[Safetensors] Make safetensors the default way of saving weights (#4235)

* make safetensors default

* set default save method as safetensors

* update tests

* update to support saving safetensors

* update test to account for safetensors default

* update example tests to use safetensors

* update example to support safetensors

* update unet tests for safetensors

* fix failing loader tests

* fix qc issues

* fix pipeline tests

* fix example test

---------

Co-authored-by: Dhruv Nair <[email protected]>

examples/custom_diffusion/train_custom_diffusion.py

@@ -26,6 +26,7 @@
 from pathlib import Path
 
 import numpy as np
+import safetensors
 import torch
 import torch.nn.functional as F
 import torch.utils.checkpoint
@@ -296,14 +297,19 @@ def __getitem__(self, index):
         return example
 
 
-def save_new_embed(text_encoder, modifier_token_id, accelerator, args, output_dir):
+def save_new_embed(text_encoder, modifier_token_id, accelerator, args, output_dir, safe_serialization=True):
     """Saves the new token embeddings from the text encoder."""
     logger.info("Saving embeddings")
     learned_embeds = accelerator.unwrap_model(text_encoder).get_input_embeddings().weight
     for x, y in zip(modifier_token_id, args.modifier_token):
         learned_embeds_dict = {}
         learned_embeds_dict[y] = learned_embeds[x]
-        torch.save(learned_embeds_dict, f"{output_dir}/{y}.bin")
+        filename = f"{output_dir}/{y}.bin"
+
+        if safe_serialization:
+            safetensors.torch.save_file(learned_embeds_dict, filename, metadata={"format": "pt"})
+        else:
+            torch.save(learned_embeds_dict, filename)
 
 
 def parse_args(input_args=None):
@@ -605,6 +611,11 @@ def parse_args(input_args=None):
         action="store_true",
         help="Dont apply augmentation during data augmentation when this flag is enabled.",
     )
+    parser.add_argument(
+        "--no_safe_serialization",
+        action="store_true",
+        help="If specified save the checkpoint not in `safetensors` format, but in original PyTorch format instead.",
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1244,8 +1255,15 @@ def main(args):
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = unet.to(torch.float32)
-        unet.save_attn_procs(args.output_dir)
-        save_new_embed(text_encoder, modifier_token_id, accelerator, args, args.output_dir)
+        unet.save_attn_procs(args.output_dir, safe_serialization=not args.no_safe_serialization)
+        save_new_embed(
+            text_encoder,
+            modifier_token_id,
+            accelerator,
+            args,
+            args.output_dir,
+            safe_serialization=not args.no_safe_serialization,
+        )
 
         # Final inference
         # Load previous pipeline
@@ -1256,9 +1274,15 @@ def main(args):
         pipeline = pipeline.to(accelerator.device)
 
         # load attention processors
-        pipeline.unet.load_attn_procs(args.output_dir, weight_name="pytorch_custom_diffusion_weights.bin")
+        weight_name = (
+            "pytorch_custom_diffusion_weights.safetensors"
+            if not args.no_safe_serialization
+            else "pytorch_custom_diffusion_weights.bin"
+        )
+        pipeline.unet.load_attn_procs(args.output_dir, weight_name=weight_name)
         for token in args.modifier_token:
-            pipeline.load_textual_inversion(args.output_dir, weight_name=f"{token}.bin")
+            token_weight_name = f"{token}.safetensors" if not args.no_safe_serialization else f"{token}.bin"
+            pipeline.load_textual_inversion(args.output_dir, weight_name=token_weight_name)
 
         # run inference
         if args.validation_prompt and args.num_validation_images > 0:

examples/dreambooth/train_dreambooth_lora.py

@@ -1374,7 +1374,7 @@ def compute_text_embeddings(prompt):
         pipeline = pipeline.to(accelerator.device)
 
         # load attention processors
-        pipeline.load_lora_weights(args.output_dir, weight_name="pytorch_lora_weights.bin")
+        pipeline.load_lora_weights(args.output_dir, weight_name="pytorch_lora_weights.safetensors")
 
         # run inference
         images = []

examples/test_examples.py

@@ -23,7 +23,7 @@
 import unittest
 from typing import List
 
-import torch
+import safetensors
 from accelerate.utils import write_basic_config
 
 from diffusers import DiffusionPipeline, UNet2DConditionModel
@@ -93,7 +93,7 @@ def test_train_unconditional(self):
 
             run_command(self._launch_args + test_args, return_stdout=True)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
 
     def test_textual_inversion(self):
@@ -144,7 +144,7 @@ def test_dreambooth(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
 
     def test_dreambooth_if(self):
@@ -170,7 +170,7 @@ def test_dreambooth_if(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
 
     def test_dreambooth_checkpointing(self):
@@ -272,10 +272,10 @@ def test_dreambooth_lora(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 
@@ -305,10 +305,10 @@ def test_dreambooth_lora_with_text_encoder(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # check `text_encoder` is present at all.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             keys = lora_state_dict.keys()
             is_text_encoder_present = any(k.startswith("text_encoder") for k in keys)
             self.assertTrue(is_text_encoder_present)
@@ -341,10 +341,10 @@ def test_dreambooth_lora_if_model(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 
@@ -373,10 +373,10 @@ def test_dreambooth_lora_sdxl(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 
@@ -406,10 +406,10 @@ def test_dreambooth_lora_sdxl_with_text_encoder(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 
@@ -437,6 +437,7 @@ def test_custom_diffusion(self):
                 --lr_scheduler constant
                 --lr_warmup_steps 0
                 --modifier_token <new1>
+                --no_safe_serialization
                 --output_dir {tmpdir}
                 """.split()
 
@@ -466,7 +467,7 @@ def test_text_to_image(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
 
     def test_text_to_image_checkpointing(self):
@@ -778,7 +779,7 @@ def test_text_to_image_sdxl(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
 
     def test_text_to_image_lora_checkpointing_checkpoints_total_limit(self):
@@ -1373,7 +1374,7 @@ def test_controlnet_sdxl(self):
 
             run_command(self._launch_args + test_args)
 
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "diffusion_pytorch_model.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "diffusion_pytorch_model.safetensors")))
 
     def test_custom_diffusion_checkpointing_checkpoints_total_limit(self):
         with tempfile.TemporaryDirectory() as tmpdir:
@@ -1390,6 +1391,7 @@ def test_custom_diffusion_checkpointing_checkpoints_total_limit(self):
             --max_train_steps=6
             --checkpoints_total_limit=2
             --checkpointing_steps=2
+            --no_safe_serialization
             """.split()
 
             run_command(self._launch_args + test_args)
@@ -1413,6 +1415,7 @@ def test_custom_diffusion_checkpointing_checkpoints_total_limit_removes_multiple
             --dataloader_num_workers=0
             --max_train_steps=9
             --checkpointing_steps=2
+            --no_safe_serialization
             """.split()
 
             run_command(self._launch_args + test_args)
@@ -1436,6 +1439,7 @@ def test_custom_diffusion_checkpointing_checkpoints_total_limit_removes_multiple
             --checkpointing_steps=2
             --resume_from_checkpoint=checkpoint-8
             --checkpoints_total_limit=3
+            --no_safe_serialization
             """.split()
 
             run_command(self._launch_args + resume_run_args)
@@ -1464,10 +1468,10 @@ def test_text_to_image_lora_sdxl(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 
@@ -1491,10 +1495,10 @@ def test_text_to_image_lora_sdxl_with_text_encoder(self):
 
             run_command(self._launch_args + test_args)
             # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
 
             # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin"))
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
             is_lora = all("lora" in k for k in lora_state_dict.keys())
             self.assertTrue(is_lora)
 

examples/textual_inversion/textual_inversion.py

@@ -24,6 +24,7 @@
 
 import numpy as np
 import PIL
+import safetensors
 import torch
 import torch.nn.functional as F
 import torch.utils.checkpoint
@@ -157,15 +158,19 @@ def log_validation(text_encoder, tokenizer, unet, vae, args, accelerator, weight
     return images
 
 
-def save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path):
+def save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path, safe_serialization=True):
     logger.info("Saving embeddings")
     learned_embeds = (
         accelerator.unwrap_model(text_encoder)
         .get_input_embeddings()
         .weight[min(placeholder_token_ids) : max(placeholder_token_ids) + 1]
     )
     learned_embeds_dict = {args.placeholder_token: learned_embeds.detach().cpu()}
-    torch.save(learned_embeds_dict, save_path)
+
+    if safe_serialization:
+        safetensors.torch.save_file(learned_embeds_dict, save_path, metadata={"format": "pt"})
+    else:
+        torch.save(learned_embeds_dict, save_path)
 
 
 def parse_args():
@@ -409,6 +414,11 @@ def parse_args():
     parser.add_argument(
         "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
     )
+    parser.add_argument(
+        "--no_safe_serialization",
+        action="store_true",
+        help="If specified save the checkpoint not in `safetensors` format, but in original PyTorch format instead.",
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -878,7 +888,14 @@ def main():
                 global_step += 1
                 if global_step % args.save_steps == 0:
                     save_path = os.path.join(args.output_dir, f"learned_embeds-steps-{global_step}.bin")
-                    save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path)
+                    save_progress(
+                        text_encoder,
+                        placeholder_token_ids,
+                        accelerator,
+                        args,
+                        save_path,
+                        safe_serialization=not args.no_safe_serialization,
+                    )
 
                 if accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
@@ -936,7 +953,14 @@ def main():
             pipeline.save_pretrained(args.output_dir)
         # Save the newly trained embeddings
         save_path = os.path.join(args.output_dir, "learned_embeds.bin")
-        save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path)
+        save_progress(
+            text_encoder,
+            placeholder_token_ids,
+            accelerator,
+            args,
+            save_path,
+            safe_serialization=not args.no_safe_serialization,
+        )
 
         if args.push_to_hub:
             save_model_card(

src/diffusers/loaders.py

@@ -497,7 +497,8 @@ def save_attn_procs(
         is_main_process: bool = True,
         weight_name: str = None,
         save_function: Callable = None,
-        safe_serialization: bool = False,
+        safe_serialization: bool = True,
+        **kwargs,
     ):
         r"""
         Save an attention processor to a directory so that it can be reloaded using the
@@ -514,7 +515,8 @@ def save_attn_procs(
                 The function to use to save the state dictionary. Useful during distributed training when you need to
                 replace `torch.save` with another method. Can be configured with the environment variable
                 `DIFFUSERS_SAVE_MODE`.
-
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
         """
         from .models.attention_processor import (
             CustomDiffusionAttnProcessor,
@@ -1414,7 +1416,7 @@ def save_lora_weights(
         is_main_process: bool = True,
         weight_name: str = None,
         save_function: Callable = None,
-        safe_serialization: bool = False,
+        safe_serialization: bool = True,
     ):
         r"""
         Save the LoRA parameters corresponding to the UNet and text encoder.
@@ -1435,6 +1437,8 @@ def save_lora_weights(
                 The function to use to save the state dictionary. Useful during distributed training when you need to
                 replace `torch.save` with another method. Can be configured with the environment variable
                 `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
         """
         # Create a flat dictionary.
         state_dict = {}