
+            "Flags_mp_aysnc_allreduce is {}, which is used to support all_reduce(dx) overlap with matmul(dw) in ColumnParallelLinear.".format(
+                str(os.getenv("Flags_mp_aysnc_allreduce")).lower()
+            )
+        )
+        print(
+            "Flags_fused_linear_param_grad_add is {}, which is used to support fused_linear_param_grad_add in ColumnParallelLinear. Only works when Flags_mp_aysnc_allreduce is True.".format(
+                str(os.getenv("Flags_fused_linear_param_grad_add")).lower()
+            )
+        )
+        print(
+            "Flags_skip_mp_c_identity is {}, which is used to support skip c_identity in ColumnParallelLinear and RowParallelLinear. Only works when Flags_mp_aysnc_allreduce is True.".format(
+                str(os.getenv("Flags_skip_mp_c_identity")).lower()
+            )
+        )
+    # Model parallel environment flag.
+    # Flags_mp_aysnc_allreduce: support all_reduce(dx) overlap with matmul(dw) in ColumnParallelLinear
+    # Flags_fused_linear_param_grad_add: support fused_linear_param_grad_add in ColumnParallelLinear. Only works when Flags_mp_aysnc_allreduce is True.
+    # Flags_skip_mp_c_identity: support skip c_identity in ColumnParallelLinear and RowParallelLinear. Only works when Flags_mp_aysnc_allreduce is True.
+    assert flag in [
+        "Flags_mp_aysnc_allreduce",
+        "Flags_fused_linear_param_grad_add",
+        "Flags_skip_mp_c_identity",
+    ], "Only support set Flags_mp_aysnc_allreduce (support all_reduce(dx) overlap with matmul(dw) in ColumnParallelLinear), Flags_fused_linear_param_grad_add (support fused_linear_param_grad_add in ColumnParallelLinear) and Flags_skip_mp_c_identity (support skip c_identity in ColumnParallelLinear with Flags_mp_aysnc_allreduce=True, and skip c_identity in RowParallelLinear)"
+    return str(os.getenv(flag)).lower() in ["true", "1"]
+
 
 def _c_identity(tensor, group=None):
     """
@@ -45,6 +79,11 @@ def _c_identity(tensor, group=None):
         class c_identity_eager(PyLayer):
             @staticmethod
             def forward(ctx, tensor):
+                if _get_mp_env_flag(
+                    "Flags_mp_aysnc_allreduce"
+                ) and _get_mp_env_flag("Flags_skip_mp_c_identity"):
+                    return tensor
+                else:
                     return _legacy_C_ops.c_identity(
                         tensor,
                         'use_calc_stream',
@@ -256,6 +295,11 @@ def _mp_allreduce(
 
             @staticmethod
             def backward(ctx, dy):
+                if _get_mp_env_flag(
+                    "Flags_mp_aysnc_allreduce"
+                ) and _get_mp_env_flag("Flags_skip_mp_c_identity"):
+                    return dy
+                else:
                     return _legacy_C_ops.c_identity(
                         dy,
                         'use_calc_stream',