Merge pull request #1054 from akx/mps

Device support improvements (MPS)

fine_tune.py

@@ -2,7 +2,6 @@
 # XXX dropped option: hypernetwork training
 
 import argparse
-import gc
 import math
 import os
 from multiprocessing import Value
@@ -11,6 +10,7 @@
 from tqdm import tqdm
 import torch
 
+from library.device_utils import clean_memory
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -158,9 +158,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         with torch.no_grad():
             train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
         vae.to("cpu")
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        gc.collect()
+        clean_memory()
 
         accelerator.wait_for_everyone()
 

finetune/make_captions.py

@@ -15,8 +15,9 @@
 sys.path.append(os.path.dirname(__file__))
 from blip.blip import blip_decoder, is_url
 import library.train_util as train_util
+from library.device_utils import get_preferred_device
 
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+DEVICE = get_preferred_device()
 
 
 IMAGE_SIZE = 384

finetune/make_captions_by_git.py

@@ -10,9 +10,9 @@
 from transformers.generation.utils import GenerationMixin
 
 import library.train_util as train_util
+from library.device_utils import get_preferred_device
 
-
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+DEVICE = get_preferred_device()
 
 PATTERN_REPLACE = [
     re.compile(r'(has|with|and) the (words?|letters?|name) (" ?[^"]*"|\w+)( ?(is )?(on|in) (the |her |their |him )?\w+)?'),

finetune/prepare_buckets_latents.py

@@ -14,7 +14,9 @@
 import library.model_util as model_util
 import library.train_util as train_util
 
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+from library.device_utils import get_preferred_device
+
+DEVICE = get_preferred_device()
 
 IMAGE_TRANSFORMS = transforms.Compose(
     [

gen_img_diffusers.py

@@ -66,6 +66,7 @@
 import numpy as np
 import torch
 
+from library.device_utils import clean_memory, get_preferred_device
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -888,8 +889,7 @@ def __call__(
                     init_latent_dist = self.vae.encode(init_image).latent_dist
                     init_latents = init_latent_dist.sample(generator=generator)
                 else:
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                    clean_memory()
                     init_latents = []
                     for i in tqdm(range(0, min(batch_size, len(init_image)), vae_batch_size)):
                         init_latent_dist = self.vae.encode(
@@ -1047,8 +1047,7 @@ def __call__(
         if vae_batch_size >= batch_size:
             image = self.vae.decode(latents).sample
         else:
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
+            clean_memory()
             images = []
             for i in tqdm(range(0, batch_size, vae_batch_size)):
                 images.append(
@@ -2325,7 +2324,7 @@ def __getattr__(self, item):
         scheduler.config.clip_sample = True
 
     # deviceを決定する
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # "mps"を考量してない
+    device = get_preferred_device()
 
     # custom pipelineをコピったやつを生成する
     if args.vae_slices:

library/device_utils.py

@@ -0,0 +1,34 @@
+import functools
+import gc
+
+import torch
+
+try:
+    HAS_CUDA = torch.cuda.is_available()
+except Exception:
+    HAS_CUDA = False
+
+try:
+    HAS_MPS = torch.backends.mps.is_available()
+except Exception:
+    HAS_MPS = False
+
+
+def clean_memory():
+    gc.collect()
+    if HAS_CUDA:
+        torch.cuda.empty_cache()
+    if HAS_MPS:
+        torch.mps.empty_cache()
+
+
+@functools.lru_cache(maxsize=None)
+def get_preferred_device() -> torch.device:
+    if HAS_CUDA:
+        device = torch.device("cuda")
+    elif HAS_MPS:
+        device = torch.device("mps")
+    else:
+        device = torch.device("cpu")
+    print(f"get_preferred_device() -> {device}")
+    return device

library/sdxl_train_util.py

@@ -1,5 +1,4 @@
 import argparse
-import gc
 import math
 import os
 from typing import Optional
@@ -8,6 +7,7 @@
 from tqdm import tqdm
 from transformers import CLIPTokenizer
 from library import model_util, sdxl_model_util, train_util, sdxl_original_unet
+from library.device_utils import clean_memory
 from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
 
 TOKENIZER1_PATH = "openai/clip-vit-large-patch14"
@@ -47,8 +47,7 @@ def load_target_model(args, accelerator, model_version: str, weight_dtype):
                 unet.to(accelerator.device)
                 vae.to(accelerator.device)
 
-            gc.collect()
-            torch.cuda.empty_cache()
+            clean_memory()
         accelerator.wait_for_everyone()
 
     return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info

library/train_util.py

@@ -20,7 +20,6 @@
     Union,
 )
 from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs
-import gc
 import glob
 import math
 import os
@@ -67,6 +66,7 @@
 
 # from library.attention_processors import FlashAttnProcessor
 # from library.hypernetwork import replace_attentions_for_hypernetwork
+from library.device_utils import clean_memory
 from library.original_unet import UNet2DConditionModel
 
 # Tokenizer: checkpointから読み込むのではなくあらかじめ提供されているものを使う
@@ -2279,8 +2279,7 @@ def cache_batch_latents(
                 info.latents_flipped = flipped_latent
 
     # FIXME this slows down caching a lot, specify this as an option
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
+    clean_memory()
 
 
 def cache_batch_text_encoder_outputs(
@@ -3920,6 +3919,7 @@ def prepare_accelerator(args: argparse.Namespace):
         kwargs_handlers=kwargs_handlers,
         dynamo_backend=dynamo_backend,
     )
+    print("accelerator device:", accelerator.device)
     return accelerator
 
 
@@ -4006,8 +4006,7 @@ def load_target_model(args, weight_dtype, accelerator, unet_use_linear_projectio
                 unet.to(accelerator.device)
                 vae.to(accelerator.device)
 
-            gc.collect()
-            torch.cuda.empty_cache()
+            clean_memory()
         accelerator.wait_for_everyone()
 
     return text_encoder, vae, unet, load_stable_diffusion_format
@@ -4816,7 +4815,7 @@ def sample_images_common(
 
     # clear pipeline and cache to reduce vram usage
     del pipeline
-    torch.cuda.empty_cache()
+    clean_memory()
 
     torch.set_rng_state(rng_state)
     if cuda_rng_state is not None:

networks/lora_diffusers.py

@@ -11,6 +11,8 @@
 from transformers import CLIPTextModel
 import torch
 
+from library.device_utils import get_preferred_device
+
 
 def make_unet_conversion_map() -> Dict[str, str]:
     unet_conversion_map_layer = []
@@ -476,7 +478,7 @@ def load_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True):
     from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline
     import torch
 
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    device = get_preferred_device()
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--model_id", type=str, default=None, help="model id for huggingface")

networks/lora_interrogator.py

@@ -9,11 +9,12 @@
 
 import library.model_util as model_util
 import lora
+from library.device_utils import get_preferred_device
 
 TOKENIZER_PATH = "openai/clip-vit-large-patch14"
 V2_STABLE_DIFFUSION_PATH = "stabilityai/stable-diffusion-2"     # ここからtokenizerだけ使う
 
-DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+DEVICE = get_preferred_device()
 
 
 def interrogate(args):

sdxl_gen_img.py

@@ -18,6 +18,7 @@
 import numpy as np
 import torch
 
+from library.device_utils import clean_memory, get_preferred_device
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -640,8 +641,7 @@ def __call__(
                     init_latent_dist = self.vae.encode(init_image.to(self.vae.dtype)).latent_dist
                     init_latents = init_latent_dist.sample(generator=generator)
                 else:
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                    clean_memory()
                     init_latents = []
                     for i in tqdm(range(0, min(batch_size, len(init_image)), vae_batch_size)):
                         init_latent_dist = self.vae.encode(
@@ -780,8 +780,7 @@ def __call__(
         if vae_batch_size >= batch_size:
             image = self.vae.decode(latents.to(self.vae.dtype)).sample
         else:
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
+            clean_memory()
             images = []
             for i in tqdm(range(0, batch_size, vae_batch_size)):
                 images.append(
@@ -796,8 +795,7 @@ def __call__(
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
         image = image.cpu().permute(0, 2, 3, 1).float().numpy()
 
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        clean_memory()
 
         if output_type == "pil":
             # image = self.numpy_to_pil(image)
@@ -1497,7 +1495,7 @@ def __getattr__(self, item):
     #     scheduler.config.clip_sample = True
 
     # deviceを決定する
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # "mps"を考量してない
+    device = get_preferred_device()
 
     # custom pipelineをコピったやつを生成する
     if args.vae_slices:

sdxl_minimal_inference.py

@@ -10,6 +10,7 @@
 import numpy as np
 import torch
 
+from library.device_utils import get_preferred_device
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -85,7 +86,7 @@ def get_timestep_embedding(x, outdim):
     guidance_scale = 7
     seed = None  # 1
 
-    DEVICE = "cuda"
+    DEVICE = get_preferred_device()
     DTYPE = torch.float16  # bfloat16 may work
 
     parser = argparse.ArgumentParser()

sdxl_train.py

@@ -1,7 +1,6 @@
 # training with captions
 
 import argparse
-import gc
 import math
 import os
 from multiprocessing import Value
@@ -11,6 +10,7 @@
 from tqdm import tqdm
 import torch
 
+from library.device_utils import clean_memory
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -252,9 +252,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         with torch.no_grad():
             train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
         vae.to("cpu")
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        gc.collect()
+        clean_memory()
 
         accelerator.wait_for_everyone()
 
@@ -407,8 +405,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         # move Text Encoders for sampling images. Text Encoder doesn't work on CPU with fp16
         text_encoder1.to("cpu", dtype=torch.float32)
         text_encoder2.to("cpu", dtype=torch.float32)
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        clean_memory()
     else:
         # make sure Text Encoders are on GPU
         text_encoder1.to(accelerator.device)

sdxl_train_control_net_lllite.py

@@ -2,7 +2,6 @@
 # training code for ControlNet-LLLite with passing cond_image to U-Net's forward
 
 import argparse
-import gc
 import json
 import math
 import os
@@ -15,6 +14,7 @@
 from tqdm import tqdm
 import torch
 
+from library.device_utils import clean_memory
 from library.ipex_interop import init_ipex
 
 init_ipex()
@@ -164,9 +164,7 @@ def train(args):
                 accelerator.is_main_process,
             )
         vae.to("cpu")
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        gc.collect()
+        clean_memory()
 
         accelerator.wait_for_everyone()
 
@@ -291,8 +289,7 @@ def train(args):
         # move Text Encoders for sampling images. Text Encoder doesn't work on CPU with fp16
         text_encoder1.to("cpu", dtype=torch.float32)
         text_encoder2.to("cpu", dtype=torch.float32)
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        clean_memory()
     else:
         # make sure Text Encoders are on GPU
         text_encoder1.to(accelerator.device)