Skip to content

Experimental gpu sdxl#157

Merged
entrpn merged 7 commits intomainfrom
experimental_gpu_sdxl
Mar 17, 2025
Merged

Experimental gpu sdxl#157
entrpn merged 7 commits intomainfrom
experimental_gpu_sdxl

Conversation

@entrpn
Copy link
Copy Markdown
Collaborator

@entrpn entrpn commented Mar 13, 2025

No description provided.

@entrpn entrpn requested a review from wang2yn84 March 13, 2025 21:52
Copy link
Copy Markdown
Contributor

@wang2yn84 wang2yn84 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you very much for the fix. Lgtm with some minor questions.

Comment thread README.md
- [Inference](#inference)
- [Flux](#flux)
- [Flash Attention for GPU:](#flash-attention-for-gpu)
- [Fused Attention for GPU:](#fused-attention-for-gpu)
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is there any reason why we name it as fused attention instead of flash attention considering you are using cudnn_flash_te?

Comment thread README.md

```bash
NVTE_FUSED_ATTN=1 python -m src.maxdiffusion.train_sdxl src/maxdiffusion/configs/base_xl.yml hardware=gpu run_name='test-sdxl-train' output_dir=/tmp/ train_new_unet=true train_text_encoder=false cache_latents_text_encoder_outputs=true max_train_steps=200 weights_dtype=bfloat16 resolution=512 per_device_batch_size=1 attention="cudnn_flash_te" jit_initializers=False
```
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider jit initializer works, maybe remove this config?

@wang2yn84
Copy link
Copy Markdown
Contributor

Since we are only using TF and Pytorch on CPU only, shall we install the tf[cpu] and pytorch[cpu] instead?

@entrpn entrpn merged commit 8b77101 into main Mar 17, 2025
@entrpn entrpn deleted the experimental_gpu_sdxl branch March 17, 2025 20:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants