[tests] add tests for flux modular (t2i, i2i, kontext)#12566
[tests] add tests for flux modular (t2i, i2i, kontext)#12566
Conversation
| InputParam( | ||
| "guidance", | ||
| required=True, | ||
| required=False, |
There was a problem hiding this comment.
For Flux.1-schnell, guidance shouldn't be required.
| "image_processor", | ||
| VaeImageProcessor, | ||
| config=FrozenDict({"vae_scale_factor": 16}), | ||
| config=FrozenDict({"vae_scale_factor": 16, "vae_latent_channels": 16}), |
There was a problem hiding this comment.
Just to be consistent.
| pooled_prompt_embeds = block_state.pooled_prompt_embeds.repeat(1, block_state.num_images_per_prompt) | ||
| block_state.pooled_prompt_embeds = pooled_prompt_embeds.view( | ||
| block_state.batch_size * block_state.num_images_per_prompt, -1 | ||
| ) |
There was a problem hiding this comment.
Was left out of the expansion incorrectly.
| import torch | ||
| from PIL import Image | ||
|
|
||
| from diffusers import ( |
There was a problem hiding this comment.
Unrelated changes but I thought I would do it (happy to revert).
|
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. |
Co-authored-by: YiYi Xu <yixu310@gmail.com>
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
thanks for the PR
looks good, but I don't know why the changes in components_manager got reverted
I left a commenet there, I think we should add a warning during enable_model_oaffload(), and you can also keep the warning from the memory_reserve (or just make a note there in comment so we know that to fix in the fugure)
feel free to merge otherwise
| device_module = getattr(torch, device_type, torch.cuda) | ||
| mem_on_device = device_module.mem_get_info(execution_device.index)[0] | ||
| mem_on_device = mem_on_device - self.memory_reserve_margin | ||
|
|
There was a problem hiding this comment.
ohh, what happended here?
#12566 (comment)
It was a good catch, I think we should also add a warning sooner during enable_model_offload()
There was a problem hiding this comment.
My bad for misunderstanding!
I have added the error catching. I added a TODO in enable_auto_cpu_offload() to error out early for mem_get_info().
3 participants
What does this PR do?
The tests also helped me uncover some bugs and fix them. Some comments are in line.