remove SharedDDP as it is deprecated (intel#1103)

* remove SharedDDP as it is deprecated

.github/workflows/script/unitTest/run_unit_test_optimize.sh

@@ -21,7 +21,7 @@ function pytest() {
     mkdir -p ${coverage_log_dir}
     pip install --no-cache-dir protobuf==3.20.0
     ## install transformers==4.34.1, to work with SharedDPO API
-    pip install transformers==4.34.1
+    pip install transformers
     cd /intel-extension-for-transformers/tests/CI || exit 1
     JOB_NAME=unit_test
     ut_log_name=${LOG_DIR}/${JOB_NAME}.log

intel_extension_for_transformers/langchain/embeddings/optimized_sentence_transformers.py

@@ -144,4 +144,4 @@ def _load_sbert_model(
                 module = module_class.load(module_path)
             modules[module_config['name']] = module
 
-        return modules
+        return modules

intel_extension_for_transformers/neural_chat/examples/finetuning/instruction/finetune_on_Intel_Xeon_CPU.ipynb

@@ -384,7 +384,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",
@@ -1526,7 +1525,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",

intel_extension_for_transformers/neural_chat/examples/finetuning/instruction/finetune_on_Nvidia_GPU.ipynb

@@ -251,7 +251,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",
@@ -740,7 +739,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",
@@ -1322,7 +1320,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",
@@ -1807,7 +1804,6 @@
       "save_strategy=no,\n",
       "save_total_limit=2,\n",
       "seed=42,\n",
-      "sharded_ddp=[],\n",
       "skip_memory_metrics=True,\n",
       "tf32=None,\n",
       "torch_compile=False,\n",

intel_extension_for_transformers/neural_chat/examples/finetuning/multi_modal/llava_trainer.py

@@ -26,7 +26,6 @@
     get_parameter_names,
     has_length,
     ALL_LAYERNORM_LAYERS,
-    ShardedDDPOption,
     logger,
 )
 from typing import List, Optional
@@ -176,7 +175,7 @@ def create_optimizer(self):
         """
         if is_sagemaker_mp_enabled():
             return super().create_optimizer()
-        if self.sharded_ddp == ShardedDDPOption.SIMPLE:
+        if self.is_fsdp_enabled:
             return super().create_optimizer()
 
         opt_model = self.model
@@ -237,27 +236,20 @@ def create_optimizer(self):
 
             optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
 
-            if self.sharded_ddp == ShardedDDPOption.SIMPLE:
-                self.optimizer = OSS(
-                    params=optimizer_grouped_parameters,
-                    optim=optimizer_cls,
-                    **optimizer_kwargs,
-                )
-            else:
-                self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                if optimizer_cls.__name__ == "Adam8bit":
-                    import bitsandbytes
+            self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
 
-                    manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
 
-                    skipped = 0
-                    for module in opt_model.modules():
-                        if isinstance(module, nn.Embedding):
-                            skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
-                            logger.info(f"skipped {module}: {skipped/2**20}M params")
-                            manager.register_module_override(module, "weight", {"optim_bits": 32})
-                            logger.debug(f"bitsandbytes: will optimize {module} in fp32")
-                    logger.info(f"skipped: {skipped/2**20}M params")
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                        logger.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(module, "weight", {"optim_bits": 32})
+                        logger.debug(f"bitsandbytes: will optimize {module} in fp32")
+                logger.info(f"skipped: {skipped/2**20}M params")
 
         return self.optimizer
 
@@ -297,7 +289,6 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None):
             get_parameter_names,
             has_length,
             ALL_LAYERNORM_LAYERS,
-            ShardedDDPOption,
             logger,
             )
     from typing import List, Optional
@@ -328,7 +319,7 @@ def create_optimizer(self):
             """
             if is_sagemaker_mp_enabled():
                 return super().create_optimizer()
-            if self.sharded_ddp == ShardedDDPOption.SIMPLE:
+            if self.is_fsdp_enabled:
                 return super().create_optimizer()
 
             opt_model = self.model
@@ -401,27 +392,20 @@ def create_optimizer(self):
 
                 # optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
 
-                if self.sharded_ddp == ShardedDDPOption.SIMPLE:
-                    self.optimizer = OSS(
-                        params=optimizer_grouped_parameters,
-                        optim=optimizer_cls,
-                        **optimizer_kwargs,
-                    )
-                else:
-                    self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                    if optimizer_cls.__name__ == "Adam8bit":
-                        import bitsandbytes
-
-                        manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
-
-                        skipped = 0
-                        for module in opt_model.modules():
-                            if isinstance(module, nn.Embedding):
-                                skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
-                                logger.info(f"skipped {module}: {skipped/2**20}M params")
-                                manager.register_module_override(module, "weight", {"optim_bits": 32})
-                                logger.debug(f"bitsandbytes: will optimize {module} in fp32")
-                        logger.info(f"skipped: {skipped/2**20}M params")
+                self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+                if optimizer_cls.__name__ == "Adam8bit":
+                    import bitsandbytes
+
+                    manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                    skipped = 0
+                    for module in opt_model.modules():
+                        if isinstance(module, nn.Embedding):
+                            skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                            logger.info(f"skipped {module}: {skipped/2**20}M params")
+                            manager.register_module_override(module, "weight", {"optim_bits": 32})
+                            logger.debug(f"bitsandbytes: will optimize {module} in fp32")
+                    logger.info(f"skipped: {skipped/2**20}M params")
 
             return self.optimizer
 

intel_extension_for_transformers/transformers/modeling/llava_models/llava_mistral.py

@@ -31,7 +31,7 @@
 
 
 class LlavaConfig(MistralConfig):
-    model_type = "llava"
+    model_type = "llava_custom"
 
 
 class LlavaMistralModel(LlavaMetaModel, MistralModel):
@@ -110,5 +110,5 @@ def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_
             _inputs['images'] = images
         return _inputs
 
-AutoConfig.register("llava", LlavaConfig)
+AutoConfig.register("llava_custom", LlavaConfig)
 AutoModelForCausalLM.register(LlavaConfig, LlavaMistralForCausalLM)

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -53,6 +53,7 @@
     QUANT_CONFIG,
     WEIGHTS_NAME,
     WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
 )
 from intel_extension_for_transformers.llm.quantization.utils import replace_linear
 from transformers.configuration_utils import PretrainedConfig
@@ -727,6 +728,13 @@ def load_low_bit(cls,
                         pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
                     )
                     is_sharded = True
+                elif os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
+                ):
+                    # Load from a safetensors checkpoint
+                    archive_file = os.path.join(
+                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
+                    )
             elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
                 archive_file = pretrained_model_name_or_path
                 is_local = True

intel_extension_for_transformers/transformers/modeling/modeling_base_seq2seq.py

@@ -25,9 +25,7 @@
 from neural_compressor.utils.pytorch import load
 from transformers import AutoModel, PretrainedConfig
 from transformers.file_utils import add_start_docstrings
-from transformers.modeling_utils import no_init_weights
 from transformers.models.auto.auto_factory import _get_model_class
-from transformers.utils.generic import ContextManagers
 from optimum.exporters import TasksManager
 
 from optimum.intel.neural_compressor import INCConfig
@@ -268,9 +266,7 @@ def _from_pretrained(
                 decoder = model
             else:
                 model_class = _get_model_class(config, cls.auto_model_class._model_mapping)
-                init_contexts = [no_init_weights(_enable=True)]
-                with ContextManagers(init_contexts):
-                    model = model_class(config)
+                model = model_class(config)
 
                 # Load the model from local directory
                 if os.path.isdir(model_id):

intel_extension_for_transformers/transformers/trainer.py

@@ -49,7 +49,7 @@
 from transformers import __version__, Seq2SeqTrainer, Trainer, PreTrainedModel
 from transformers.configuration_utils import PretrainedConfig
 from transformers.debug_utils import DebugOption, DebugUnderflowOverflow
-from transformers.file_utils import (
+from transformers.utils import (
     CONFIG_NAME,
     WEIGHTS_NAME,
     is_torch_tpu_available,
@@ -67,7 +67,6 @@
 )
 from transformers.trainer_utils import (
     HPSearchBackend,
-    ShardedDDPOption,
     TrainOutput,
     EvalLoopOutput,
     EvalPrediction,
@@ -762,7 +761,8 @@ def train(
             else:
                 debug_overflow = DebugUnderflowOverflow(self.model)  # noqa
 
-        delay_optimizer_creation = self.sharded_ddp is not None and self.sharded_ddp != ShardedDDPOption.SIMPLE
+        # delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled or self.is_fsdp_enabled
+        delay_optimizer_creation = is_sagemaker_mp_enabled()
 
         if not delay_optimizer_creation:
             self.create_optimizer_and_scheduler(num_training_steps=max_steps)
@@ -1176,9 +1176,7 @@ def training_step(
             else:
                 loss.backward()
         else:
-            if self.do_grad_scaling:
-                self.scaler.scale(loss).backward()
-            elif self.use_apex:
+            if self.use_apex:
                 with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                     scaled_loss.backward()
             elif NEW_DEEPSPEED_FLAG:
@@ -1265,9 +1263,7 @@ def training_step_length_adaptive(
             else:
                 loss.backward()
         else:
-            if self.do_grad_scaling:
-                self.scaler.scale(loss).backward()
-            elif self.use_apex:
+            if self.use_apex:
                 with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                     scaled_loss.backward()
             elif NEW_DEEPSPEED_FLAG:
@@ -1360,9 +1356,7 @@ def training_step_length_adaptive(
                 else:
                     loss.backward()
             else:
-                if self.do_grad_scaling:
-                    self.scaler.scale(loss).backward()
-                elif self.use_apex:
+                if self.use_apex:
                     with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                         scaled_loss.backward()
                 elif NEW_DEEPSPEED_FLAG:

intel_extension_for_transformers/transformers/utils/utility.py

@@ -34,6 +34,7 @@
 WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
 QUANT_CONFIG = "quantization_config.json"
 SPARSITY_CONFIG = "sparsity_config.json"
+SAFE_WEIGHTS_NAME = "model.safetensors"
 
 torch = LazyImport("torch")
 

tests/CI/test_weight_only.py

@@ -153,7 +153,7 @@ def test_auto_model_saving_loading(self):
             if isinstance(module, QuantizedLinearQBits):
                 module_list.append(name)
         self.assertTrue(len(module_list) > 0)
-        model.save_pretrained(self.workspace)
+        model.save_pretrained(self.workspace, safe_serialization=False)
         loaded_model = AutoModelForCausalLM.from_pretrained(self.workspace)
         for name, module in loaded_model.named_modules():
             if isinstance(module, QuantizedLinearQBits):