Open to Work

Tolga Cangöz

tolgacangoz

standard_ai

AI & ML interests

AIGC

Recent Activity

liked a model about 5 hours ago

facebook/sam3.1

upvoted a collection 3 days ago

Modular Pipelines

liked a model 5 days ago

netflix/void-model

View all activity

Organizations

reactedto nouamanetazi's post with 🧠 5 months ago

Post

4672

After training 𝐒𝐦𝐨𝐥𝐋𝐌𝟑 on 𝟑𝟖𝟒 𝐇𝟏𝟎𝟎𝐬 for nearly a month, I've come to realize something most people overlook: 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐦𝐚𝐤𝐞-𝐨𝐫-𝐛𝐫𝐞𝐚𝐤 𝐟𝐚𝐜𝐭𝐨𝐫 𝐢𝐧 𝐋𝐋𝐌 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠. 🔥

Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️

Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?

That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.

We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.

If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.

𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS

Shared with ❤️ by the HuggingFace team

reactedto sayakpaul's post with 🚀 6 months ago

Post

3001

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

2 replies

repliedto sayakpaul's post 6 months ago

Thanks for the blog post! nunchaku when?

reactedto sayakpaul's post with 🔥 6 months ago

Post

2666

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. torch.compile
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

3 replies

reactedto jwgu's post with 🔥 6 months ago

Post

2525

🎉 NEW RELEASES: Cosmos Predict 2.5 and Transfer 2.5

Cosmos Predict 2.5:
- Combines Text2World, Image2World, and Video2World
- Multimodal, future-state video prediction

Cosmos Transfer 2.5:
- High-fidelity multicontrol world simulations
- Inputs: RGB, depth, segmentation—blended seamlessly

These updates boost development of autonomous vehicles, robotics, and video analytics.

Don’t miss Jensen Huang’s keynote at NVIDIA GTC Washington, D.C. on 10/28 to hear the latest in physical AI.

📺 Watch live: https://nvda.ws/4pUjF4x

🔗 Try Predict 2.5:
https://nvda.ws/4otReZZ

🔗 Try Transfer 2.5:
https://nvda.ws/46GEx7T

reactedto multimodalart's post with ❤️ 7 months ago

Post

18298

Self-Forcing - a real-time video distilled model from Wan 2.1 by @adobe is out, and they open sourced it 🐐

I've built a live real time demo on Spaces 📹💨

multimodalart/self-forcing

6 replies

reactedto a-r-r-o-w's post with 🧠 8 months ago

Post

2522

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 ✨)

3 replies

reactedto linoyts's post with 🔥 9 months ago

Post

19684

FramePack is hands down one of the best OS releases in video generation 🙇🏻‍♀️🤯
✅ 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 👇https://huggingface.co/spaces/tori29umai/FramePack_rotate_landscape

4 replies

reactedto a-r-r-o-w's post with 🧠 9 months ago

Post

3516

Caching is an essential technique used in diffusion inference serving for speeding up image/video generations. Diffusers just added support for another caching method: First Block Cache - a technique developed by @chengzeyi building upon the ideas of TeaCache.

The idea in short is: if the model predictions do not vary much over successive inference steps, we can skip certain steps where the prediction difference is small. To figure out whether an inference step will make a significant improvement to the overall velocity/noise prediction, we calculate the relative difference of the output of the first transformer block at timestep $t$ with $t-1$, and compare it against a selected threshold. If the difference is lower than the threshold, we skip the step. A higher threshold will lead to more steps being skipped. However, skipping many steps is bad because it can throw off the model predictions, and so we need to test and select the threshold based on level of quality-speed tradeoff for every model we use it with.

Diffusers usage with CogView4:

import torch
from diffusers import CogView4Pipeline
from diffusers.hooks import apply_first_block_cache, FirstBlockCacheConfig

pipe = CogView4Pipeline.from_pretrained("THUDM/CogView4-6B", torch_dtype=torch.bfloat16)
pipe.to("cuda")

apply_first_block_cache(pipe.transformer, FirstBlockCacheConfig(threshold=0.2))

prompt = "A photo of an astronaut riding a horse on mars"
image = pipe(prompt, generator=torch.Generator().manual_seed(42)).images[0]
image.save("output.png")

Below, you'll find the benchmarks and visualizations of the predicted output at different blocks of the Flux DiT.

Docs: https://huggingface.co/docs/diffusers/main/en/optimization/cache
PR: https://github.com/huggingface/diffusers/pull/11180

References:
- First Block Cache: https://github.com/chengzeyi/ParaAttention
- TeaCache: https://github.com/ali-vilab/TeaCache

1 reply

reactedto a-r-r-o-w's post with 🚀 10 months ago

Post

2869

As you might have already heard, FLUX.1-Kontext-dev is now released and taken the generative community by storm!

In case you haven't come across it, you can get started with Kontext using 🤗 diffusers. See the official [model]( black-forest-labs/FLUX.1-Kontext-dev) and [docs](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux#flux).

Want to know how inference companies like Fal & Replicate are able to run the model so fast and in under 2 seconds per image? Check out this [gist](https://gist.github.com/a-r-r-o-w/d08c37e8bd3e9c26b4ce80360be148c6) for some details!

1 reply

reactedto a-r-r-o-w's post with 👀 10 months ago

Post

1348

Did you know how simple it was to get started with your own custom compiler backend with torch.compile? What's stopping you from writing your own compiler?

import torch
from torch._functorch.partitioners import draw_graph

def compiler(fx_module: torch.fx.GraphModule, _):
    draw_graph(fx_module, f"compile.dot")
    return fx_module.forward

def capture(model, *inputs):
    compiled_model = torch.compile(model, backend=compiler)
    y = compiled_model(*inputs)
    y.sum().backward()

class MLP(torch.nn.Module):
    def __init__(self):
        super().__init__()
        
        self.linear_1 = torch.nn.Linear(16, 32)
        self.linear_2 = torch.nn.Linear(32, 16)
    
    def forward(self, x):
        x = self.linear_1(x)
        x = torch.nn.functional.silu(x)
        x = self.linear_2(x)
        return x

if __name__ == '__main__':
    model = MLP()
    model.to("mps")
    x = torch.randn(4, 16, device="mps", dtype=torch.float32)

    capture(model, x)

--------------

Part of https://huggingface.co/posts/a-r-r-o-w/231008365980283

1 reply

repliedto a-r-r-o-w's post 10 months ago

I sincerely support you. Writing seems like a completely different way of thinking. It is not something passive -it generates ideas! (Writing, Briefly)
Looking forward to your posts.

reactedto a-r-r-o-w's post with 🔥 10 months ago

Post

2319

New diffusion model for text-to-image and video-to-world generation: Cosmos Predict-2 👽

Model collection: nvidia/cosmos-predict2-68028efc052239369a0f2959
Diffusers support: https://github.com/huggingface/diffusers/pull/11695
Documentation: https://huggingface.co/docs/diffusers/main/en/api/pipelines/cosmos

These are results with the 2B param model. Imagine what you could do with the 14B version! Go check it out now!

1 reply

reactedto a-r-r-o-w's post with 🧠 10 months ago

Post

1348

Did you know how simple it was to get started with your own custom compiler backend with torch.compile? What's stopping you from writing your own compiler?

import torch
from torch._functorch.partitioners import draw_graph

def compiler(fx_module: torch.fx.GraphModule, _):
    draw_graph(fx_module, f"compile.dot")
    return fx_module.forward

def capture(model, *inputs):
    compiled_model = torch.compile(model, backend=compiler)
    y = compiled_model(*inputs)
    y.sum().backward()

class MLP(torch.nn.Module):
    def __init__(self):
        super().__init__()
        
        self.linear_1 = torch.nn.Linear(16, 32)
        self.linear_2 = torch.nn.Linear(32, 16)
    
    def forward(self, x):
        x = self.linear_1(x)
        x = torch.nn.functional.silu(x)
        x = self.linear_2(x)
        return x

if __name__ == '__main__':
    model = MLP()
    model.to("mps")
    x = torch.randn(4, 16, device="mps", dtype=torch.float32)

    capture(model, x)

--------------

Part of https://huggingface.co/posts/a-r-r-o-w/231008365980283

1 reply

reactedto a-r-r-o-w's post with 🔥 10 months ago

Post

2272

Recently, I've been focusing my learning on the following topics:
- Pytorch internals, specifically the inductor system (roughly ~1 month of experience)
- Triton internals (~8 moe)
- CUDA (~3 moe)
- Understanding fusion patterns in compilers and how to improve them (~1 moe)
- Parallelism strategies for large scale inference optimization (~6-7 moe)

I thought it would be nice to document it somewhere for no particular reason. Maybe someone will find it useful? It's also because I want to get into the habit of writing, but had no motivation to do so. Maybe writing short informal posts will help build the habit.

Since I don't have a personal site, and don't plan to create one in the near future, I think HF posts are best suited for short and informal documentation to share my little discoveries and learnings. If you're interested, strap in!

First post in this series will be on basic study of Pytorch's float32 matmuls and their Triton implementation (nothing much, just the tutorial available on the website), short dive into TF32 and their TFLOPS comparison on an A100 machine.

3 replies

reactedto Kseniase's post with ❤️ about 1 year ago

Post

5843

8 types of RoPE

As we always use Transformers, it's helpful to understand RoPE—Rotary Position Embedding. Since token order matters, RoPE encodes it by rotating token embeddings based on their position, so the model knows how to interpret which token comes first, second, and so on.

Here are 8 types of RoPE that can be implemented in different cases:

1. Original RoPE -> RoFormer: Enhanced Transformer with Rotary Position Embedding (2104.09864)
Encodes token positions by rotating token embeddings in the complex plane via a position-based rotation matrix, thereby providing the self-attention mechanism with relative positional info.

2. LongRoPE -> LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens (2402.13753)
Extends the context window of pre-trained LLMs to 2048k tokens, leveraging non-uniformities in positional interpolation with an efficient search.

3. LongRoPE2 -> LongRoPE2: Near-Lossless LLM Context Window Scaling (2502.20082)
Extends the effective context window of pre-trained LLMs to the target! length, rescaling RoPE guided by “needle-driven” perplexity.

4. Multimodal RoPE (MRoPE) -> Qwen2.5-VL Technical Report (2502.13923)
Decomposes positional embedding into 3 components: temporal, height and width, so that positional features are aligned across modalities: text, images and videos.

5. Directional RoPE (DRoPE) -> DRoPE: Directional Rotary Position Embedding for Efficient Agent Interaction Modeling (2503.15029)
Adds an identity scalar, improving how angles are handled without extra complexity. It helps balance accuracy, speed, and memory usage.

6. VideoRoPE -> VideoRoPE: What Makes for Good Video Rotary Position Embedding? (2502.05173)
Adapts RoPE for video, featuring 3D structure, low-frequency temporal allocation, diagonal layout, and adjustable spacing.

7. VRoPE -> VRoPE: Rotary Position Embedding for Video Large Language Models (2502.11664)
An another RoPE for video, which restructures positional indices and balances encoding for uniform spatial focus.

8. XPos (Extrapolatable Position Embedding) -> https://huggingface.co/papers/2212.10
Introduces an exponential decay factor into the rotation matrix, improving stability on long sequences.

1 reply

reactedto tomaarsen's post with ❤️ about 1 year ago

Post

6941

An assembly of 18 European companies, labs, and universities have banded together to launch 🇪🇺 EuroBERT! It's a state-of-the-art multilingual encoder for 15 European languages, designed to be finetuned for retrieval, classification, etc.

🇪🇺 15 Languages: English, French, German, Spanish, Chinese, Italian, Russian, Polish, Portuguese, Japanese, Vietnamese, Dutch, Arabic, Turkish, Hindi
3️⃣ 3 model sizes: 210M, 610M, and 2.1B parameters - very very useful sizes in my opinion
➡️ Sequence length of 8192 tokens! Nice to see these higher sequence lengths for encoders becoming more common.
⚙️ Architecture based on Llama, but with bi-directional (non-causal) attention to turn it into an encoder. Flash Attention 2 is supported.
🔥 A new Pareto frontier (stronger *and* smaller) for multilingual encoder models
📊 Evaluated against mDeBERTa, mGTE, XLM-RoBERTa for Retrieval, Classification, and Regression (after finetuning for each task separately): EuroBERT punches way above its weight.
📝 Detailed paper with all details, incl. data: FineWeb for English and CulturaX for multilingual data, The Stack v2 and Proof-Pile-2 for code.

Check out the release blogpost here: https://huggingface.co/blog/EuroBERT/release
* EuroBERT/EuroBERT-210m
* EuroBERT/EuroBERT-610m
* EuroBERT/EuroBERT-2.1B

The next step is for researchers to build upon the 3 EuroBERT base models and publish strong retrieval, zero-shot classification, etc. models for all to use. I'm very much looking forward to it!

1 reply

reactedto kadirnar's post with 🔥 about 1 year ago

Post

4505

I created my own AI image and video from scratch using the fal.ai platform 💫

Workflow: Flux Lora Training + Upscale + Kling AI(1.6)

6 replies

reactedto reach-vb's post with 🔥 over 1 year ago

Post

3548

NEW: Open Source Text/ Image to video model is out - MIT licensed - Rivals Gen-3, Pika & Kling 🔥

> Pyramid Flow: Training-efficient Autoregressive Video Generation method
> Utilizes Flow Matching
> Trains on open-source datasets
> Generates high-quality 10-second videos
> Video resolution: 768p
> Frame rate: 24 FPS
> Supports image-to-video generation

> Model checkpoints available on the hub 🤗: rain1011/pyramid-flow-sd3

reactedto kadirnar's post with 🔥 over 1 year ago

Post

4630

I am training a controlnet model for Flux. And some of my experiences:

Checkpoint-10000:

https://x.com/kadirnar_ai/status/1829831750471606668

Checkpoint-12000:

https://x.com/kadirnar_ai/status/1829889524962640001

Checkpoint-14000:

https://x.com/kadirnar_ai/status/1829989622878744711

Checkpoint (16000-18000):

https://x.com/kadirnar_ai/status/1830179551407665654

Dataset: kadirnar/fluxdev_controlnet_16k
GPU: 1xA100(80GB)
GPU Hours: 65

1 reply

Tolga Cangöz

AI & ML interests

Recent Activity

Organizations

tolgacangoz's activity