Use DS callable API to allow hf_scheduler + ds_optimizer (huggingface#13216)

tjruwase · stas00 · web-flow · commit 42f359d015ae · 2021-08-30T10:01:06.000-07:00
* Use DS callable API to allow hf_scheduler + ds_optimizer

* Preserve backward-compatibility

* Restore backward compatibility

* Tweak arg positioning

* Tweak arg positioning

* bump the required version

* Undo indent

* Update src/transformers/trainer.py

* style

Co-authored-by: Stas Bekman &lt;stas@stason.org&gt;
Co-authored-by: Stas Bekman &lt;stas00@users.noreply.github.com&gt;
diff --git a/setup.py b/setup.py
@@ -91,7 +91,7 @@
     "cookiecutter==1.7.2",
     "dataclasses",
     "datasets",
-    "deepspeed>=0.4.3",
+    "deepspeed>=0.5.1",
     "docutils==0.16.0",
     "fairscale>0.3",
     "faiss-cpu",
diff --git a/src/transformers/deepspeed.py b/src/transformers/deepspeed.py
@@ -311,13 +311,13 @@ def deepspeed_init(trainer, num_training_steps, resume_from_checkpoint=None):
     # 1. DS scheduler + DS optimizer: Yes
     # 2. HF scheduler + HF optimizer: Yes
     # 3. DS scheduler + HF optimizer: Yes
-    # 4. HF scheduler + DS optimizer: No
+    # 4. HF scheduler + DS optimizer: Yes
     #
     # Unless Offload is enabled in which case it's:
     # 1. DS scheduler + DS optimizer: Yes
     # 2. HF scheduler + HF optimizer: Mostly*
     # 3. DS scheduler + HF optimizer: Mostly*
-    # 4. HF scheduler + DS optimizer: No
+    # 4. HF scheduler + DS optimizer: Yes
     #
     # Mostly*: All non-native DeepSpeed optimizers that have both CPU and GPU implementation should work (except LAMB)
 
@@ -336,28 +336,20 @@ def deepspeed_init(trainer, num_training_steps, resume_from_checkpoint=None):
 
         # ds supports Adam, OneBitAdam, and Lamb optimizers and can import other optimizers from torch.
         # But trainer uses AdamW by default.
-        trainer.create_optimizer()
-        optimizer = trainer.optimizer
+        optimizer = trainer.create_optimizer()
         # To use other optimizers requires voiding warranty with: `zero_allow_untested_optimizer`
         config["zero_allow_untested_optimizer"] = True
 
-    # DS schedulers (deepspeed/runtime/lr_schedules.py):
-    #
-    # DS name      | --lr_scheduler_type  | HF func                           | Notes
-    # -------------| ---------------------|-----------------------------------|--------------------
-    # LRRangeTest  | na                   | na                                | LRRT
-    # OneCycle     | na                   | na                                | 1CLR
-    # WarmupLR     | constant_with_warmup | get_constant_schedule_with_warmup | w/ warmup_min_lr=0
-    # WarmupDecayLR| linear               | get_linear_schedule_with_warmup   |
+    def _lr_scheduler_callable(optimizer):
+        return trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
+
     lr_scheduler = None
     if "scheduler" not in config:
-        if "optimizer" in config:
-            # to make this option work, we need to init DS optimizer first, then init HS scheduler,
-            # then pass the HS scheduler to DS init, which is not possible at the moment
-            raise ValueError("At the moment HF scheduler + DeepSpeed optimizer combination is not possible")
+        if optimizer is None:
+            # Optimizer is not available, so use callable to defer lr_scheduler creation to DS init
+            lr_scheduler = _lr_scheduler_callable
         else:
-            trainer.create_scheduler(num_training_steps=num_training_steps)
-            lr_scheduler = trainer.lr_scheduler
+            lr_scheduler = trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
 
     # keep for quick debug:
     # from pprint import pprint; pprint(config)
diff --git a/src/transformers/dependency_versions_table.py b/src/transformers/dependency_versions_table.py
@@ -8,7 +8,7 @@
     "cookiecutter": "cookiecutter==1.7.2",
     "dataclasses": "dataclasses",
     "datasets": "datasets",
-    "deepspeed": "deepspeed>=0.4.3",
+    "deepspeed": "deepspeed>=0.5.1",
     "docutils": "docutils==0.16.0",
     "fairscale": "fairscale>0.3",
     "faiss-cpu": "faiss-cpu",
diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
@@ -768,7 +768,7 @@ def create_optimizer_and_scheduler(self, num_training_steps: int):
         and/or :obj:`create_scheduler`) in a subclass.
         """
         self.create_optimizer()
-        self.create_scheduler(num_training_steps)
+        self.create_scheduler(num_training_steps=num_training_steps, optimizer=self.optimizer)
 
     def create_optimizer(self):
         """
@@ -813,20 +813,24 @@ def create_optimizer(self):
         if is_sagemaker_mp_enabled():
             self.optimizer = smp.DistributedOptimizer(self.optimizer)
 
-    def create_scheduler(self, num_training_steps: int):
+        return self.optimizer
+
+    def create_scheduler(self, num_training_steps: int, optimizer: torch.optim.Optimizer = None):
         """
-        Setup the scheduler. The optimizer of the trainer must have been set up before this method is called.
+        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        passed as an argument.
 
         Args:
             num_training_steps (int): The number of training steps to do.
         """
         if self.lr_scheduler is None:
             self.lr_scheduler = get_scheduler(
                 self.args.lr_scheduler_type,
-                self.optimizer,
+                optimizer=self.optimizer if optimizer is None else optimizer,
                 num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
                 num_training_steps=num_training_steps,
             )
+        return self.lr_scheduler
 
     def num_examples(self, dataloader: DataLoader) -> int:
         """
diff --git a/tests/deepspeed/test_deepspeed.py b/tests/deepspeed/test_deepspeed.py
@@ -292,19 +292,16 @@ def test_ds_scheduler_hf_optimizer(self):
         self.assertNotEqual(new_a, a)
 
     def test_hf_scheduler_ds_optimizer(self):
-        # this combo is not possible at the moment
+        a = 0
         with mockenv_context(**self.dist_env_1_gpu):
             ds_config_zero2_dict = self.get_config_dict(ZERO2)
             del ds_config_zero2_dict["scheduler"]  # force default HF Trainer scheduler
             ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
             ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
             trainer = get_regression_trainer(local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
-            with self.assertRaises(Exception) as context:
-                trainer.train()
-        self.assertTrue(
-            "HF scheduler + DeepSpeed optimizer combination is not possible" in str(context.exception),
-            f"got exception: {context.exception}",
-        )
+            trainer.train()
+        new_a = trainer.model.a.item()
+        self.assertNotEqual(new_a, a)
 
     @require_deepspeed_aio
     def test_stage3_nvme_offload(self):
