SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations
Paper
•
2512.14080
•
Published
•
5
LLHF
community
AI & ML interests
None defined yet.
Recent Activity
View all activity
eliebak
submitted
a
paper
to Daily Papers
17 days ago
Post
2814
Nvidia is on a roll lately. Nemotron 3 Nano is my new fav local model, but here's the real flex: they published the entire evaluation setup. Configs, prompts, logs, all of it. This is how you do open models 🔥
https://huggingface.co/blog/nvidia/nemotron-3-nano-evaluation-recipe
https://huggingface.co/blog/nvidia/nemotron-3-nano-evaluation-recipe
hfkwr
authored
a
paper
3 months ago
Post
3876
Super excited to announce that our research team at Hugging Face will be doing an AMA on reddit r/LocalLLaMA.
Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe there’ll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST 🤗
science
Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe there’ll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST 🤗
Post
712
Motif 2.6B tech report is pretty insane, first time i see a model with differential attention and polynorm trained at scale!
> It's trained on 2.5T of token, with a "data mixture schedule" to continuously adjust the mixture over training.
> They use WSD with a "Simple moving average" averaging the last 6 ckpt every 8B token.
> They trained on Finemath, Fineweb2, DCLM, TxT360.
> Lot of details in the finetuning data they used, for instance they used EvolKit and did some "dataset fusion" to have more compressed knowledge into the data.
> They mention they also tried Normalized GPT, QK-Norm and Cross Layer Attention.
Motif-Technologies/Motif-2.6B
> It's trained on 2.5T of token, with a "data mixture schedule" to continuously adjust the mixture over training.
> They use WSD with a "Simple moving average" averaging the last 6 ckpt every 8B token.
> They trained on Finemath, Fineweb2, DCLM, TxT360.
> Lot of details in the finetuning data they used, for instance they used EvolKit and did some "dataset fusion" to have more compressed knowledge into the data.
> They mention they also tried Normalized GPT, QK-Norm and Cross Layer Attention.
Motif-Technologies/Motif-2.6B
Post
3296
Say hello to
We are glad to announce a long-awaited quality-of-life improvement: the Hugging Face CLI has been officially renamed from huggingface-cli to hf!
So... why this change?
Typing huggingface-cli constantly gets old fast. More importantly, the CLI’s command structure became messy as new features were added over time (upload, download, cache management, repo management, etc.). Renaming the CLI is a chance to reorganize commands into a clearer, more consistent format.
We decided not to reinvent the wheel and instead follow a well-known CLI pattern: hf <resource> <action>. Isn't
The full rationale, implementation details, and migration notes are in the blog post: https://huggingface.co/blog/hf-cli
hf: a faster, friendlier Hugging Face CLI ✨We are glad to announce a long-awaited quality-of-life improvement: the Hugging Face CLI has been officially renamed from huggingface-cli to hf!
So... why this change?
Typing huggingface-cli constantly gets old fast. More importantly, the CLI’s command structure became messy as new features were added over time (upload, download, cache management, repo management, etc.). Renaming the CLI is a chance to reorganize commands into a clearer, more consistent format.
We decided not to reinvent the wheel and instead follow a well-known CLI pattern: hf <resource> <action>. Isn't
hf auth login easier to type and remember?The full rationale, implementation details, and migration notes are in the blog post: https://huggingface.co/blog/hf-cli
Post
4793
Kimi K2 tech report is full of gems as always. Here are my notes on it:
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
dvilasuero
posted
an
update
7 months ago
Post
3263
Super excited to launch Hugging Face Sheets: Spreadsheets meet AI and unstructured data.
A few months ago, we started imagining new ways to build and transform datasets with the latest open-source models.
Today, I'm thrilled to introduce our first step in this direction.
In a nutshell:
📁 Effortlessly run prompts and models over your data.
🌐 Agentic search for accuracy and real-time information.
🖼️ Familiar, minimalistic interface for interacting with data.
🎯 Human feedback 2.0: Your input directly improves generated data.
💯 Access hundreds of open models and leading inference providers.
Go to this space to try it out!
aisheets/sheets
Leave your questions below, we're just getting started!
A few months ago, we started imagining new ways to build and transform datasets with the latest open-source models.
Today, I'm thrilled to introduce our first step in this direction.
In a nutshell:
📁 Effortlessly run prompts and models over your data.
🌐 Agentic search for accuracy and real-time information.
🖼️ Familiar, minimalistic interface for interacting with data.
🎯 Human feedback 2.0: Your input directly improves generated data.
💯 Access hundreds of open models and leading inference providers.
Go to this space to try it out!
aisheets/sheets
Leave your questions below, we're just getting started!
-
- 3 replies
Post
7589
Open Source Avengers, Assemble! Ask an expert AI agent team to solve complex problems together 🔥
Consilium brings together multiple agents that debate and use live research (web, arXiv, SEC) to reach a consensus. You set the strategy, they find the answer.
Credit to @azettl for this awesome demo: Agents-MCP-Hackathon/consilium_mcp
Consilium brings together multiple agents that debate and use live research (web, arXiv, SEC) to reach a consensus. You set the strategy, they find the answer.
Credit to @azettl for this awesome demo: Agents-MCP-Hackathon/consilium_mcp
- 2 replies
eliebak
authored
a
paper
7 months ago
Post
756
Let's go! Custom generation code has landed in
Have you designed a new cool KV cache? Maybe you're comparing new test-time compute ideas you've been researching? Have you found a way to do diffusion with existing models? You can now easily share your findings with the community with custom generation code, sharing the well-known
In a nutshell, we have expanded the support of custom modeling code on the Hub with *model-agnostic* custom generation code. Write for one model, reuse with any model -- hopefully, this will democratize access to new generation ideas 🫡
As a creator, you gain the ability to get your ideas in
💎 Resources 💎
- docs: https://huggingface.co/docs/transformers/generation_strategies#custom-decoding-methods
- minimal example: transformers-community/custom_generate_example
- discussion: transformers-community/support#10
transformers 🚀Have you designed a new cool KV cache? Maybe you're comparing new test-time compute ideas you've been researching? Have you found a way to do diffusion with existing models? You can now easily share your findings with the community with custom generation code, sharing the well-known
generate interface 🤓In a nutshell, we have expanded the support of custom modeling code on the Hub with *model-agnostic* custom generation code. Write for one model, reuse with any model -- hopefully, this will democratize access to new generation ideas 🫡
As a creator, you gain the ability to get your ideas in
transformers with minimal effort. You'll also have access to all Hub features: a landing page for your creation, discussions, usage metrics, ... 🤓💎 Resources 💎
- docs: https://huggingface.co/docs/transformers/generation_strategies#custom-decoding-methods
- minimal example: transformers-community/custom_generate_example
- discussion: transformers-community/support#10
clefourrier
posted
an
update
8 months ago
Post
2130
Always surprised that so few people actually read the FineTasks blog, on
✨how to select training evals with the highest signal✨
If you're serious about training models without wasting compute on shitty runs, you absolutely should read it!!
An high signal eval actually tells you precisely, during training, how wel & what your model is learning, allowing you to discard the bad runs/bad samplings/...!
The blog covers in depth prompt choice, metrics, dataset, across languages/capabilities, and my fave section is "which properties should evals have"👌
(to know on your use case how to select the best evals for you)
Blog: HuggingFaceFW/blogpost-fine-tasks
✨how to select training evals with the highest signal✨
If you're serious about training models without wasting compute on shitty runs, you absolutely should read it!!
An high signal eval actually tells you precisely, during training, how wel & what your model is learning, allowing you to discard the bad runs/bad samplings/...!
The blog covers in depth prompt choice, metrics, dataset, across languages/capabilities, and my fave section is "which properties should evals have"👌
(to know on your use case how to select the best evals for you)
Blog: HuggingFaceFW/blogpost-fine-tasks
-
- 2 replies
Post
5138
- 1 reply
philschmid
posted
an
update
9 months ago
Post
5162
Gemini 2.5 Flash is here! We excited launch our first hybrid reasoning Gemini model. In Flash 2.5 developer can turn thinking off.
**TL;DR:**
- 🧠 Controllable "Thinking" with thinking budget with up to 24k token
- 🌌 1 Million multimodal input context for text, image, video, audio, and pdf
- 🛠️ Function calling, structured output, google search & code execution.
- 🏦 $0.15 1M input tokens; $0.6 or $3.5 (thinking on) per million output tokens (thinking tokens are billed as output tokens)
- 💡 Knowledge cut of January 2025
- 🚀 Rate limits - Free 10 RPM 500 req/day
- 🏅Outperforms 2.0 Flash on every benchmark
Try it ⬇️
https://aistudio.google.com/prompts/new_chat?model=gemini-2.5-flash-preview-04-17
**TL;DR:**
- 🧠 Controllable "Thinking" with thinking budget with up to 24k token
- 🌌 1 Million multimodal input context for text, image, video, audio, and pdf
- 🛠️ Function calling, structured output, google search & code execution.
- 🏦 $0.15 1M input tokens; $0.6 or $3.5 (thinking on) per million output tokens (thinking tokens are billed as output tokens)
- 💡 Knowledge cut of January 2025
- 🚀 Rate limits - Free 10 RPM 500 req/day
- 🏅Outperforms 2.0 Flash on every benchmark
Try it ⬇️
https://aistudio.google.com/prompts/new_chat?model=gemini-2.5-flash-preview-04-17
- 1 reply
pcuenq
authored
a
paper
9 months ago
eliebak
authored
a
paper
9 months ago
Post
2336
‼️ huggingface_hub's v0.30.0 is out with our biggest update of the past two years!
Full release notes: https://github.com/huggingface/huggingface_hub/releases/tag/v0.30.0.
🚀 Ready. Xet. Go!
Xet is a groundbreaking new protocol for storing large objects in Git repositories, designed to replace Git LFS. Unlike LFS, which deduplicates files, Xet operates at the chunk level—making it a game-changer for AI builders collaborating on massive models and datasets. Our Python integration is powered by [xet-core](https://github.com/huggingface/xet-core), a Rust-based package that handles all the low-level details.
You can start using Xet today by installing the optional dependency:
With that, you can seamlessly download files from Xet-enabled repositories! And don’t worry—everything remains fully backward-compatible if you’re not ready to upgrade yet.
Blog post: https://huggingface.co/blog/xet-on-the-hub
Docs: https://huggingface.co/docs/hub/en/storage-backends#xet
⚡ Inference Providers
- We’re thrilled to introduce Cerebras and Cohere as official inference providers! This expansion strengthens the Hub as the go-to entry point for running inference on open-weight models.
- Novita is now our 3rd provider to support text-to-video task after Fal.ai and Replicate.
- Centralized billing: manage your budget and set team-wide spending limits for Inference Providers! Available to all Enterprise Hub organizations.
- No more timeouts when generating videos, thanks to async calls. Available right now for Fal.ai, expecting more providers to leverage the same structure very soon!
Full release notes: https://github.com/huggingface/huggingface_hub/releases/tag/v0.30.0.
🚀 Ready. Xet. Go!
Xet is a groundbreaking new protocol for storing large objects in Git repositories, designed to replace Git LFS. Unlike LFS, which deduplicates files, Xet operates at the chunk level—making it a game-changer for AI builders collaborating on massive models and datasets. Our Python integration is powered by [xet-core](https://github.com/huggingface/xet-core), a Rust-based package that handles all the low-level details.
You can start using Xet today by installing the optional dependency:
pip install -U huggingface_hub[hf_xet]With that, you can seamlessly download files from Xet-enabled repositories! And don’t worry—everything remains fully backward-compatible if you’re not ready to upgrade yet.
Blog post: https://huggingface.co/blog/xet-on-the-hub
Docs: https://huggingface.co/docs/hub/en/storage-backends#xet
⚡ Inference Providers
- We’re thrilled to introduce Cerebras and Cohere as official inference providers! This expansion strengthens the Hub as the go-to entry point for running inference on open-weight models.
- Novita is now our 3rd provider to support text-to-video task after Fal.ai and Replicate.
- Centralized billing: manage your budget and set team-wide spending limits for Inference Providers! Available to all Enterprise Hub organizations.
from huggingface_hub import InferenceClient
client = InferenceClient(provider="fal-ai", bill_to="my-cool-company")
image = client.text_to_image(
"A majestic lion in a fantasy forest",
model="black-forest-labs/FLUX.1-schnell",
)
image.save("lion.png")- No more timeouts when generating videos, thanks to async calls. Available right now for Fal.ai, expecting more providers to leverage the same structure very soon!
- 6 replies
philschmid
authored
a
paper
9 months ago
philschmid
posted
an
update
10 months ago
Post
4426
Gemini 2.5 Pro, thinking by default! We excited launch our best Gemini model for reasoning, multimodal and coding yet! #1 on LMSYS, Humanity’s Last Exam, AIME and GPQA and more!
TL;DR:
- 💻 Best Gemini coding model yet, particularly for web development (excels on LiveCodeBench).
- 🧠 Default "Thinking" with up to 64k token output
- 🌌 1 Million multimodal input context for text, image, video, audio, and pdf
- 🛠️ Function calling, structured output, google search & code execution.
- 🏆 #1 on LMArena & sota on AIME, GPQA, Humanity's Last Exam
- 💡 Knowledge cut of January 2025
- 🤗 Available for free as Experimental in AI Studio, Gemini API & Gemini APP
- 🚀 Rate limits - Free 2 RPM 50 req/day
Try it ⬇️
https://aistudio.google.com/?model=gemini-2.5-pro-exp-03-25
TL;DR:
- 💻 Best Gemini coding model yet, particularly for web development (excels on LiveCodeBench).
- 🧠 Default "Thinking" with up to 64k token output
- 🌌 1 Million multimodal input context for text, image, video, audio, and pdf
- 🛠️ Function calling, structured output, google search & code execution.
- 🏆 #1 on LMArena & sota on AIME, GPQA, Humanity's Last Exam
- 💡 Knowledge cut of January 2025
- 🤗 Available for free as Experimental in AI Studio, Gemini API & Gemini APP
- 🚀 Rate limits - Free 2 RPM 50 req/day
Try it ⬇️
https://aistudio.google.com/?model=gemini-2.5-pro-exp-03-25
- 3 replies
Post
2112
Google just dropped an exciting technical report for the brand-new Gemma3 model! 🚀 Here are my personal notes highlighting the most intriguing architectural innovations, design choices, and insights from this release:
1) Architecture choices:
> No more softcaping, replace by QK-Norm
> Both Pre AND Post Norm
> Wider MLP than Qwen2.5, ~ same depth
> SWA with 5:1 and 1024 (very small and cool ablation on the paper!)
> No MLA to save KV cache, SWA do the job!
2) Long context
> Only increase the rope in the global layer (to 1M)
> Confirmation that it's harder to do long context for smol models, no 128k for the 1B
> Pretrained with 32k context? seems very high
> No yarn nor llama3 like rope extension
3) Distillation
> Only keep te first 256 logits for the teacher
> Ablation on the teacher gap (tl;dr you need some "patience" to see that using a small teacher is better)
> On policy distillation yeahh (by
@agarwl_
et al), not sure if the teacher gap behave the same here, curious if someone have more info?
4) Others
> Checkpoint with QAT, that's very cool
> RL using improve version of BOND, WARM/WARP good excuse to look at
@ramealexandre
papers
> Only use Zero3, no TP/PP if i understand correctly ?
> Training budget relatively similar than gemma2
1) Architecture choices:
> No more softcaping, replace by QK-Norm
> Both Pre AND Post Norm
> Wider MLP than Qwen2.5, ~ same depth
> SWA with 5:1 and 1024 (very small and cool ablation on the paper!)
> No MLA to save KV cache, SWA do the job!
2) Long context
> Only increase the rope in the global layer (to 1M)
> Confirmation that it's harder to do long context for smol models, no 128k for the 1B
> Pretrained with 32k context? seems very high
> No yarn nor llama3 like rope extension
3) Distillation
> Only keep te first 256 logits for the teacher
> Ablation on the teacher gap (tl;dr you need some "patience" to see that using a small teacher is better)
> On policy distillation yeahh (by
@agarwl_
et al), not sure if the teacher gap behave the same here, curious if someone have more info?
4) Others
> Checkpoint with QAT, that's very cool
> RL using improve version of BOND, WARM/WARP good excuse to look at
@ramealexandre
papers
> Only use Zero3, no TP/PP if i understand correctly ?
> Training budget relatively similar than gemma2
- 1 reply