Article
Introducing Waypoint-1: Real-time interactive video diffusion from Overworld
- +3
ā¢
33
Abdennacer Badaoui
badaoui
AI & ML interests
None yet
Recent Activity
upvoted
an
article
4 days ago
Introducing Waypoint-1: Real-time interactive video diffusion from Overworld
upvoted
an
article
4 days ago
Nemotron-Personas-Brazil: Co-Designed Data for Sovereign AI
upvoted
an
article
4 days ago
Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective
Organizations
upvoted
3
articles
4 days ago
Article
Nemotron-Personas-Brazil: Co-Designed Data for Sovereign AI
ā¢
10
Article
Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective
ā¢
43
liked
a
dataset
4 days ago
reacted to
IlyasMoutawwakil's
post with š„
6 days ago
Post
2890
Transformers v5 just landed! š
It significantly unifies and reduces modeling code across architectures, while opening the door to a whole new class of performance optimizations.
My favorite new feature? š¤
The new dynamic weight loader + converter. Hereās why š
Over the last few months, the core Transformers maintainers built an incredibly fast weight loader, capable of converting tensors on the fly while loading them in parallel threads. This means weāre no longer constrained by how parameters are laid out inside the safetensors weight files.
In practice, this unlocks two big things:
- Much more modular modeling code. You can now clearly see how architectures build on top of each other (DeepSeek v2 ā v3, Qwen v2 ā v3 ā MoE, etc.). This makes shared bottlenecks obvious and lets us optimize the right building blocks once, for all model families.
- Performance optimizations beyond what torch.compile can do alone. torch.compile operates on the computation graph, but it canāt change parameter layouts. With the new loader, we can restructure weights at load time: fusing MoE expert projections, merging attention QKV projections, and enabling more compute-dense kernels that simply werenāt possible before.
Personally, I'm honored to have contributed in this direction, including the work on optimizing MoE implementations and making modeling code more torch-exportable, so these optimizations can be ported cleanly across runtimes.
Overall, Transformers v5 is a strong signal of where the community and industry are converging: Modularity and Performance, without sacrificing Flexibility.
Transformers v5 makes its signature from_pretrained an entrypoint where you can mix and match:
- Parallelism
- Quantization
- Custom kernels
- Flash/Paged attention
- Continuous batching
- ...
Kudos to everyone involved! I highly recommend the:
Release notes: https://github.com/huggingface/transformers/releases/tag/v5.0.0
Blog post: https://huggingface.co/blog/transformers-v5
It significantly unifies and reduces modeling code across architectures, while opening the door to a whole new class of performance optimizations.
My favorite new feature? š¤
The new dynamic weight loader + converter. Hereās why š
Over the last few months, the core Transformers maintainers built an incredibly fast weight loader, capable of converting tensors on the fly while loading them in parallel threads. This means weāre no longer constrained by how parameters are laid out inside the safetensors weight files.
In practice, this unlocks two big things:
- Much more modular modeling code. You can now clearly see how architectures build on top of each other (DeepSeek v2 ā v3, Qwen v2 ā v3 ā MoE, etc.). This makes shared bottlenecks obvious and lets us optimize the right building blocks once, for all model families.
- Performance optimizations beyond what torch.compile can do alone. torch.compile operates on the computation graph, but it canāt change parameter layouts. With the new loader, we can restructure weights at load time: fusing MoE expert projections, merging attention QKV projections, and enabling more compute-dense kernels that simply werenāt possible before.
Personally, I'm honored to have contributed in this direction, including the work on optimizing MoE implementations and making modeling code more torch-exportable, so these optimizations can be ported cleanly across runtimes.
Overall, Transformers v5 is a strong signal of where the community and industry are converging: Modularity and Performance, without sacrificing Flexibility.
Transformers v5 makes its signature from_pretrained an entrypoint where you can mix and match:
- Parallelism
- Quantization
- Custom kernels
- Flash/Paged attention
- Continuous batching
- ...
Kudos to everyone involved! I highly recommend the:
Release notes: https://github.com/huggingface/transformers/releases/tag/v5.0.0
Blog post: https://huggingface.co/blog/transformers-v5
- 3 replies
upvoted
an
article
about 1 month ago
Article
Shadow AI - Where are the CIOs?
ā¢
31
liked
3
Spaces
about 2 months ago
upvoted
an
article
2 months ago
Article
Transformers v5: Simple model definitions powering the AI ecosystem
- +2
ā¢
288
liked
a
Space
2 months ago
upvoted
4
articles
2 months ago
Article
Continuous batching from first principles
-
- +1
ā¢
315
Article
Building Deep Research: How we Achieved State of the Art
ā¢
34
Article
Google Cloud C4 Brings a 70% TCO improvement on GPT OSS with Intel and Hugging Face
- +2
ā¢
18
Article
Join the AMD Open Robotics Hackathon
ā¢
13
Fix historical data not updating after automatic reloads
#9 opened 3 months ago
by
badaoui
updated
a
Space
3 months ago
commented on
Easily Build and Share ROCm Kernels with Hugging Face
3 months ago
Yes, absolutely!
You can find the documentation for Apple Silicon (Metal) here: https://github.com/huggingface/kernel-builder/blob/main/docs/metal.md
And hereās an example of a kernel built for Metal: https://huggingface.co/kernels-community/gpt-oss-metal-kernels
posted
an
update
3 months ago
Post
490
Building high-performance, reproducible kernels forĀ AMD ROCmĀ just got a lot easier.
I've put together a guide on building, testing, and sharingĀ ROCm-compatible kernelsĀ using the Hugging FaceĀ
Learn how to:
- Use Nix for reproducible builds
- Integrate kernels as native PyTorch operators
- Share your kernels on the Hub for anyone to use withĀ
We use the š award-winning RadeonFlow GEMM kernel as a practical example.
š Check out the full guide here : https://huggingface.co/blog/build-rocm-kernels
I've put together a guide on building, testing, and sharingĀ ROCm-compatible kernelsĀ using the Hugging FaceĀ
kernel-builder and kernelsĀ libraries; so you can focus onĀ optimizing performanceĀ rather than spending time on setup.Learn how to:
- Use Nix for reproducible builds
- Integrate kernels as native PyTorch operators
- Share your kernels on the Hub for anyone to use withĀ
kernels.get_kernel()We use the š award-winning RadeonFlow GEMM kernel as a practical example.
š Check out the full guide here : https://huggingface.co/blog/build-rocm-kernels
upvoted
an
article
3 months ago
Article
Easily Build and Share ROCm Kernels with Hugging Face
-
- +2
ā¢
36