[HybridParallel]Fix precision problem of model parallel (#32897) (#33087)

* fix precision of mp

* fix bug of seed

* fix dp

* print group

paddle/fluid/framework/distributed_strategy.proto

@@ -141,6 +141,7 @@ message PipelineConfig {
 
 message TensorParallelConfig {
   optional int32 tensor_parallel_degree = 1 [ default = 1 ];
+  optional int32 tensor_init_seed = 2 [ default = -1 ];
 }
 
 message DistributedStrategy {

python/paddle/distributed/collective.py

@@ -98,6 +98,13 @@ class Group():
         else:
             return -1
 
+    def __repr__(self):
+        debug_str = "rank: {}, nranks: {}, id: {}, ranks: ".format(
+            self.rank, self.nranks, self.id)
+        debug_str += ", ".join(map(str, self.ranks))
+        debug_str += ". "
+        return debug_str
+
 
 _global_env = None
 

python/paddle/distributed/fleet/base/distributed_strategy.py

@@ -923,6 +923,8 @@ class DistributedStrategy(object):
         **Notes**:
             **Detailed arguments for tensor_parallel_configs**
             **tensor_parallel_degree**: degree of tensor parallel
+            **tensor_init_seed**: parameter initialization random seed
+
 
         Examples:
 
@@ -931,7 +933,8 @@ class DistributedStrategy(object):
             import paddle.distributed.fleet as fleet
             strategy = fleet.DistributedStrategy()
             strategy.tensor_parallel = True
-            strategy.tensor_parallel_configs = {"tensor_parallel_degree": 4}
+            strategy.tensor_parallel_configs = {"tensor_parallel_degree": 4,
+                                                "tensor_init_seed": 123}
 
         """
         return get_msg_dict(self.strategy.tensor_parallel_configs)

python/paddle/distributed/fleet/base/fleet_base.py

@@ -17,6 +17,7 @@ import copy
 import warnings
 import paddle
 import os
+import numpy as np
 from paddle.fluid.framework import dygraph_only
 from paddle.fluid import compiler
 from .role_maker import UserDefinedRoleMaker, PaddleCloudRoleMaker, RoleMakerBase
@@ -28,7 +29,7 @@ from paddle.fluid.wrapped_decorator import wrap_decorator
 from paddle.fluid.dygraph import parallel_helper
 from . import topology as tp
 from .topology import ParallelMode
-from ..meta_parallel import ModelParallel
+from ..meta_parallel import TensorParallel, model_parallel_random_seed
 from ..meta_parallel import PipelineParallel
 from ..meta_optimizers import HybridParallelOptimizer
 from ..meta_optimizers import HybridParallelGradScaler
@@ -279,6 +280,14 @@ class Fleet(object):
 
         self._hcg = tp.HybridCommunicateGroup(self._topology)
 
+        if self.mp_degree > 1:
+            tensor_parallel_configs = self._user_defined_strategy.tensor_parallel_configs
+            tensor_init_seed = tensor_parallel_configs["tensor_init_seed"]
+            if tensor_init_seed == -1:
+                model_parallel_random_seed()
+            else:
+                model_parallel_random_seed(tensor_init_seed)
+
     def get_hybrid_communicate_group(self):
         assert self._hcg is not None
         return self._hcg
@@ -780,8 +789,8 @@ class Fleet(object):
                 last_comm_group_size_MB,
                 find_unused_parameters=self._user_defined_strategy.
                 find_unused_parameters)
-        elif self._hcg.get_parallel_mode() == ParallelMode.MODEL_PARALLEL:
-            distributed_model = ModelParallel(
+        elif self._hcg.get_parallel_mode() == ParallelMode.TENSOR_PARALLEL:
+            distributed_model = TensorParallel(
                 model, self._hcg, strategy=self._user_defined_strategy)
         elif self._hcg.get_parallel_mode() == ParallelMode.PIPELINE_PARALLEL:
             distributed_model = PipelineParallel(

python/paddle/distributed/fleet/base/topology.py

@@ -28,7 +28,7 @@ _HYBRID_PARALLEL_GROUP = None
 
 class ParallelMode(object):
     DATA_PARALLEL = 0
-    MODEL_PARALLEL = 1
+    TENSOR_PARALLEL = 1
     PIPELINE_PARALLEL = 2
 
 
@@ -155,12 +155,12 @@ class HybridCommunicateGroup(object):
         _HYBRID_PARALLEL_GROUP = self
 
     def get_parallel_mode(self):
-        # there are three modes : DataParallel / ModelParallel / PipelineParallel
+        # there are three modes : DataParallel / TensorParallel / PipelineParallel
         if self._mp_degree == 1 and self._pp_degree == 1:
             return ParallelMode.DATA_PARALLEL
         elif self._mp_degree > 1 and self._pp_degree == 1:
             # initialize the seed
-            return ParallelMode.MODEL_PARALLEL
+            return ParallelMode.TENSOR_PARALLEL
         elif self._pp_degree > 1:
             return ParallelMode.PIPELINE_PARALLEL
 

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_gradscaler.py

@@ -31,7 +31,7 @@ class HybridParallelGradScaler:
         self._scaler = scaler
         self._hcg = hcg
         self._is_mp = (
-            self._hcg.get_parallel_mode() == ParallelMode.MODEL_PARALLEL)
+            self._hcg.get_parallel_mode() == ParallelMode.TENSOR_PARALLEL)
 
     def scale(self, var):
         return self._scaler.scale(var)

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_optimizer.py

@@ -90,12 +90,12 @@ class HybridParallelOptimizer:
         self._strategy = strategy
         self._hcg = hcg
         self._is_mp = (
-            self._hcg.get_parallel_mode() == ParallelMode.MODEL_PARALLEL)
+            self._hcg.get_parallel_mode() == ParallelMode.TENSOR_PARALLEL)
         self._need_dp = (self._hcg.get_data_parallel_world_size() > 1)
 
         if isinstance(self._inner_opt._grad_clip,
                       ClipGradByGlobalNorm) and self._is_mp:
-            logger.warning("using ClipGradByGlobalNorm in ModelParallel, the origin " \
+            logger.warning("using ClipGradByGlobalNorm in TensorParallel, the origin " \
                   "optmizer'grad clip will be changed.")
             self._inner_opt._grad_clip = HybridParallelClipGrad(
                 self._inner_opt._grad_clip, hcg)

python/paddle/distributed/fleet/meta_parallel/__init__.py

@@ -20,7 +20,7 @@ from .parallel_layers import PipelineLayer  # noqa: F401
 from .parallel_layers import RNGStatesTracker  # noqa: F401
 from .parallel_layers import model_parallel_random_seed  # noqa: F401
 from .parallel_layers import get_rng_state_tracker  # noqa: F401
-from .model_parallel import ModelParallel  # noqa: F401
+from .tensor_parallel import TensorParallel  # noqa: F401
 from .pipeline_parallel import PipelineParallel  # noqa: F401
 
 __all__ = []

python/paddle/distributed/fleet/meta_parallel/parallel_layers/mp_layers.py

@@ -41,6 +41,7 @@ class VocabParallelEmbedding(Layer):
         self.rank = tp._HYBRID_PARALLEL_GROUP.get_model_parallel_rank()
 
         self.origin_num_embeddings = num_embeddings
+        self.is_mp = (self.world_size > 1)
 
         per_part_size = (
             num_embeddings + self.world_size - 1) // self.world_size
@@ -50,16 +51,36 @@ class VocabParallelEmbedding(Layer):
         per_part_size += 1  # make the last row as the padding index
         self.per_part_size = per_part_size
 
-        self.embedding = paddle.nn.Embedding(
-            per_part_size,
-            embedding_dim,
-            padding_idx=per_part_size - 1,
-            sparse=False,
-            weight_attr=weight_attr,
-            name=name)
-        self.embedding.weight.is_distributed = True
+        self._dtype = self._helper.get_default_dtype()
+        self._size = [per_part_size, embedding_dim]
+        self._weight_attr = weight_attr
+        self._name = name
+
+        if self.is_mp:
+            with get_rng_state_tracker().rng_state():
+                self.weight = self.create_parameter(
+                    attr=self._weight_attr,
+                    shape=self._size,
+                    dtype=self._dtype,
+                    is_bias=False)
+            self.weight[per_part_size - 1] = 0.0
+            self.weight.is_distributed = True
+        else:
+            self.weight = self.create_parameter(
+                attr=self._weight_attr,
+                shape=[num_embeddings, embedding_dim],
+                dtype=self._dtype,
+                is_bias=False)
 
     def forward(self, x):
+        if not self.is_mp:
+            return F.embedding(
+                x,
+                weight=self.weight,
+                padding_idx=None,
+                sparse=False,
+                name=self._name)
+
         origin_input_shape = x.shape
         if len(origin_input_shape) == 2:
             x = paddle.unsqueeze(x, axis=-1)
@@ -72,8 +93,13 @@ class VocabParallelEmbedding(Layer):
         if len(origin_input_shape) == 2:
             x_shard = paddle.squeeze(x_shard, axis=-1)
 
-        emb_out = self.embedding(x_shard)
-        if self.world_size > 1:
+        emb_out = F.embedding(
+            x_shard,
+            weight=self.weight,
+            padding_idx=self.per_part_size - 1,
+            sparse=False,
+            name=self._name)
+
         emb_out = paddle.distributed.collective._mp_allreduce(
             emb_out,
             group=self.model_parallel_group,
@@ -96,8 +122,9 @@ class ColumnParallelLinear(Layer):
         )
         self.world_size = tp._HYBRID_PARALLEL_GROUP.get_model_parallel_world_size(
         )
+        self._name = name
+        self.is_mp = (self.world_size > 1)
 
-        self.name = name
         self.gather_output = gather_output
         assert out_features % self.world_size == 0, (
             "Number of column of the weight for linear ({}) must be"
@@ -108,10 +135,20 @@ class ColumnParallelLinear(Layer):
         self._weight_attr = weight_attr
         self._dtype = self._helper.get_default_dtype()
 
+        if self.is_mp:
+            with get_rng_state_tracker().rng_state():
                 self.weight = self.create_parameter(
                     shape=[in_features, self.output_size_per_partition],
                     attr=self._weight_attr,
-            dtype=self._dtype)
+                    dtype=self._dtype,
+                    is_bias=False)
+        else:
+            self.weight = self.create_parameter(
+                shape=[in_features, self.output_size_per_partition],
+                attr=self._weight_attr,
+                dtype=self._dtype,
+                is_bias=False)
+
         self.weight.is_distributed = True
 
         if has_bias:
@@ -119,18 +156,24 @@ class ColumnParallelLinear(Layer):
             self.bias = self.create_parameter(
                 shape=[self.output_size_per_partition],
                 attr=paddle.nn.initializer.Constant(value=0.0),
-                dtype=self._dtype)
+                dtype=self._dtype,
+                is_bias=True)
             self.bias.is_distributed = True
         else:
             self.bias = None
 
     def forward(self, x):
         # use inner api to process identity
+        if self.is_mp:
             input_parallel = paddle.distributed.collective._c_identity(
                 x, group=self.model_parallel_group)
+        else:
+            input_parallel = x
+
         output_parallel = F.linear(
-            input_parallel, self.weight, self.bias, name=self.name)
-        if self.gather_output:
+            input_parallel, self.weight, self.bias, name=self._name)
+
+        if self.gather_output and self.is_mp:
             output = paddle.distributed.collective._c_concat(
                 output_parallel,
                 nranks=self.world_size,
@@ -155,7 +198,7 @@ class RowParallelLinear(Layer):
         self.input_is_parallel = input_is_parallel
         self._weight_attr = weight_attr
         self._dtype = self._helper.get_default_dtype()
-        self.name = name
+        self._name = name
 
         self.model_parallel_group = tp._HYBRID_PARALLEL_GROUP.get_model_parallel_group(
         )
@@ -163,6 +206,7 @@ class RowParallelLinear(Layer):
         )
         self.rank = tp._HYBRID_PARALLEL_GROUP.get_model_parallel_rank()
 
+        self.is_mp = (self.world_size > 1)
         assert in_features % self.world_size == 0, (
             "Number of row of the weight for linear ({}) must be"
             " divisible by model parallel size ({})".format(in_features,
@@ -170,22 +214,33 @@ class RowParallelLinear(Layer):
 
         self.input_size_per_partition = in_features // self.world_size
 
+        if self.is_mp:
+            with get_rng_state_tracker().rng_state():
+                self.weight = self.create_parameter(
+                    shape=[self.input_size_per_partition, self.out_features],
+                    attr=self._weight_attr,
+                    dtype=self._dtype,
+                    is_bias=False)
+        else:
             self.weight = self.create_parameter(
                 shape=[self.input_size_per_partition, self.out_features],
                 attr=self._weight_attr,
-            dtype=self._dtype)
+                dtype=self._dtype,
+                is_bias=False)
+
         self.weight.is_distributed = True
 
         if has_bias:
             self.bias = self.create_parameter(
                 shape=[self.out_features],
                 attr=paddle.nn.initializer.Constant(value=0.0),
-                dtype=self._dtype)
+                dtype=self._dtype,
+                is_bias=True)
         else:
             self.bias = None
 
     def forward(self, x):
-        if self.input_is_parallel:
+        if self.input_is_parallel or (not self.is_mp):
             input_parallel = x
         else:
             # split last dim
@@ -195,12 +250,16 @@ class RowParallelLinear(Layer):
                 nranks=self.world_size,
                 group=self.model_parallel_group)
 
-        output_parallel = F.linear(input_parallel, self.weight, name=self.name)
+        output_parallel = F.linear(input_parallel, self.weight, name=self._name)
+
+        if self.is_mp:
             output_ = paddle.distributed.collective._mp_allreduce(
                 output_parallel,
                 group=self.model_parallel_group,
                 use_calc_stream=True,
                 use_model_parallel=True)
+        else:
+            output_ = output_parallel
 
         output = output_ + self.bias if self.bias is not None else output_
         return output

python/paddle/distributed/fleet/meta_parallel/parallel_layers/random.py

@@ -14,6 +14,7 @@
 
 import paddle
 import contextlib
+import numpy as np
 
 __all__ = []
 
@@ -65,14 +66,18 @@ def get_rng_state_tracker():
     return RNG_STATE_TRACKER
 
 
-def model_parallel_random_seed(seed=2048):
+def model_parallel_random_seed(seed=None):
     import paddle.distributed.fleet as fleet
     hcg = fleet.get_hybrid_communicate_group()
     rank = hcg.get_model_parallel_rank()
 
-    local_seed = seed + 1024 + rank
+    if seed:
         global_seed = seed
+        local_seed = seed * 1024 + rank * 100
+    else:
+        global_seed = np.random.randint(0, 655350)
+        local_seed = np.random.randint(rank * 10000, (rank + 1) * 10000 - 1)
 
     RNG_STATE_TRACKER.reset()
-    paddle.seed(global_seed)
     RNG_STATE_TRACKER.add(MODEL_PARALLEL_RNG, local_seed)
+    paddle.seed(global_seed)

python/paddle/distributed/fleet/meta_parallel/model_parallel.py → python/paddle/distributed/fleet/meta_parallel/tensor_parallel.py

@@ -22,15 +22,15 @@ from ..utils.log_util import logger
 __all__ = []
 
 
-class ModelParallel(MetaParallelBase):
+class TensorParallel(MetaParallelBase):
     def __init__(self, layers, hcg, **kwargs):
-        super(ModelParallel, self).__init__(layers, hcg, **kwargs)
+        super(TensorParallel, self).__init__(layers, hcg, **kwargs)
 
     def _prepare_for_model(self):
         logger.info("start broadcast mp parameters")
         broadcast_mp_parameters(self._layers, self._hcg)
 
-        logger.info("start broadcast mp parameters")
+        logger.info("start broadcast dp parameters")
         broadcast_dp_parameters(self._layers, self._hcg)
 
         logger.info("mp's parameters is ready")

python/paddle/distributed/fleet/utils/hybrid_parallel_util.py

@@ -44,7 +44,15 @@ def _apply_collective_grads(parameters, comm_group):
 
     for coalesced_grad, _, _ in coalesced_grads_and_vars:
         # need to div nranks
-        coalesced_grad = coalesced_grad / comm_group.nranks
+        div_factor = paddle.to_tensor(
+            comm_group.nranks, dtype=coalesced_grad.dtype)
+        paddle.fluid.framework._dygraph_tracer().trace_op(
+            type="elementwise_div",
+            inputs={'X': coalesced_grad,
+                    'Y': div_factor},
+            outputs={'Out': coalesced_grad},
+            attrs={'axis': -1})
+
         paddle.distributed.all_reduce(coalesced_grad, group=comm_group)
 
     _split_tensors(coalesced_grads_and_vars)

python/paddle/fluid/tests/unittests/hybrid_parallel_mp_layers.py

@@ -231,7 +231,7 @@ class TestDistTraning(unittest.TestCase):
         # model_b
         check_group = dist.new_group(list(range(self.model_parallel_size)))
         integral_w = []
-        partial_w = model_a.embedding.embedding.weight.clone().detach()
+        partial_w = model_a.embedding.weight.clone().detach()
         paddle.distributed.all_gather(integral_w, partial_w, group=check_group)
         result_w = []
         for idx in range(len(integral_w)):

python/paddle/fluid/tests/unittests/new_group.py

@@ -27,6 +27,7 @@ class TestNewGroupAPI(object):
 
     def test_all(self):
         gp = paddle.distributed.new_group([0, 1])
+        print("gp info:", gp)
         print("test new group api ok")
 
         tmp = np.array([0, 0, 0])