7529 do not disable autocast for cuda devices (huggingface#7530)

* 7529 do not disable autocast for cuda devices

* Remove typecasting error check for non-mps platforms, as a correct autocast implementation makes it a non-issue

* add autocast fix to other training examples

* disable native_amp for dreambooth (sdxl)

* disable native_amp for pix2pix (sdxl)

* remove tests from remaining files

* disable native_amp on huggingface accelerator for every training example that uses it

* convert more usages of autocast to nullcontext, make style fixes

* make style fixes

* style.

* Empty-Commit

---------

Co-authored-by: bghira <bghira@users.github.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -23,6 +23,7 @@
 import re
 import shutil
 import warnings
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Optional
 
@@ -1844,7 +1845,12 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
 
-                with torch.cuda.amp.autocast():
+            if torch.backends.mps.is_available():
+                autocast_ctx = nullcontext()
+            else:
+                autocast_ctx = torch.autocast(accelerator.device.type)
+
+                with autocast_ctx:
                     images = [
                         pipeline(**pipeline_args, generator=generator).images[0]
                         for _ in range(args.num_validation_images)

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
 import gc
 import hashlib
 import itertools
@@ -26,6 +25,7 @@
 import re
 import shutil
 import warnings
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Optional
 
@@ -2192,13 +2192,12 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 # run inference
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
-                inference_ctx = (
-                    contextlib.nullcontext()
-                    if "playground" in args.pretrained_model_name_or_path
-                    else torch.cuda.amp.autocast()
-                )
+                if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
+                    autocast_ctx = nullcontext()
+                else:
+                    autocast_ctx = torch.autocast(accelerator.device.type)
 
-                with inference_ctx:
+                with autocast_ctx:
                     images = [
                         pipeline(**pipeline_args, generator=generator).images[0]
                         for _ in range(args.num_validation_images)

examples/amused/train_amused.py

@@ -430,6 +430,9 @@ def main(args):
         log_with=args.report_to,
         project_config=accelerator_project_config,
     )
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
 
     if accelerator.is_main_process:
         os.makedirs(args.output_dir, exist_ok=True)

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -23,6 +23,7 @@
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Union
 
@@ -238,6 +239,10 @@ def train_dataloader(self):
 
 def log_validation(vae, unet, args, accelerator, weight_dtype, step):
     logger.info("Running validation... ")
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
 
     unet = accelerator.unwrap_model(unet)
     pipeline = StableDiffusionPipeline.from_pretrained(
@@ -274,7 +279,7 @@ def log_validation(vae, unet, args, accelerator, weight_dtype, step):
 
     for _, prompt in enumerate(validation_prompts):
         images = []
-        with torch.autocast("cuda", dtype=weight_dtype):
+        with autocast_ctx:
             images = pipeline(
                 prompt=prompt,
                 num_inference_steps=4,
@@ -1172,6 +1177,11 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok
     ).input_ids.to(accelerator.device)
     uncond_prompt_embeds = text_encoder(uncond_input_ids)[0]
 
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
+
     # 16. Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
@@ -1300,7 +1310,7 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok
                 # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
                 # solver timestep.
                 with torch.no_grad():
-                    with torch.autocast("cuda"):
+                    with autocast_ctx:
                         # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
@@ -1359,7 +1369,7 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok
                 # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 # Note that we do not use a separate target network for LCM-LoRA distillation.
                 with torch.no_grad():
-                    with torch.autocast("cuda", dtype=weight_dtype):
+                    with autocast_ctx:
                         target_noise_pred = unet(
                             x_prev.float(),
                             timesteps,

examples/consistency_distillation/train_lcm_distill_lora_sdxl.py

@@ -22,6 +22,7 @@
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 
 import accelerate
@@ -146,7 +147,12 @@ def log_validation(vae, args, accelerator, weight_dtype, step, unet=None, is_fin
 
     for _, prompt in enumerate(validation_prompts):
         images = []
-        with torch.autocast("cuda", dtype=weight_dtype):
+        if torch.backends.mps.is_available():
+            autocast_ctx = nullcontext()
+        else:
+            autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
+
+        with autocast_ctx:
             images = pipeline(
                 prompt=prompt,
                 num_inference_steps=4,

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -24,6 +24,7 @@
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Union
 
@@ -256,6 +257,10 @@ def train_dataloader(self):
 
 def log_validation(vae, unet, args, accelerator, weight_dtype, step):
     logger.info("Running validation... ")
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
 
     unet = accelerator.unwrap_model(unet)
     pipeline = StableDiffusionXLPipeline.from_pretrained(
@@ -291,7 +296,7 @@ def log_validation(vae, unet, args, accelerator, weight_dtype, step):
 
     for _, prompt in enumerate(validation_prompts):
         images = []
-        with torch.autocast("cuda", dtype=weight_dtype):
+        with autocast_ctx:
             images = pipeline(
                 prompt=prompt,
                 num_inference_steps=4,
@@ -1353,7 +1358,12 @@ def compute_embeddings(
                 # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
                 # solver timestep.
                 with torch.no_grad():
-                    with torch.autocast("cuda"):
+                    if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type)
+
+                    with autocast_ctx:
                         # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
@@ -1416,7 +1426,12 @@ def compute_embeddings(
                 # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 # Note that we do not use a separate target network for LCM-LoRA distillation.
                 with torch.no_grad():
-                    with torch.autocast("cuda", enabled=True, dtype=weight_dtype):
+                    if torch.backends.mps.is_available():
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
+
+                    with autocast_ctx:
                         target_noise_pred = unet(
                             x_prev.float(),
                             timesteps,

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -23,6 +23,7 @@
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Union
 
@@ -252,7 +253,12 @@ def log_validation(vae, unet, args, accelerator, weight_dtype, step, name="targe
 
     for _, prompt in enumerate(validation_prompts):
         images = []
-        with torch.autocast("cuda"):
+        if torch.backends.mps.is_available():
+            autocast_ctx = nullcontext()
+        else:
+            autocast_ctx = torch.autocast(accelerator.device.type)
+
+        with autocast_ctx:
             images = pipeline(
                 prompt=prompt,
                 num_inference_steps=4,
@@ -1257,7 +1263,12 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok
                 # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
                 # solver timestep.
                 with torch.no_grad():
-                    with torch.autocast("cuda"):
+                    if torch.backends.mps.is_available():
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type)
+
+                    with autocast_ctx:
                         # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
@@ -1315,7 +1326,12 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok
 
                 # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 with torch.no_grad():
-                    with torch.autocast("cuda", dtype=weight_dtype):
+                    if torch.backends.mps.is_available():
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
+
+                    with autocast_ctx:
                         target_noise_pred = target_unet(
                             x_prev.float(),
                             timesteps,

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -24,6 +24,7 @@
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 from typing import List, Union
 
@@ -270,7 +271,12 @@ def log_validation(vae, unet, args, accelerator, weight_dtype, step, name="targe
 
     for _, prompt in enumerate(validation_prompts):
         images = []
-        with torch.autocast("cuda"):
+        if torch.backends.mps.is_available():
+            autocast_ctx = nullcontext()
+        else:
+            autocast_ctx = torch.autocast(accelerator.device.type)
+
+        with autocast_ctx:
             images = pipeline(
                 prompt=prompt,
                 num_inference_steps=4,
@@ -1355,7 +1361,12 @@ def compute_embeddings(
                 # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
                 # solver timestep.
                 with torch.no_grad():
-                    with torch.autocast("cuda"):
+                    if torch.backends.mps.is_available():
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type)
+
+                    with autocast_ctx:
                         # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
@@ -1417,7 +1428,12 @@ def compute_embeddings(
 
                 # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 with torch.no_grad():
-                    with torch.autocast("cuda", dtype=weight_dtype):
+                    if torch.backends.mps.is_available():
+                        autocast_ctx = nullcontext()
+                    else:
+                        autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)
+
+                    with autocast_ctx:
                         target_noise_pred = target_unet(
                             x_prev.float(),
                             timesteps,

examples/controlnet/train_controlnet.py

@@ -752,6 +752,10 @@ def main(args):
         project_config=accelerator_project_config,
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",

examples/controlnet/train_controlnet_sdxl.py

@@ -14,14 +14,14 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
 import functools
 import gc
 import logging
 import math
 import os
 import random
 import shutil
+from contextlib import nullcontext
 from pathlib import Path
 
 import accelerate
@@ -125,11 +125,10 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
         )
 
     image_logs = []
-    inference_ctx = (
-        contextlib.nullcontext()
-        if (is_final_validation or torch.backends.mps.is_available())
-        else torch.autocast("cuda")
-    )
+    if is_final_validation or torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
 
     for validation_prompt, validation_image in zip(validation_prompts, validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
@@ -138,7 +137,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
         images = []
 
         for _ in range(args.num_validation_images):
-            with inference_ctx:
+            with autocast_ctx:
                 image = pipeline(
                     prompt=validation_prompt, image=validation_image, num_inference_steps=20, generator=generator
                 ).images[0]
@@ -811,6 +810,10 @@ def main(args):
         project_config=accelerator_project_config,
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",

examples/custom_diffusion/train_custom_diffusion.py

@@ -676,6 +676,10 @@ def main(args):
         project_config=accelerator_project_config,
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     if args.report_to == "wandb":
         if not is_wandb_available():
             raise ImportError("Make sure to install wandb if you want to use it for logging during training.")

examples/dreambooth/train_dreambooth.py

@@ -821,6 +821,10 @@ def main(args):
         project_config=accelerator_project_config,
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     if args.report_to == "wandb":
         if not is_wandb_available():
             raise ImportError("Make sure to install wandb if you want to use it for logging during training.")

examples/dreambooth/train_dreambooth_lora.py

@@ -749,6 +749,10 @@ def main(args):
         project_config=accelerator_project_config,
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     if args.report_to == "wandb":
         if not is_wandb_available():
             raise ImportError("Make sure to install wandb if you want to use it for logging during training.")