[Deepspeed Inference] HF Integration

src/transformers/deepspeed.py

@@ -33,6 +33,54 @@
 logger = logging.get_logger(__name__)
 
 
+inference_custom_map = dict(
+    electra=dict(ElectraLayer=("output.dense")),
+    roberta=dict(RobertaLayer=("output.dense")),
+    t5=dict(T5Block=("SelfAttention.o", "EncDecAttention.o", "DenseReluDense.wo")),
+    albert=dict(AlbertLayer=("attention.dense", "ffn_output")),
+    bart=dict(BartEncoderLayer=("self_attn.out_proj", "fc2")),
+    deberta=dict(DebertaLayer=("output.dense")),
+    deberta_v2=dict(DebertaV2Layer=("output.dense")),
+    wav2vec2=dict(Wav2Vec2EncoderLayer=("attention.out_proj", "feed_forward.output_dense")),
+)
+
+inference_auto_map = ["gpt_neo", "gptj", "gpt2", "bert"]
+
+
+def deepspeed_inference_init(trainer):
+    """
+    XXX:
+    """
+
+    dep_version_check("deepspeed")
+    import deepspeed
+
+    args = trainer.args
+
+    model_arch = trainer.model.config.model_type
+
+    if model_arch in inference_auto_map:
+        kwargs = dict(
+            replace_method="auto",
+            replace_with_kernel_inject=True,
+        )
+    elif model_arch in inference_custom_map:
+        kwargs = dict(injection_policy=inference_custom_map[model_arch])
+    else:
+        raise ValueError(
+            f"[Deepspeed Inference] {model_arch} hasn't yet been mapped out, please file an Issue to request support for it"
+        )
+
+    deepspeed_inference_engine = deepspeed.init_inference(
+        trainer.model,
+        mp_size=args.world_size,
+        dtype=torch.half if args.fp16 else torch.float,  # XXX: add bf16 once ds supports it
+        **kwargs,
+    )
+
+    return deepspeed_inference_engine
+
+
 def is_deepspeed_available():
     return importlib.util.find_spec("deepspeed") is not None
 

src/transformers/trainer.py

@@ -61,7 +61,7 @@
 from .configuration_utils import PretrainedConfig
 from .data.data_collator import DataCollator, DataCollatorWithPadding, default_data_collator
 from .debug_utils import DebugOption, DebugUnderflowOverflow
-from .deepspeed import deepspeed_init, deepspeed_reinit, is_deepspeed_zero3_enabled
+from .deepspeed import deepspeed_inference_init, deepspeed_init, deepspeed_reinit, is_deepspeed_zero3_enabled
 from .dependency_versions_check import dep_version_check
 from .file_utils import (
     CONFIG_NAME,
@@ -362,6 +362,7 @@ def __init__(
         if (
             self.is_model_parallel
             or args.deepspeed
+            or args.deepspeed_inference
             or ((args.fp16_full_eval or args.bf16_full_eval) and not args.do_train)
             or (self.sharded_ddp in [ShardedDDPOption.ZERO_DP_2, ShardedDDPOption.ZERO_DP_3])
         ):
@@ -1829,6 +1830,11 @@ def _prepare_input(self, data: Union[torch.Tensor, Any]) -> Union[torch.Tensor,
             return type(data)(self._prepare_input(v) for v in data)
         elif isinstance(data, torch.Tensor):
             kwargs = dict(device=self.args.device)
+            # if self.args.deepspeed_inference:
+            #     print(data.dtype)
+            #     print(kwargs)
+            #     return data.to("cuda:0")
+
             if self.deepspeed and data.dtype != torch.int64:
                 # NLP models inputs are int64 and those get adjusted to the right dtype of the
                 # embedding. Other models such as wav2vec2's inputs are already float and thus
@@ -2274,6 +2280,12 @@ def evaluation_loop(
             self.model_wrapped = deepspeed_engine
             self.deepspeed = deepspeed_engine
 
+        if self.args.deepspeed_inference:
+            deepspeed_inference_engine = deepspeed_inference_init(self)
+            self.model = deepspeed_inference_engine.module
+            self.model_wrapped = deepspeed_inference_engine
+            self.deepspeed = deepspeed_inference_engine
+
         model = self._wrap_model(self.model, training=False)
 
         # if full fp16 or bf16 eval is wanted and this ``evaluation`` or ``predict`` isn't called
@@ -2437,6 +2449,10 @@ def _pad_across_processes(self, tensor, pad_index=-100):
         Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so
         they can safely be gathered.
         """
+        # XXX: hangs here with 2 gpus if we don't return
+        # if self.args.deepspeed_inference:
+        #     return tensor
+
         if isinstance(tensor, (list, tuple)):
             return type(tensor)(self._pad_across_processes(t, pad_index=pad_index) for t in tensor)
         elif isinstance(tensor, dict):

src/transformers/training_args.py

@@ -313,6 +313,9 @@ class TrainingArguments:
             Use [Deepspeed](https://github.com/microsoft/deepspeed). This is an experimental feature and its API may
             evolve in the future. The value is either the location of DeepSpeed json config file (e.g.,
             `ds_config.json`) or an already loaded json file as a `dict`"
+        deepspeed_inference (`bool`, *optional*):
+            Enable [Deepspeed Inference](https://www.deepspeed.ai/tutorials/inference-tutorial). This is an
+            experimental feature and its API may change in the future.
         label_smoothing_factor (`float`, *optional*, defaults to 0.0):
             The label smoothing factor to use. Zero means no label smoothing, otherwise the underlying onehot-encoded
             labels are changed from 0s and 1s to `label_smoothing_factor/num_labels` and `1 - label_smoothing_factor +
@@ -635,9 +638,13 @@ class TrainingArguments:
     deepspeed: Optional[str] = field(
         default=None,
         metadata={
-            "help": "Enable deepspeed and pass the path to deepspeed json config file (e.g. ds_config.json) or an already loaded json file as a dict"
+            "help": "Enable DeepSpeed and pass the path to deepspeed json config file (e.g. ds_config.json) or an already loaded json file as a dict"
         },
     )
+    deepspeed_inference: bool = field(
+        default=False,
+        metadata={"help": "Enable DeepSpeed Inference"},
+    )
     label_smoothing_factor: float = field(
         default=0.0, metadata={"help": "The label smoothing epsilon to apply (zero means no label smoothing)."}
     )
@@ -992,7 +999,7 @@ def _setup_devices(self) -> "torch.device":
             self.local_rank = sm_dist.get_local_rank()
             device = torch.device("cuda", self.local_rank)
             self._n_gpu = 1
-        elif self.deepspeed:
+        elif self.deepspeed or self.deepspeed_inference:
             # deepspeed inits torch.distributed internally
             from .deepspeed import is_deepspeed_available