Fix for OOM Errors during Ultrachat200k Finetuning (#2180) (#2181)

* testing fix

* get rid of repeated log

* revert yaml

Author: Sara Adkins

src/sparseml/modifiers/pruning/constant/pytorch.py

@@ -71,6 +71,9 @@ def on_update(self, state: State, event: Event, **kwargs):
             def apply_masks(module):
                 mask_name = param_mask_name()
                 if hasattr(module, mask_name):
+                    mask = getattr(module, mask_name)
+                    if mask.device != module.weight.device:
+                        setattr(module, mask_name, mask.to(module.weight.device))
                     module.weight *= getattr(module, mask_name)
 
             state.model.model.apply(apply_masks)

src/sparseml/transformers/finetune/runner.py

@@ -40,9 +40,7 @@
 )
 from sparseml.transformers.finetune.model_args import ModelArguments
 from sparseml.transformers.finetune.training_args import TrainingArguments
-from sparseml.utils.fsdp.context import summon_full_params_context
 from sparseml.utils.fsdp.helpers import is_fsdp_model, unwrap_and_export_model
-from sparseml.utils.pytorch import qat_active
 
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)
@@ -287,12 +285,6 @@ def run_sequential_stages(self, checkpoint: Optional[str] = None):
             session = session_manager.active_session()
             session.reset_stage()
 
-            # log model sparsity
-            with summon_full_params_context(self.trainer.model):
-                if self.trainer.accelerator.is_main_process:
-                    if not qat_active(self.trainer.model):
-                        self.trainer.log_model_sparsification()
-
             # synchronize and clean up memory
             self.trainer.accelerator.wait_for_everyone()
             self.trainer.model = get_session_model()

src/sparseml/transformers/finetune/session_mixin.py

@@ -40,6 +40,7 @@
 )
 from sparseml.utils.fsdp.context import summon_full_params_context
 from sparseml.utils.fsdp.helpers import is_fsdp_model, save_pretrained_fsdp
+from sparseml.utils.pytorch import qat_active
 
 
 __all__ = [
@@ -137,7 +138,7 @@ def initialize_session(
         train_data = self.get_train_dataloader()
 
         self.accelerator.wait_for_everyone()
-        with summon_full_params_context(self.model):
+        with summon_full_params_context(self.model, offload_to_cpu=True):
             session_manager.initialize(
                 model=self.model,
                 teacher_model=self.teacher,  # TODO: what about for self/disable?
@@ -370,9 +371,13 @@ def train(self, *args, stage: Optional[str] = None, **kwargs):
 
         self.accelerator.wait_for_everyone()
 
-        # Need to gather parameters across the GPUs before accessing layer weights
-        with summon_full_params_context(self.model):
-            self.log_model_sparsification()
+        # log model sparsity
+        with summon_full_params_context(self.model, offload_to_cpu=True):
+            if self.accelerator.is_main_process:
+                if not qat_active(self.model):
+                    self.log_model_sparsification()
+
+        self.accelerator.wait_for_everyone()
 
         return output
 
@@ -434,6 +439,12 @@ def one_shot(self, calib_data: DataLoader, stage: Optional[str] = None):
             accelerator=self.accelerator,
         )
 
+        # log model sparsity
+        with summon_full_params_context(self.model, offload_to_cpu=True):
+            if self.accelerator.is_main_process:
+                if not qat_active(self.model):
+                    self.log_model_sparsification()
+
         self.accelerator.wait_for_everyone()
 
     def save_model(