Add offload option in flux-control training (#10225)

* Add offload option in flux-control training

* Update examples/flux-control/train_control_flux.py
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* modify help message

* fix format

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

examples/flux-control/README.md

@@ -36,6 +36,7 @@ accelerate launch train_control_lora_flux.py \
   --max_train_steps=5000 \
   --validation_image="openpose.png" \
   --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
   --seed="0" \
   --push_to_hub
 ```
@@ -154,6 +155,7 @@ accelerate launch --config_file=accelerate_ds2.yaml train_control_flux.py \
   --validation_steps=200 \
   --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
   --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
   --seed="0" \
   --push_to_hub
 ```

examples/flux-control/train_control_flux.py

@@ -541,6 +541,11 @@ def parse_args(input_args=None):
         default=1.29,
         help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
     )
+    parser.add_argument(
+        "--offload",
+        action="store_true",
+        help="Whether to offload the VAE and the text encoders to CPU when they are not used.",
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -999,6 +1004,7 @@ def main(args):
                 control_latents = encode_images(
                     batch["conditioning_pixel_values"], vae.to(accelerator.device), weight_dtype
                 )
+                if args.offload:
                     # offload vae to CPU.
                     vae.cpu()
 
@@ -1064,6 +1070,7 @@ def main(args):
                 if args.proportion_empty_prompts and random.random() < args.proportion_empty_prompts:
                     prompt_embeds.zero_()
                     pooled_prompt_embeds.zero_()
+                if args.offload:
                     text_encoding_pipeline = text_encoding_pipeline.to("cpu")
 
                 # Predict.

examples/flux-control/train_control_lora_flux.py

@@ -573,6 +573,11 @@ def parse_args(input_args=None):
         default=1.29,
         help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
     )
+    parser.add_argument(
+        "--offload",
+        action="store_true",
+        help="Whether to offload the VAE and the text encoders to CPU when they are not used.",
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1140,6 +1145,8 @@ def main(args):
                 control_latents = encode_images(
                     batch["conditioning_pixel_values"], vae.to(accelerator.device), weight_dtype
                 )
+
+                if args.offload:
                     # offload vae to CPU.
                     vae.cpu()
 
@@ -1205,6 +1212,7 @@ def main(args):
                 if args.proportion_empty_prompts and random.random() < args.proportion_empty_prompts:
                     prompt_embeds.zero_()
                     pooled_prompt_embeds.zero_()
+                if args.offload:
                     text_encoding_pipeline = text_encoding_pipeline.to("cpu")
 
                 # Predict.