diff comment

src/transformers/trainer.py

@@ -436,8 +436,13 @@ class Trainer:
         # Will reach this branch if the user has
         # 1. Used `.from_pretrained` or `.from_config` to initialize their model
         # 2. Did not configure Zero-3 via `TrainingArguments` or `accelerate launch` beforehand
+        # 3. Also not using quantization
         # New models init such as `MyModel()` will not hit this step
-        if is_deepspeed_zero3_enabled() and not getattr(model, "_transformers_zero3_init_used", True):
+        if (
+            not (hasattr(model, "hf_quantizer") and model.hf_quantizer.is_trainable)
+            and is_deepspeed_zero3_enabled()
+            and not getattr(model, "_transformers_zero3_init_used", True)
+        ):
             raise ValueError(
                 "Model was not initialized with `Zero-3` despite being configured for DeepSpeed Zero-3. Please re-initialize your model via `Model.from_pretrained(...)` or `Model.from_config(...)` after creating your `TrainingArguments`!"
             )

tests/deepspeed/test_deepspeed.py

@@ -42,9 +42,11 @@ from transformers.testing_utils import (
     backend_device_count,
     execute_subprocess_async,
     mockenv_context,
+    require_bitsandbytes,
     require_deepspeed,
     require_optuna,
     require_torch_accelerator,
+    require_torch_gpu,
     require_torch_multi_accelerator,
     slow,
     torch_device,
@@ -734,6 +736,23 @@ class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, T
             assert trainer.is_deepspeed_enabled
             assert model._transformers_zero3_init_used
 
+    @require_torch_gpu
+    @require_bitsandbytes
+    def test_missed_zero3_init_quantized(self):
+        from transformers import Trainer  # noqa
+
+        with mockenv_context(**self.dist_env_1_gpu):
+            model = AutoModel.from_pretrained(T5_TINY, load_in_4bit=True)
+            training_args = TrainingArguments(
+                output_dir="./test_missed_zero3_init",
+                deepspeed=self.get_config_dict(ZERO3),
+            )
+            # Shouldn't raise an error in this case
+            _ = Trainer(
+                model=model,
+                args=training_args,
+            )
+
     def check_saved_checkpoints_deepspeed(self, output_dir, freq, total, stage, dtype):
         # adapted from TrainerIntegrationCommon.check_saved_checkpoints
         file_list = [SAFE_WEIGHTS_NAME, "training_args.bin", "trainer_state.json", "config.json"]