[flux dreambooth lora training] make LoRA target modules configurable + small bug fix#9646
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
linoytsaban
changed the title
| f"Running validation... \n Generating {args.num_validation_images} images with prompt:" | ||
| f" {args.validation_prompt}." | ||
| ) | ||
| pipeline = pipeline.to(accelerator.device, dtype=torch_dtype) |
There was a problem hiding this comment.
Let's provide the author courtesy here.
| target_modules = [ | ||
| "attn.to_k", | ||
| "attn.to_q", | ||
| "attn.to_v", | ||
| "attn.to_out.0", | ||
| "attn.add_k_proj", | ||
| "attn.add_q_proj", | ||
| "attn.add_v_proj", | ||
| "attn.to_add_out", | ||
| "ff.net.0.proj", | ||
| "ff.net.2", | ||
| "ff_context.net.0.proj", | ||
| "ff_context.net.2", | ||
| ] |
There was a problem hiding this comment.
Seems like a bit breaking no? Better to not do it and instead make a note from the README?
WDYT?
There was a problem hiding this comment.
Breaking or just changing default behavior? I think it's geared more towards the latter, but I think it's in line with the other trainers & makes sense for Transformer based models, so maybe a Warning note and a guide on how to train it the old way for e.g.?
There was a problem hiding this comment.
Yeah maybe a warning note at the beginning of the README should cut it.
With this change, we're likely also increasing the total training wall-clock time in the default setting, so, that is worth noting.
| --cache_latents | ||
| --learning_rate 5.0e-04 | ||
| --scale_lr | ||
| --lora_layers single_transformer_blocks.0.attn.to_k |
There was a problem hiding this comment.
(nit): could make "single_transformer_blocks.0.attn.to_k" a constant and then supply it here. This way we know what we're testing for immediately. WDYT?
… + small bug fix (#9646) * make lora target modules configurable and change the default * style * make lora target modules configurable and change the default * fix bug when using prodigy and training te * fix mixed precision training as proposed in #9565 for full dreambooth as well * add test and notes * style * address sayaks comments * style * fix test --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
4 participants
new feature for the Flux dreambooth lora training script:
make LoRA target modules configurable through
--lora_layerschange the current default target modules to not be attention layers only
& small fix to mixed precision training for dreambooth script, as proposed in #9565