[deepspeed] zero inference (#14253)

* [deepspeed] zero inference

* only z3 makes sense for inference

* fix and style

* docs

* rework

* fix test

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* responding to suggestions

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

docs/source/main_classes/deepspeed.rst

@@ -46,6 +46,20 @@ won't be possible on a single GPU.
    parts of DeepSpeed like ``zero.Init`` for ZeRO stage 3 and higher. To tap into this feature read the docs on
    :ref:`deepspeed-non-trainer-integration`.
 
+What is integrated:
+
+Training:
+
+1. DeepSpeed ZeRO training supports the full ZeRO stages 1, 2 and 3 with ZeRO-Infinity (CPU and NVME offload).
+
+Inference:
+
+1. DeepSpeed ZeRO Inference supports ZeRO stage 3 with ZeRO-Infinity. It uses the same ZeRO protocol as training, but
+   it doesn't use an optimizer and a lr scheduler and only stage 3 is relevant. For more details see:
+   :ref:`deepspeed-zero-inference`.
+
+There is also DeepSpeed Inference - this is a totally different technology which uses Tensor Parallelism instead of
+ZeRO (coming soon).
 
 
 
@@ -1628,6 +1642,47 @@ larger multi-dimensional shape, this means that the parameter is partitioned and
 
 
 
+.. _deepspeed-zero-inference:
+
+
+ZeRO Inference
+=======================================================================================================================
+
+ZeRO Inference uses the same config as ZeRO-3 Training. You just don't need the optimizer and scheduler sections. In
+fact you can leave these in the config file if you want to share the same one with the training. They will just be
+ignored.
+
+Otherwise you just need to pass the usual :class:`~transformers.TrainingArguments` arguments. For example:
+
+.. code-block:: bash
+
+    deepspeed --num_gpus=2 your_program.py <normal cl args> --do_eval --deepspeed ds_config.json
+
+The only important thing is that you need to use a ZeRO-3 configuration, since ZeRO-2 provides no benefit whatsoever
+for the inference as only ZeRO-3 performs sharding of parameters, whereas ZeRO-1 shards gradients and optimizer states.
+
+Here is an example of running ``run_translation.py`` under DeepSpeed deploying all available GPUs:
+
+.. code-block:: bash
+
+    deepspeed examples/pytorch/translation/run_translation.py \
+    --deepspeed tests/deepspeed/ds_config_zero3.json \
+    --model_name_or_path t5-small --output_dir output_dir \
+    --do_eval --max_eval_samples 50 --warmup_steps 50  \
+    --max_source_length 128 --val_max_target_length 128 \
+    --overwrite_output_dir --per_device_eval_batch_size 4 \
+    --predict_with_generate --dataset_config "ro-en" --fp16 \
+    --source_lang en --target_lang ro --dataset_name wmt16 \
+    --source_prefix "translate English to Romanian: "
+
+Since for inference there is no need for additional large memory used by the optimizer states and the gradients you
+should be able to fit much larger batches and/or sequence length onto the same hardware.
+
+
+Additionally DeepSpeed is currently developing a related product called Deepspeed-Inference which has no relationship
+to the ZeRO technology, but instead uses tensor parallelism to scale models that can't fit onto a single GPU. This is a
+work in progress and we will provide the integration once that product is complete.
+
 
 Filing Issues
 =======================================================================================================================

setup.py

@@ -97,7 +97,7 @@
     "cookiecutter==1.7.2",
     "dataclasses",
     "datasets",
-    "deepspeed>=0.5.3",
+    "deepspeed>=0.5.7",
     "docutils==0.16.0",
     "fairscale>0.3",
     "faiss-cpu",

src/transformers/deepspeed.py

@@ -111,6 +111,29 @@ def get_value(self, ds_key_long, default=None):
             return default
         return config.get(ds_key, default)
 
+    def del_config_sub_tree(self, ds_key_long, must_exist=False):
+        """
+        Deletes a sub-section of the config file if it's found.
+
+        Unless ``must_exist`` is :obj:`True` the section doesn't have to exist.
+        """
+        config = self.config
+
+        # find the config node of interest if it exists
+        nodes = ds_key_long.split(".")
+        for node in nodes:
+            parent_config = config
+            config = config.get(node)
+            if config is None:
+                if must_exist:
+                    raise ValueError(f"Can't find {ds_key_long} entry in the config: {self.config}")
+                else:
+                    return
+
+        # if found remove it
+        if parent_config is not None:
+            parent_config.pop(node)
+
     def is_true(self, ds_key_long):
         """
         Returns :obj:`True`/:obj:`False` only if the value is set, always :obj:`False` otherwise. So use this method to
@@ -280,30 +303,10 @@ def deepspeed_config():
         return None
 
 
-def deepspeed_init(trainer, num_training_steps, resume_from_checkpoint=None):
+def deepspeed_optim_sched(trainer, hf_deepspeed_config, args, num_training_steps):
     """
-    Init DeepSpeed, after updating the DeepSpeed configuration with any relevant Trainer's args.
-
-    If ``resume_from_checkpoint`` was passed then an attempt to resume from a previously saved checkpoint will be made.
-
-    Args:
-        trainer: Trainer object
-        num_training_steps: per single gpu
-        resume_from_checkpoint: path to a checkpoint if to resume from after normal DeepSpeedEngine load
-
-    Returns: model, optimizer, lr_scheduler
-
+    A convenience wrapper that deals with optimizer and lr scheduler configuration.
     """
-    import deepspeed
-    from deepspeed.utils import logger as ds_logger
-
-    model = trainer.model
-    args = trainer.args
-
-    hf_deepspeed_config = args.hf_deepspeed_config
-    hf_deepspeed_config.trainer_config_finalize(args, model, num_training_steps)
-
-    # resume config update - some bits like `model` and `num_training_steps` only become available during train
     config = hf_deepspeed_config.config
 
     # Optimizer + Scheduler
@@ -351,13 +354,54 @@ def _lr_scheduler_callable(optimizer):
         else:
             lr_scheduler = trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
 
-    # keep for quick debug:
-    # from pprint import pprint; pprint(config)
+    return optimizer, lr_scheduler
+
+
+def deepspeed_init(trainer, num_training_steps, resume_from_checkpoint=None, inference=False):
+    """
+    Init DeepSpeed, after updating the DeepSpeed configuration with any relevant Trainer's args.
+
+    If ``resume_from_checkpoint`` was passed then an attempt to resume from a previously saved checkpoint will be made.
+
+    Args:
+        trainer: Trainer object
+        num_training_steps: per single gpu
+        resume_from_checkpoint: path to a checkpoint if to resume from after normal DeepSpeedEngine load
+        inference: launch in inference mode (no optimizer and no lr scheduler)
+
+    Returns: model, optimizer, lr_scheduler
+
+    """
+    import deepspeed
+    from deepspeed.utils import logger as ds_logger
+
+    model = trainer.model
+    args = trainer.args
+
+    # resume config update - some bits like `model` and `num_training_steps` only become available during train
+    hf_deepspeed_config = args.hf_deepspeed_config
+    hf_deepspeed_config.trainer_config_finalize(args, model, num_training_steps)
+    config = hf_deepspeed_config.config
 
-    # set the Deepspeed log level consistent with the trainer
+    # set the Deepspeed log level consistent with the Trainer
     ds_logger.setLevel(args.get_process_log_level())
 
-    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
+    if inference:
+        # only Z3 makes sense for the inference
+        if not hf_deepspeed_config.is_zero3():
+            raise ValueError("ZeRO inference only makes sense with ZeRO Stage 3 - please adjust your config")
+
+        # in case the training config is re-used for inference
+        hf_deepspeed_config.del_config_sub_tree("optimizer")
+        hf_deepspeed_config.del_config_sub_tree("lr_scheduler")
+        optimizer, lr_scheduler = None, None
+        model_parameters = None
+    else:
+        optimizer, lr_scheduler = deepspeed_optim_sched(trainer, hf_deepspeed_config, args, num_training_steps)
+        model_parameters = filter(lambda p: p.requires_grad, model.parameters())
+
+    # keep for quick debug:
+    # from pprint import pprint; pprint(config)
 
     model, optimizer, _, lr_scheduler = deepspeed.initialize(
         model=model,

src/transformers/dependency_versions_table.py

@@ -8,7 +8,7 @@
     "cookiecutter": "cookiecutter==1.7.2",
     "dataclasses": "dataclasses",
     "datasets": "datasets",
-    "deepspeed": "deepspeed>=0.5.3",
+    "deepspeed": "deepspeed>=0.5.7",
     "docutils": "docutils==0.16.0",
     "fairscale": "fairscale>0.3",
     "faiss-cpu": "faiss-cpu",

src/transformers/trainer.py

@@ -2229,15 +2229,12 @@ def evaluation_loop(
 
             # XXX: eval doesn't have `resume_from_checkpoint` arg but we should be able to do eval
             # from the checkpoint eventually
-            deepspeed_engine, _, _ = deepspeed_init(self, num_training_steps=0, resume_from_checkpoint=None)
+            deepspeed_engine, _, _ = deepspeed_init(
+                self, num_training_steps=0, resume_from_checkpoint=None, inference=True
+            )
             self.model = deepspeed_engine.module
             self.model_wrapped = deepspeed_engine
             self.deepspeed = deepspeed_engine
-            # XXX: we don't need optim/sched for inference, but this needs to be sorted out, since
-            # for example the Z3-optimizer is a must for zero3 to work even for inference - what we
-            # don't need is the deepspeed basic optimizer which is self.optimizer.optimizer
-            deepspeed_engine.optimizer.optimizer = None
-            deepspeed_engine.lr_scheduler = None
 
         model = self._wrap_model(self.model, training=False)
 

tests/deepspeed/test_deepspeed.py

@@ -697,11 +697,10 @@ class TestDeepSpeedWithLauncher(TestCasePlus):
     def test_basic_distributed(self, stage):
         self.run_and_check(stage=stage, distributed=True)
 
-    @parameterized.expand(stages)
-    def test_do_eval_no_train(self, stage):
-        # we should not fail if train is skipped
+    def test_do_eval_no_train(self):
+        # testing only zero3 since zero2 makes no sense with inference
         self.run_and_check(
-            stage=stage,
+            stage=ZERO3,
             eval_steps=1,
             distributed=False,
             do_train=False,
@@ -755,6 +754,22 @@ def test_resume_train_not_from_ds_checkpoint(self, stage):
 
         self.do_checks(output_dir, do_train=do_train, do_eval=do_eval)
 
+    @require_torch_multi_gpu
+    @parameterized.expand(["fp16", "fp32"])
+    def test_inference(self, dtype):
+        # this is just inference, so no optimizer should be loaded
+        # it only works for z3 (makes no sense with z1-z2)
+        fp16 = True if dtype == "fp16" else False
+        self.run_and_check(
+            stage=ZERO3,
+            model_name=T5_TINY,
+            distributed=True,
+            do_train=False,
+            do_eval=True,
+            quality_checks=False,
+            fp16=fp16,
+        )
+
     def do_checks(self, output_dir, do_train=True, do_eval=True, quality_checks=True):
 
         if do_train: