25
FIBO
β¨
ZeroGPU AoTI
community
AI & ML interests
AoT compilation, ZeroGPU inference optimization
Recent Activity
View all activity
linoytsΒ
updated
a
Space
5 days ago
multimodalartΒ
posted
an
update
18 days ago
Post
2249
Want to iterate on a Hugging Face Space with an LLM?
Now you can easily convert any HF entire repo (Model, Dataset or Space) to a text file and feed it to a language model!
multimodalart/repo2txt
Now you can easily convert any HF entire repo (Model, Dataset or Space) to a text file and feed it to a language model!
multimodalart/repo2txt
sayakpaulΒ
authored
a
paper
27 days ago
Post
1645
Fast LoRA inference for Flux with Diffusers and PEFT π¨
There are great materials that demonstrate how to optimize inference for popular image generation models, such as Flux. However, very few cover how to serve LoRAs fast, despite LoRAs being an inseparable part of their adoption.
In our latest post, @BenjaminB and I show different techniques to optimize LoRA inference for the Flux family of models for image generation. Our recipe includes the use of:
1.
2. Flash Attention 3 (when compatible)
3. Dynamic FP8 weight quantization (when compatible)
4. Hotswapping for avoiding recompilation during swapping new LoRAs π€―
We have tested our recipe with Flux.1-Dev on both H100 and RTX 4090. We achieve at least a *2x speedup* in either of the GPUs. We believe our recipe is grounded in the reality of how LoRA-based use cases are generally served. So, we hope this will be beneficial to the community π€
Even though our recipe was tested primarily with NVIDIA GPUs, it should also work with AMD GPUs.
Learn the details and the full code here:
https://huggingface.co/blog/lora-fast
There are great materials that demonstrate how to optimize inference for popular image generation models, such as Flux. However, very few cover how to serve LoRAs fast, despite LoRAs being an inseparable part of their adoption.
In our latest post, @BenjaminB and I show different techniques to optimize LoRA inference for the Flux family of models for image generation. Our recipe includes the use of:
1.
torch.compile2. Flash Attention 3 (when compatible)
3. Dynamic FP8 weight quantization (when compatible)
4. Hotswapping for avoiding recompilation during swapping new LoRAs π€―
We have tested our recipe with Flux.1-Dev on both H100 and RTX 4090. We achieve at least a *2x speedup* in either of the GPUs. We believe our recipe is grounded in the reality of how LoRA-based use cases are generally served. So, we hope this will be beneficial to the community π€
Even though our recipe was tested primarily with NVIDIA GPUs, it should also work with AMD GPUs.
Learn the details and the full code here:
https://huggingface.co/blog/lora-fast
multimodalartΒ
posted
an
update
5 months ago
Post
17666
Self-Forcing - a real-time video distilled model from Wan 2.1 by
@adobe
is out, and they open sourced it π
I've built a live real time demo on Spaces πΉπ¨
multimodalart/self-forcing
I've built a live real time demo on Spaces πΉπ¨
multimodalart/self-forcing
-
- 6 replies
Post
3930
π ZeroGPU now supports PyTorch native quantization via
While it hasnβt been battle-tested yet,
Let us know if you run into any issues β and weβre excited to see what the community will build!
torchaoWhile it hasnβt been battle-tested yet,
Int8WeightOnlyConfig is already working flawlessly in our tests.Let us know if you run into any issues β and weβre excited to see what the community will build!
import spaces
from diffusers import FluxPipeline
from torchao.quantization.quant_api import Int8WeightOnlyConfig, quantize_
pipeline = FluxPipeline.from_pretrained(...).to('cuda')
quantize_(pipeline.transformer, Int8WeightOnlyConfig()) # Or any other component(s)
@spaces.GPU
def generate(prompt: str):
return pipeline(prompt).images[0]
- 5 replies
Post
2870
Diffusers supports a good variety of quantization backends. It can be challenging to navigate through them, given the complex nature of diffusion pipelines in general.
So, @derekl35 set out to write a comprehensive guide that puts users in the front seat. Explore the different backends we support, learn the trade-offs they offer, and finally, check out the cool space we built that lets you compare quantization results.
Give it a go here:
https://lnkd.in/gf8Pi4-2
So, @derekl35 set out to write a comprehensive guide that puts users in the front seat. Explore the different backends we support, learn the trade-offs they offer, and finally, check out the cool space we built that lets you compare quantization results.
Give it a go here:
https://lnkd.in/gf8Pi4-2
-
- 2 replies
Post
1816
Despite the emergence of combining LLM and DiT architectures for T2I synthesis, its design remains severely understudied.
This was done long ago and got into CVPR25 -- super excited to finally share it now, along with the data and code β₯οΈ
We explore several architectural choices that affect this design. We provide an open & reproducible training recipe that works at scale.
Works like Playground v3 have already explored a deep fusion between an LLM and a DiT, sharing their representations through layerwise attention. They exhibit excellent performance on T2I.
Despite its compelling results and other performance virtues, it remains unexplored, which is what we want to improve in our work. Specifically, we take a pre-trained LLM (Gemma-2B) and trainable DiT, and set out to explore what makes a "good deep fusion" between the two for T2I.
We explore several key questions in the work, such as:
Q1: How should we do attention? We considered several alternatives. PixArt-Alpha like attention (cross-attention) is very promising.
Q2: Should we incorporate additional text modulation?
Q3: Can we eliminate timestep conditioning?
Q4: How do we do positional encodings?
Q5: Do instruction-tuned LLMs help deep fusion?
Q6: Would using a decoder LLM from a multimodal model be helpful?
Q7: Does using a better variant of Gemma help?
Based on the above findings, we arrive at FuseDiT with the following components on top of the base architecture from the findings of our experiments.
* No AdaLN-Zero modules
* 1D + 2D-RoPE
* Gemma 2 2B, adjusting DiT configurations accordingly
We trained FuseDiT on a mixture from CC12M, JourneyDB, & SA (~26M image-text pairs) for 800 steps. While not the best model, it's encouraging to develop something in a guided manner using open datasets.
To know more (code, models, all are available), please check out the paper:
https://lnkd.in/gg6qyqZX.
This was done long ago and got into CVPR25 -- super excited to finally share it now, along with the data and code β₯οΈ
We explore several architectural choices that affect this design. We provide an open & reproducible training recipe that works at scale.
Works like Playground v3 have already explored a deep fusion between an LLM and a DiT, sharing their representations through layerwise attention. They exhibit excellent performance on T2I.
Despite its compelling results and other performance virtues, it remains unexplored, which is what we want to improve in our work. Specifically, we take a pre-trained LLM (Gemma-2B) and trainable DiT, and set out to explore what makes a "good deep fusion" between the two for T2I.
We explore several key questions in the work, such as:
Q1: How should we do attention? We considered several alternatives. PixArt-Alpha like attention (cross-attention) is very promising.
Q2: Should we incorporate additional text modulation?
Q3: Can we eliminate timestep conditioning?
Q4: How do we do positional encodings?
Q5: Do instruction-tuned LLMs help deep fusion?
Q6: Would using a decoder LLM from a multimodal model be helpful?
Q7: Does using a better variant of Gemma help?
Based on the above findings, we arrive at FuseDiT with the following components on top of the base architecture from the findings of our experiments.
* No AdaLN-Zero modules
* 1D + 2D-RoPE
* Gemma 2 2B, adjusting DiT configurations accordingly
We trained FuseDiT on a mixture from CC12M, JourneyDB, & SA (~26M image-text pairs) for 800 steps. While not the best model, it's encouraging to develop something in a guided manner using open datasets.
To know more (code, models, all are available), please check out the paper:
https://lnkd.in/gg6qyqZX.
Post
6058
π ZeroGPU
Nothing too fancy for nowβZeroGPU Spaces still default to
- π° size-based quotas / pricing (
- 𦣠the upcoming
You can as of now control GPU size via a Space variable. Accepted values:
-
-
-
The auto mode checks total CUDA tensor size during startup:
- More than 30GB β
- Otherwise β
medium size is now available as a power-user featureNothing too fancy for nowβZeroGPU Spaces still default to
large (70GB VRAM)βbut this paves the way for:- π° size-based quotas / pricing (
medium will offer significantly more usage than large)- 𦣠the upcoming
xlarge size (141GB VRAM)You can as of now control GPU size via a Space variable. Accepted values:
-
auto (future default)-
medium-
large (current default)The auto mode checks total CUDA tensor size during startup:
- More than 30GB β
large- Otherwise β
medium
-
- 3 replies
sayakpaulΒ
authored
a
paper
6 months ago
Post
6059
FramePack is hands down one of the best OS releases in video generation ππ»ββοΈπ€―
β fully open sourced + amazing quality + reduced memory + improved speed
but more even - its gonna facilitate *soooo* many downstream applications
like this version adapted for landscape rotation πhttps://huggingface.co/spaces/tori29umai/FramePack_rotate_landscape
β fully open sourced + amazing quality + reduced memory + improved speed
but more even - its gonna facilitate *soooo* many downstream applications
like this version adapted for landscape rotation πhttps://huggingface.co/spaces/tori29umai/FramePack_rotate_landscape
- 2 replies
sayakpaulΒ
authored
a
paper
6 months ago
Post
3179
We just shipped HiDream Image LoRA fine-tuning to diffusersπ§¨
HiDream's sota capabilities (and mit license) bring a lot of potential to explore with fine-tunes π₯
- more upgrades and features soon!
- code, weights and config example π
π§ΆYarn art lora: linoyts/HiDream-yarn-art-LoRA
code: https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_hidream.md
HiDream's sota capabilities (and mit license) bring a lot of potential to explore with fine-tunes π₯
- more upgrades and features soon!
- code, weights and config example π
π§ΆYarn art lora: linoyts/HiDream-yarn-art-LoRA
code: https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_hidream.md
- 2 replies
sayakpaulΒ
authored
a
paper
8 months ago
Post
3941
Inference-time scaling meets Flux.1-Dev (and others) π₯
Presenting a simple re-implementation of "Inference-time scaling diffusion models beyond denoising steps" by Ma et al.
I did the simplest random search strategy, but results can potentially be improved with better-guided search methods.
Supports Gemini 2 Flash & Qwen2.5 as verifiers for "LLMGrading" π€
The steps are simple:
For each round:
1> Starting by sampling 2 starting noises with different seeds.
2> Score the generations w.r.t a metric.
3> Obtain the best generation from the current round.
If you have more compute budget, go to the next search round. Scale the noise pool (
This constitutes the random search method as done in the paper by Google DeepMind.
Code, more results, and a bunch of other stuff are in the repository. Check it out here: https://github.com/sayakpaul/tt-scale-flux/ π€
Presenting a simple re-implementation of "Inference-time scaling diffusion models beyond denoising steps" by Ma et al.
I did the simplest random search strategy, but results can potentially be improved with better-guided search methods.
Supports Gemini 2 Flash & Qwen2.5 as verifiers for "LLMGrading" π€
The steps are simple:
For each round:
1> Starting by sampling 2 starting noises with different seeds.
2> Score the generations w.r.t a metric.
3> Obtain the best generation from the current round.
If you have more compute budget, go to the next search round. Scale the noise pool (
2 ** search_round) and repeat 1 - 3.This constitutes the random search method as done in the paper by Google DeepMind.
Code, more results, and a bunch of other stuff are in the repository. Check it out here: https://github.com/sayakpaul/tt-scale-flux/ π€
Post
2152
We have been cooking a couple of fine-tuning runs on CogVideoX with finetrainers, smol datasets, and LoRA to generate cool video effects like crushing, dissolving, etc.
We are also releasing a LoRA extraction utility from a fully fine-tuned checkpoint. I know that kind of stuff has existed since eternity, but the quality on video models was nothing short of spectacular. Below are some links:
* Models and datasets:
finetrainers
* finetrainers: https://github.com/a-r-r-o-w/finetrainers
* LoRA extraction: https://github.com/huggingface/diffusers/blob/main/scripts/extract_lora_from_model.py
We are also releasing a LoRA extraction utility from a fully fine-tuned checkpoint. I know that kind of stuff has existed since eternity, but the quality on video models was nothing short of spectacular. Below are some links:
* Models and datasets:
finetrainers
* finetrainers: https://github.com/a-r-r-o-w/finetrainers
* LoRA extraction: https://github.com/huggingface/diffusers/blob/main/scripts/extract_lora_from_model.py
- 1 reply
Post
2084
We have authored a post to go over the state of video generation in the Diffusers ecosystem π§¨
We cover the models supported, the knobs of optims our users can fire, fine-tuning, and more π₯
5-6GBs for HunyuanVideo, sky is the limit π π€
https://huggingface.co/blog/video_gen
We cover the models supported, the knobs of optims our users can fire, fine-tuning, and more π₯
5-6GBs for HunyuanVideo, sky is the limit π π€
https://huggingface.co/blog/video_gen
Post
4495
Commits speak louder than words π€ͺ
* 4 new video models
* Multiple image models, including SANA & Flux Control
* New quantizers -> GGUF & TorchAO
* New training scripts
Enjoy this holiday-special Diffusers release π€
Notes: https://github.com/huggingface/diffusers/releases/tag/v0.32.0
* 4 new video models
* Multiple image models, including SANA & Flux Control
* New quantizers -> GGUF & TorchAO
* New training scripts
Enjoy this holiday-special Diffusers release π€
Notes: https://github.com/huggingface/diffusers/releases/tag/v0.32.0
Post
2301
In the past seven days, the Diffusers team has shipped:
1. Two new video models
2. One new image model
3. Two new quantization backends
4. Three new fine-tuning scripts
5. Multiple fixes and library QoL improvements
Coffee on me if someone can guess 1 - 4 correctly.
1. Two new video models
2. One new image model
3. Two new quantization backends
4. Three new fine-tuning scripts
5. Multiple fixes and library QoL improvements
Coffee on me if someone can guess 1 - 4 correctly.
- 1 reply
Post
2220
Introducing a high-quality open-preference dataset to further this line of research for image generation.
Despite being such an inseparable component for modern image generation, open preference datasets are a rarity!
So, we decided to work on one with the community!
Check it out here:
https://huggingface.co/blog/image-preferences
Despite being such an inseparable component for modern image generation, open preference datasets are a rarity!
So, we decided to work on one with the community!
Check it out here:
https://huggingface.co/blog/image-preferences
- 8 replies