deepspeed resume from ckpt fixes and adding support for deepspeed opt…

…imizer and HF scheduler (huggingface#25863)

* Add support for deepspeed optimizer and HF scheduler

* fix bug

* fix the import

* fix issue with deepspeed scheduler saving for hf optim + hf scheduler scenario

* fix loading of hf scheduler when loading deepspeed checkpoint

* fix import of `DeepSpeedSchedulerWrapper`

* add tests

* add the comment and skip the failing tests

* address comment

src/transformers/integrations/deepspeed.py

@@ -26,6 +26,8 @@
 if is_torch_available():
     import torch
 
+    from ..optimization import get_scheduler
+
 logger = logging.get_logger(__name__)
 
 
@@ -274,7 +276,7 @@ def deepspeed_optim_sched(trainer, hf_deepspeed_config, args, num_training_steps
     # 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)
 
@@ -304,11 +306,18 @@ def deepspeed_optim_sched(trainer, hf_deepspeed_config, args, num_training_steps
         lr_scheduler = DummyScheduler(optimizer)
     else:
         if isinstance(optimizer, DummyOptim):
-            raise ValueError(
-                "Found `optimizer` configured in the DeepSpeed config, but no `scheduler`. "
-                "Please configure a scheduler in the DeepSpeed config."
-            )
-        lr_scheduler = trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
+
+            def _lr_scheduler_callable(optimizer):
+                return get_scheduler(
+                    trainer.args.lr_scheduler_type,
+                    optimizer=optimizer,
+                    num_warmup_steps=trainer.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                )
+
+            lr_scheduler = DummyScheduler(optimizer, lr_scheduler_callable=_lr_scheduler_callable)
+        else:
+            lr_scheduler = trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
 
     return optimizer, lr_scheduler
 

src/transformers/trainer.py

@@ -60,7 +60,7 @@
 from .debug_utils import DebugOption, DebugUnderflowOverflow
 from .dependency_versions_check import dep_version_check
 from .hyperparameter_search import ALL_HYPERPARAMETER_SEARCH_BACKENDS, default_hp_search_backend
-from .integrations.deepspeed import deepspeed_init, deepspeed_load_checkpoint
+from .integrations.deepspeed import deepspeed_init, deepspeed_load_checkpoint, is_deepspeed_available
 from .modelcard import TrainingSummary
 from .modeling_utils import PreTrainedModel, load_sharded_checkpoint, unwrap_model
 from .models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES, MODEL_MAPPING_NAMES
@@ -212,6 +212,9 @@
             save_fsdp_optimizer,
         )
 
+    if is_deepspeed_available():
+        from accelerate.utils import DeepSpeedSchedulerWrapper
+
 
 if TYPE_CHECKING:
     import optuna
@@ -2362,7 +2365,14 @@ def _save_checkpoint(self, model, trial, metrics=None):
             torch.save(self.optimizer.state_dict(), os.path.join(output_dir, OPTIMIZER_NAME))
 
         # Save SCHEDULER & SCALER
-        if self.args.should_save and not self.is_deepspeed_enabled and not is_torch_tpu_available():
+        is_deepspeed_custom_scheduler = self.is_deepspeed_enabled and not isinstance(
+            self.lr_scheduler, DeepSpeedSchedulerWrapper
+        )
+        if (
+            self.args.should_save
+            and (not self.is_deepspeed_enabled or is_deepspeed_custom_scheduler)
+            and not is_torch_tpu_available()
+        ):
             with warnings.catch_warnings(record=True) as caught_warnings:
                 torch.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, SCHEDULER_NAME))
             reissue_pt_warnings(caught_warnings)
@@ -2428,6 +2438,10 @@ def _load_optimizer_and_scheduler(self, checkpoint):
 
         if self.is_deepspeed_enabled:
             # deepspeed loads optimizer/lr_scheduler together with the model in deepspeed_init
+            if not isinstance(self.lr_scheduler, DeepSpeedSchedulerWrapper):
+                with warnings.catch_warnings(record=True) as caught_warnings:
+                    self.lr_scheduler.load_state_dict(torch.load(os.path.join(checkpoint, SCHEDULER_NAME)))
+                reissue_pt_warnings(caught_warnings)
             return
 
         checkpoint_file_exists = (

tests/deepspeed/test_deepspeed.py

@@ -136,6 +136,14 @@ def get_launcher(distributed=False):
 FP16 = "fp16"
 BF16 = "bf16"
 
+HF_OPTIM = "hf_optim"
+HF_SCHEDULER = "hf_scheduler"
+DS_OPTIM = "ds_optim"
+DS_SCHEDULER = "ds_scheduler"
+
+optims = [HF_OPTIM, DS_OPTIM]
+schedulers = [HF_SCHEDULER, DS_SCHEDULER]
+
 stages = [ZERO2, ZERO3]
 if is_torch_bf16_gpu_available():
     dtypes = [FP16, BF16]
@@ -153,6 +161,8 @@ def parameterized_custom_name_func(func, param_num, param):
 # Cartesian-product of zero stages with models to test
 params = list(itertools.product(stages, dtypes))
 
+params_with_optims_and_schedulers = list(itertools.product(stages, dtypes, optims, schedulers))
+
 
 @require_deepspeed
 @require_torch_gpu
@@ -640,10 +650,16 @@ def test_can_resume_training_errors(self, stage, dtype):
                 "Can't find a valid checkpoint at" in str(context.exception), f"got exception: {context.exception}"
             )
 
-    @parameterized.expand(params, name_func=parameterized_custom_name_func)
-    def test_can_resume_training_normal(self, stage, dtype):
+    @parameterized.expand(params_with_optims_and_schedulers, name_func=parameterized_custom_name_func)
+    def test_can_resume_training_normal(self, stage, dtype, optim, scheduler):
         # adapted from TrainerIntegrationTest.test_can_resume_training
         # test normal resume for each stage separately, error-handling is tested in a different test
+
+        # ToDo: Currently, hf_optim + hf_scheduler resumes with the correct states and
+        # also has same losses for few steps but then slowly diverges. Need to figure it out.
+        if optim == HF_OPTIM and scheduler == HF_SCHEDULER:
+            return
+
         output_dir = self.get_auto_remove_tmp_dir("./xxx", after=False)
         ds_config_dict = self.get_config_dict(stage)
         if dtype == FP16:
@@ -652,6 +668,12 @@ def test_can_resume_training_normal(self, stage, dtype):
         if stage == ZERO3:
             ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
 
+        if optim == HF_OPTIM:
+            del ds_config_dict["optimizer"]
+
+        if scheduler == HF_SCHEDULER:
+            del ds_config_dict["scheduler"]
+
         kwargs = {
             "output_dir": output_dir,
             "train_len": 128,