Skip to content

[advanced flux training] bug fix + reduce memory cost as in #9829 #9838

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Nov 19, 2024
Merged

[advanced flux training] bug fix + reduce memory cost as in #9829 #9838

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
dda1cb8
memory improvement as done here: https://github.com/huggingface/diffu…
linoytsaban Nov 1, 2024
6a9c5ed
fix bug
linoytsaban Nov 1, 2024
5aa1db6
fix bug
linoytsaban Nov 1, 2024
7feb480
style
linoytsaban Nov 1, 2024
026f9da
Merge branch 'main' into flux-training-mem
linoytsaban Nov 3, 2024
ef54dd2
Merge branch 'main' into flux-training-mem
linoytsaban Nov 18, 2024
97689a3
Merge branch 'main' into flux-training-mem
sayakpaul Nov 19, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 12 additions & 2 deletions examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2154,6 +2154,7 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):

# encode batch prompts when custom prompts are provided for each image -
if train_dataset.custom_instance_prompts:
elems_to_repeat = 1
if freeze_text_encoder:
prompt_embeds, pooled_prompt_embeds, text_ids = compute_text_embeddings(
prompts, text_encoders, tokenizers
Expand All @@ -2168,17 +2169,21 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
max_sequence_length=args.max_sequence_length,
add_special_tokens=add_special_tokens_t5,
)
else:
elems_to_repeat = len(prompts)

if not freeze_text_encoder:
prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
text_encoders=[text_encoder_one, text_encoder_two],
tokenizers=[None, None],
text_input_ids_list=[tokens_one, tokens_two],
text_input_ids_list=[
tokens_one.repeat(elems_to_repeat, 1),
tokens_two.repeat(elems_to_repeat, 1),
],
max_sequence_length=args.max_sequence_length,
device=accelerator.device,
prompt=prompts,
)

# Convert images to latent space
if args.cache_latents:
model_input = latents_cache[step].sample()
Expand Down Expand Up @@ -2371,6 +2376,9 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
epoch=epoch,
torch_dtype=weight_dtype,
)
images = None
del pipeline

if freeze_text_encoder:
del text_encoder_one, text_encoder_two
free_memory()
Expand Down Expand Up @@ -2448,6 +2456,8 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
images = None
del pipeline

accelerator.end_training()

Expand Down
6 changes: 5 additions & 1 deletion examples/dreambooth/train_dreambooth_lora_flux.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1648,11 +1648,15 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
prompt=prompts,
)
else:
elems_to_repeat = len(prompts)
if args.train_text_encoder:
prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
text_encoders=[text_encoder_one, text_encoder_two],
tokenizers=[None, None],
text_input_ids_list=[tokens_one, tokens_two],
text_input_ids_list=[
tokens_one.repeat(elems_to_repeat, 1),
tokens_two.repeat(elems_to_repeat, 1),
],
max_sequence_length=args.max_sequence_length,
device=accelerator.device,
prompt=args.instance_prompt,
Expand Down
Loading