Phenomenal post!
Maybe it could help future readers to have some clarity on the CUDA programming model itself so that the hierarchy of where each of the components (SM, thread blocks, registers, etc.) sits is clear.
Sayak Paul
sayakpaul
AI & ML interests
Diffusion models, representation learning
Recent Activity
updated
a dataset
about 4 hours ago
huggingface/diffusers-metadata
liked
a model
about 9 hours ago
nunchaku-tech/nunchaku-qwen-image
updated
a model
2 days ago
diffusers-internal-dev/gemini-prompt-expander
Organizations
reacted to
a-r-r-o-w's
post with 👍
17 days ago
Post
2065
You would've implemented the 3-loop matrix multiplication many times as a ML practitioner, but the naive implementation is terrible for GPU performance. Modern GPUs achieve peak performance through careful memory access patterns and minimizing scheduling overhead.
In naive matmul (MxK . KxN), the computation happens in tiles - both for the output matrix and for how you read chunks from the input matrices. Each thread-block processes one output tile by loading corresponding tiles from input (for sum-reduction across K dimension), performing the computation, then terminating. The GPU launches many thread-blocks and schedules them across available streaming multiprocessors (SMs). When an SM finishes one tile, it gets assigned a new thread-block for the next uncomputed tile. This way, multiple output tiles are computed in parallel across the SMs, but we pay the cost for launching thread-blocks each time a new tile is computed.
Persistent matmul changes this approach. Instead of launching thread-blocks to compute some output tiles, computing the results on SMs in parallel, and repeating until all output tiles are computed, you launch only as many thread-blocks as you have SMs available (typically 80-132 on modern GPUs). These thread-blocks stay alive until all output tiles are computed, looping through multiple tiles sequentially. Each persistent thread-block may handle multiple output tiles.
The key benefit is the reduced thread-block launch latency. This persistence strategy, combined with other optimizations like coalesced memory loads/stores, block-tiling, warp-tiling, warp-specialization, double-buffering, ping-pong scheduling and other tricks, helps achieve peak performance. More on this in the future!
Code snippet for testing: https://gist.github.com/a-r-r-o-w/28339b442d164084506c0967029968a8
(Bonus: Since I've wanted to learn Manim for a while, this was a great opportunity to make a visualization for Naive VS Persistent matmul. Enjoy ✨)
In naive matmul (MxK . KxN), the computation happens in tiles - both for the output matrix and for how you read chunks from the input matrices. Each thread-block processes one output tile by loading corresponding tiles from input (for sum-reduction across K dimension), performing the computation, then terminating. The GPU launches many thread-blocks and schedules them across available streaming multiprocessors (SMs). When an SM finishes one tile, it gets assigned a new thread-block for the next uncomputed tile. This way, multiple output tiles are computed in parallel across the SMs, but we pay the cost for launching thread-blocks each time a new tile is computed.
Persistent matmul changes this approach. Instead of launching thread-blocks to compute some output tiles, computing the results on SMs in parallel, and repeating until all output tiles are computed, you launch only as many thread-blocks as you have SMs available (typically 80-132 on modern GPUs). These thread-blocks stay alive until all output tiles are computed, looping through multiple tiles sequentially. Each persistent thread-block may handle multiple output tiles.
The key benefit is the reduced thread-block launch latency. This persistence strategy, combined with other optimizations like coalesced memory loads/stores, block-tiling, warp-tiling, warp-specialization, double-buffering, ping-pong scheduling and other tricks, helps achieve peak performance. More on this in the future!
Code snippet for testing: https://gist.github.com/a-r-r-o-w/28339b442d164084506c0967029968a8
(Bonus: Since I've wanted to learn Manim for a while, this was a great opportunity to make a visualization for Naive VS Persistent matmul. Enjoy ✨)
-
- 3 replies
posted
an
update
29 days ago
Post
1032
Fast LoRA inference for Flux with Diffusers and PEFT 🚨
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 🤯
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 🤗
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 🤯
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 🤗
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
reacted to
burtenshaw's
post with ❤️
about 2 months ago
Post
2885
Inference for generative ai models looks like a mine field, but there’s a simple protocol for picking the best inference:
🌍 95% of users >> If you’re using open (large) models and need fast online inference, then use Inference providers on auto mode, and let it choose the best provider for the model. https://huggingface.co/docs/inference-providers/index
👷 fine-tuners/ bespoke >> If you’ve got custom setups, use Inference Endpoints to define a configuration from AWS, Azure, GCP. https://endpoints.huggingface.co/
🦫 Locals >> If you’re trying to stretch everything you can out of a server or local machine, use Llama.cpp, Jan, LMStudio or vLLM. https://huggingface.co/settings/local-apps#local-apps
🪟 Browsers >> If you need open models running right here in the browser, use transformers.js. https://github.com/huggingface/transformers.js
Let me know what you’re using, and if you think it’s more complex than this.
🌍 95% of users >> If you’re using open (large) models and need fast online inference, then use Inference providers on auto mode, and let it choose the best provider for the model. https://huggingface.co/docs/inference-providers/index
👷 fine-tuners/ bespoke >> If you’ve got custom setups, use Inference Endpoints to define a configuration from AWS, Azure, GCP. https://endpoints.huggingface.co/
🦫 Locals >> If you’re trying to stretch everything you can out of a server or local machine, use Llama.cpp, Jan, LMStudio or vLLM. https://huggingface.co/settings/local-apps#local-apps
🪟 Browsers >> If you need open models running right here in the browser, use transformers.js. https://github.com/huggingface/transformers.js
Let me know what you’re using, and if you think it’s more complex than this.
posted
an
update
3 months ago
Post
2790
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
posted
an
update
3 months ago
Post
1759
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 ♥️
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 ♥️
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.
reacted to
clem's
post with ❤️
4 months ago
Post
4577
Hugging Face is becoming the best place to share the most viral AI apps with spaces.
Kolors Virtual Try-on just crossed 6,000,000 unique visitors & is now the #5 most popular space. Congrats to the Kwai Kolors team!
Kwai-Kolors/Kolors-Virtual-Try-On
Kolors Virtual Try-on just crossed 6,000,000 unique visitors & is now the #5 most popular space. Congrats to the Kwai Kolors team!
Kwai-Kolors/Kolors-Virtual-Try-On
- 2 replies
posted
an
update
6 months ago
Post
3913
Inference-time scaling meets Flux.1-Dev (and others) 🔥
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" 🤗
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/ 🤗
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" 🤗
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/ 🤗
posted
an
update
7 months ago
Post
2128
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
posted
an
update
7 months ago
Post
2064
We have authored a post to go over the state of video generation in the Diffusers ecosystem 🧨
We cover the models supported, the knobs of optims our users can fire, fine-tuning, and more 🔥
5-6GBs for HunyuanVideo, sky is the limit 🌌 🤗
https://huggingface.co/blog/video_gen
We cover the models supported, the knobs of optims our users can fire, fine-tuning, and more 🔥
5-6GBs for HunyuanVideo, sky is the limit 🌌 🤗
https://huggingface.co/blog/video_gen
posted
an
update
8 months ago
Post
4473
Commits speak louder than words 🤪
* 4 new video models
* Multiple image models, including SANA & Flux Control
* New quantizers -> GGUF & TorchAO
* New training scripts
Enjoy this holiday-special Diffusers release 🤗
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 🤗
Notes: https://github.com/huggingface/diffusers/releases/tag/v0.32.0
posted
an
update
8 months ago
Post
2285
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
posted
an
update
9 months ago
Post
2206
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
posted
an
update
9 months ago
Post
2227
The Control family of Flux from
@black-forest-labs
should be discussed more!
It enables structural controls like ControlNets while being significantly less expensive to run!
So, we're working on a Control LoRA training script 🤗
It's still WIP, so go easy:
https://github.com/huggingface/diffusers/pull/10130
It enables structural controls like ControlNets while being significantly less expensive to run!
So, we're working on a Control LoRA training script 🤗
It's still WIP, so go easy:
https://github.com/huggingface/diffusers/pull/10130
posted
an
update
9 months ago
Post
1585
Let 2024 be the year of video model fine-tunes!
Check it out here:
https://github.com/a-r-r-o-w/cogvideox-factory/tree/main/training/mochi-1
Check it out here:
https://github.com/a-r-r-o-w/cogvideox-factory/tree/main/training/mochi-1
posted
an
update
9 months ago
Post
2742
It's been a while we shipped native quantization support in
We currently support
This post is just a reminder of what's possible:
1. Loading a model with a quantization config
2. Saving a model with quantization config
3. Loading a pre-quantized model
4.
5. Training and loading LoRAs into quantized checkpoints
Docs:
https://huggingface.co/docs/diffusers/main/en/quantization/bitsandbytes
diffusers 🧨We currently support
bistandbytes as the official backend but using others like torchao is already very simple. This post is just a reminder of what's possible:
1. Loading a model with a quantization config
2. Saving a model with quantization config
3. Loading a pre-quantized model
4.
enable_model_cpu_offload()5. Training and loading LoRAs into quantized checkpoints
Docs:
https://huggingface.co/docs/diffusers/main/en/quantization/bitsandbytes
- 1 reply
posted
an
update
11 months ago
Post
2820
Did some little experimentation to resize pre-trained LoRAs on Flux. I explored two themes:
* Decrease the rank of a LoRA
* Increase the rank of a LoRA
The first one is helpful in reducing memory requirements if the LoRA is of a high rank, while the second one is merely an experiment. Another implication of this study is in the unification of LoRA ranks when you would like to
Check it out here:
sayakpaul/flux-lora-resizing
* Decrease the rank of a LoRA
* Increase the rank of a LoRA
The first one is helpful in reducing memory requirements if the LoRA is of a high rank, while the second one is merely an experiment. Another implication of this study is in the unification of LoRA ranks when you would like to
torch.compile() them. Check it out here:
sayakpaul/flux-lora-resizing
- 1 reply
reacted to
dn6's
post with ❤️
12 months ago
Post
3079
Sharing for anyone using Diffusers
If you have
If you do not have the model repo saved in your cache, then automatically inferring the pipeline config will not work since the reference repo
You can use an alternative SD1.5 repo id to still configure your pipeline.
We're working on resolving the issue ASAP.
from_single_file loading and affected by the Runway SD 1.5 issue.If you have
runwayml/stable-diffusion-v1-5 saved locally in your HF cache then loading single file checkpoints in the following way should still work. from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_single_file("<url or path to single file checkpoint>")If you do not have the model repo saved in your cache, then automatically inferring the pipeline config will not work since the reference repo
runwayml/stable-diffusion-v1-5 doesn't exist anymore. You can use an alternative SD1.5 repo id to still configure your pipeline.
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_single_file("<url or path to single file checkpoint>", config="Lykon/DreamShaper")We're working on resolving the issue ASAP.
-
- 2 replies
posted
an
update
about 1 year ago
Post
2988
Here is a hackable and minimal implementation showing how to perform distributed text-to-image generation with Diffusers and Accelerate.
Full snippet is here: https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba
With @JW17
Full snippet is here: https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba
With @JW17
posted
an
update
about 1 year ago
Post
4548
Flux.1-Dev like images but in fewer steps.
Merging code (very simple), inference code, merged params: sayakpaul/FLUX.1-merged
Enjoy the Monday 🤗
Merging code (very simple), inference code, merged params: sayakpaul/FLUX.1-merged
Enjoy the Monday 🤗
- 4 replies