[Dygraph] Support param groups in grad_clip (#39175)

* support param groups in grad_clip

* update

* modify for review

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

@@ -49,8 +49,6 @@ class HybridParallelClipGrad:
 
     @imperative_base.no_grad
     def _dygraph_clip(self, params_grads):
-        params_and_grads = []
-
         sum_square_dist_fp16 = []
         sum_square_dist_fp32 = []
         sum_square_not_dist_fp16 = []
@@ -153,15 +151,14 @@ class HybridParallelClipGrad:
             if g is None:
                 continue
             if getattr(p, 'need_clip', True) is False:
-                params_and_grads.append((p, g))
                 continue
             if p.dtype == paddle.float16:
-                new_grad = layers.elementwise_mul(x=g, y=clip_var_fp16)
+                g.scale_(clip_var_fp16)
             else:
-                new_grad = layers.elementwise_mul(x=g, y=clip_var)
-            params_and_grads.append((p, new_grad))
+                g.scale_(clip_var)
+            p._reset_grad_inplace_version(True)
 
-        return params_and_grads
+        return params_grads
 
     def __getattr__(self, item):
         return getattr(self._clip, item)
@@ -201,6 +198,12 @@ class HybridParallelOptimizer:
             else:
                 self._inner_opt._grad_clip = HybridParallelClipGrad(
                     self._inner_opt._grad_clip, hcg)
+                if self._inner_opt._parameter_list and isinstance(
+                        self._inner_opt._parameter_list[0], dict):
+                    for item in self._inner_opt._param_groups:
+                        if "grad_clip" in item.keys():
+                            item["grad_clip"] = HybridParallelClipGrad(
+                                self._inner_opt._grad_clip, hcg)
 
     @imperative_base.no_grad
     @framework.dygraph_only

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/sharding_optimizer_stage2.py

@@ -109,6 +109,13 @@ class ShardingOptimizerStage2(Optimizer):
             self._optim._grad_clip = ShardingClipGrad(self._optim._grad_clip,
                                                       paddle.get_device(),
                                                       self.group)
+            if self._optim._parameter_list and isinstance(
+                    self._optim._parameter_list[0], dict):
+                for item in self._optim._param_groups:
+                    if "grad_clip" in item.keys():
+                        item["grad_clip"] = ShardingClipGrad(
+                            self._optim._grad_clip,
+                            paddle.get_device(), self.group)
 
         if offload:
             assert self._pfp16, "Only support offload strategy while using \'Adam\', \'AdamW\' and \'Momentum\' optimizer with AMP/Pure FP16"

python/paddle/distributed/fleet/meta_parallel/sharding/sharding_utils.py

@@ -57,8 +57,6 @@ class ShardingClipGrad:
 
     @imperative_base.no_grad
     def _dygraph_clip(self, params_grads):
-        params_and_grads = []
-
         sum_square_fp16 = []
         sum_square_fp32 = []
 
@@ -114,15 +112,14 @@ class ShardingClipGrad:
             if g is None:
                 continue
             if getattr(p, 'need_clip', True) is False:
-                params_and_grads.append((p, g))
                 continue
             if p.dtype == paddle.float16:
-                new_grad = layers.elementwise_mul(x=g, y=clip_var_fp16)
+                g.scale_(clip_var_fp16)
             else:
-                new_grad = layers.elementwise_mul(x=g, y=clip_var)
-            params_and_grads.append((p, new_grad))
+                g.scale_(clip_var)
+            p._reset_grad_inplace_version(True)
 
-        return params_and_grads
+        return params_grads
 
     def __getattr__(self, item):
         return getattr(self._clip, item)

python/paddle/fluid/tests/unittests/dygraph_sharding_stage2.py

@@ -159,10 +159,13 @@ def test_dp_stage2():
     mlp2 = MLP()
     mlp3 = MLP()
     mlp4 = MLP()
+    mlp5 = MLP()
     mlp1.set_state_dict(state_dict)
     mlp2.set_state_dict(state_dict)
     mlp3.set_state_dict(state_dict)
     mlp4.set_state_dict(state_dict)
+    mlp5.set_state_dict(state_dict)
+
     dp_params = train_mlp(
         mlp1, sharding_stage="dp", use_pure_fp16=False, opt_group=False)
     stage2_params = train_mlp(
@@ -181,6 +184,11 @@ def test_dp_stage2():
             rtol=1e-5,
             atol=1e-5)
 
+    stage2_params = train_mlp(
+        mlp2, sharding_stage=2, use_pure_fp16=False, opt_group=True)
+    for i in range(len(dp_params)):
+        np.testing.assert_allclose(
+            dp_params[i].numpy(), stage2_params[i].numpy(), rtol=1e-6)
     return
 
 

python/paddle/fluid/tests/unittests/dygraph_sharding_stage2_offload.py

@@ -49,7 +49,7 @@ def train_mlp(model, offload=False):
     optimizer = ShardingOptimizerStage2(
         params=model.parameters(), optim=optimizer, offload=offload)
     model = ShardingStage2(
-        model, optimizer, buffer_max_size=2**21, accumulate_grads=True)
+        model, optimizer, buffer_max_size=2**21, accumulate_grads=False)
 
     train_reader = paddle.batch(
         reader_decorator(linear_size), batch_size=batch_size, drop_last=True)
@@ -98,12 +98,11 @@ def test_sharding_stage2_offload():
     mlp_offload_params = train_mlp(mlp_offload, offload=True)
 
     for i in range(len(mlp_params)):
-        for j in range(len(mlp_offload_params)):
-            if mlp_params[i].name == mlp_offload_params[j].name:
         np.testing.assert_allclose(
             mlp_params[i].numpy(),
-                    mlp_offload_params[j].numpy(),
-                    rtol=1e-6)
+            mlp_offload_params[i].numpy(),
+            rtol=5e-3,
+            atol=5e-3)
     return
 
 

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

@@ -31,5 +31,19 @@ class TestPPClipGrad(TestDistPPTraning):
         return scheduler, optimizer
 
 
+class TestPPClipGradParamGroup(TestDistPPTraning):
+    def build_optimizer(self, model):
+        grad_clip = paddle.nn.ClipGradByGlobalNorm(0.5)
+        scheduler = paddle.optimizer.lr.PiecewiseDecay(
+            boundaries=[2], values=[0.001, 0.002], verbose=True)
+        optimizer = paddle.optimizer.Momentum(
+            learning_rate=scheduler,
+            grad_clip=grad_clip,
+            parameters=[{
+                "params": model.parameters()
+            }])
+        return scheduler, optimizer
+
+
 if __name__ == "__main__":
     unittest.main()