Add FSDP config for CPU RAM efficient loading through accelerate (#30002)

* Add FSDP config for CPU RAM efficient loading

* Style fix

* Update src/transformers/training_args.py

Co-authored-by: Zach Mueller <muellerzr@gmail.com>

* Update src/transformers/training_args.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Add sync_module_states and cpu_ram_efficient_loading validation logic

* Update src/transformers/training_args.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Style

---------

Co-authored-by: Zach Mueller <muellerzr@gmail.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/training_args.py

@@ -513,6 +513,11 @@ class TrainingArguments:
                 - sync_module_states (`bool`, *optional*, defaults to `True`)
                     If `"True"`, each individually wrapped FSDP unit will broadcast module parameters from rank 0 to
                     ensure they are the same across all ranks after initialization
+                - cpu_ram_efficient_loading (`bool`, *optional*, defaults to `False`)
+                    If `"True"`, only the first process loads the pretrained model checkpoint while all other processes
+                    have empty weights.  When this setting as `"True"`, `sync_module_states` also must to be `"True"`,
+                    otherwise all the processes except the main process would have random weights leading to unexpected
+                    behaviour during training.
                 - activation_checkpointing (`bool`, *optional*, defaults to `False`):
                     If `"True"`, activation checkpointing is a technique to reduce memory usage by clearing activations of
                     certain layers and recomputing them during a backward pass. Effectively, this trades extra
@@ -1826,7 +1831,18 @@ class TrainingArguments:
             prefetch_policy = self.fsdp_config.get("backward_prefetch", "NO_PREFETCH")
             os.environ[f"{prefix}BACKWARD_PREFETCH"] = prefetch_policy.upper()
             os.environ[f"{prefix}FORWARD_PREFETCH"] = self.fsdp_config.get("forward_prefetch", "false")
-            os.environ[f"{prefix}SYNC_MODULE_STATES"] = self.fsdp_config.get("sync_module_states", "true")
+
+            sync_module_states = self.fsdp_config.get("sync_module_states", "true")
+            cpu_ram_efficient_loading = self.fsdp_config.get("cpu_ram_efficient_loading", "false")
+
+            if str(sync_module_states).lower() == "false" and str(cpu_ram_efficient_loading).lower() == "true":
+                # In this case, all the processes except the main process would have random weights leading
+                # to unexpected behaviour during training, thus throwing error here to prevent it.
+                raise ValueError('`sync_module_states` must be `"True"` if `cpu_ram_efficient_loading` is `"True"`')
+
+            os.environ[f"{prefix}SYNC_MODULE_STATES"] = sync_module_states
+            os.environ[f"{prefix}CPU_RAM_EFFICIENT_LOADING"] = cpu_ram_efficient_loading
+
             os.environ[f"{prefix}USE_ORIG_PARAMS"] = self.fsdp_config.get("use_orig_params", "true")
 
         if is_accelerate_available():

tests/fsdp/test_fsdp.py

@@ -144,6 +144,7 @@ class TrainerIntegrationFSDP(TestCasePlus, TrainerIntegrationCommon):
             "limit_all_gathers": "False",
             "use_orig_params": "True",
             "sync_module_states": "True",
+            "cpu_ram_efficient_loading": "True",
             "activation_checkpointing": "False",
             "min_num_params": 1,
         }
@@ -208,6 +209,9 @@ class TrainerIntegrationFSDP(TestCasePlus, TrainerIntegrationCommon):
             self.assertEqual(os.environ[f"{prefix}FORWARD_PREFETCH"], fsdp_config["forward_prefetch"])
             self.assertEqual(os.environ[f"{prefix}USE_ORIG_PARAMS"], fsdp_config["use_orig_params"])
             self.assertEqual(os.environ[f"{prefix}SYNC_MODULE_STATES"], fsdp_config["sync_module_states"])
+            self.assertEqual(
+                os.environ[f"{prefix}CPU_RAM_EFFICIENT_LOADING"], fsdp_config["cpu_ram_efficient_loading"]
+            )
             self.assertEqual(os.environ.get("ACCELERATE_USE_FSDP", "false"), "true")
 
     @parameterized.expand(params, name_func=_parameterized_custom_name_func)