SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations
Paper
ā¢
2512.14080
ā¢
Published
ā¢
6
Nanotron Research
community
AI & ML interests
Large scale distributed AI model training, model parallelisation, low-level GPU acceleration, make GPUs go brrrrr
Recent Activity
View all activity
eliebakĀ
submitted
a
paper
to Daily Papers
20 days ago
nouamanetaziĀ
posted
an
update
2 months ago
Post
4197
After training šš¦šØš„ššš on ššš ššššš¬ for nearly a month, I've come to realize something most people overlook: š¢š§šš«šš¬šš«š®ššš®š«š š¢š¬ šš”š š¦šš¤š-šØš«-šš«ššš¤ šššššØš« š¢š§ ššš šš«šš¢š§š¢š§š . š„
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 šššš šš«š«šØš«š¬, or when your expensive GPU cluster is running at šš% šššš¢šš¢šš§šš², the problem isn't your model. It's most probably a š¦š¢š¬š®š¬š šØš šš”š š”šš«šš°šš«š. š ļø
Questions that seemed simple but had no clear answers: Why is ššØš šš«šš¢š§š¢š§š š¬š„šØš°šš« šš”šš§ ššš§š¬š š¦šØššš„š¬? Which šššš šš„šš š¬ should we actually set? How often should we checkpoint without killing throughput?
That's why we built šš”š šš¦šØš„ šš«šš¢š§š¢š§š šš„šš²ššØšØš¤ š: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the š¢š§šš«šš¬šš«š®ššš®š«š š„šš²šš« that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: šššš š”š¢ššš¢š§š š šš/š¬, šššš¢š§š¤ š.š š«šššš”š¢š§š ššš šš/š¬, šššš ššš§š šš šš.š šš/š¬. Then we ran collective operations across ššš šššš¬ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ššš šš/š¬ on a single node to ššš-ššš šš/š¬ 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.
šš”š šš¦šØš„ šš«šš¢š§š¢š§š šš„šš²ššØšØš¤: https://lnkd.in/e5MKXUHS
Shared with ā¤ļø by the HuggingFace team
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 šššš šš«š«šØš«š¬, or when your expensive GPU cluster is running at šš% šššš¢šš¢šš§šš², the problem isn't your model. It's most probably a š¦š¢š¬š®š¬š šØš šš”š š”šš«šš°šš«š. š ļø
Questions that seemed simple but had no clear answers: Why is ššØš šš«šš¢š§š¢š§š š¬š„šØš°šš« šš”šš§ ššš§š¬š š¦šØššš„š¬? Which šššš šš„šš š¬ should we actually set? How often should we checkpoint without killing throughput?
That's why we built šš”š šš¦šØš„ šš«šš¢š§š¢š§š šš„šš²ššØšØš¤ š: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the š¢š§šš«šš¬šš«š®ššš®š«š š„šš²šš« that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: šššš š”š¢ššš¢š§š š šš/š¬, šššš¢š§š¤ š.š š«šššš”š¢š§š ššš šš/š¬, šššš ššš§š šš šš.š šš/š¬. Then we ran collective operations across ššš šššš¬ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ššš šš/š¬ on a single node to ššš-ššš šš/š¬ 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.
šš”š šš¦šØš„ šš«šš¢š§š¢š§š šš„šš²ššØšØš¤: https://lnkd.in/e5MKXUHS
Shared with ā¤ļø by the HuggingFace team
thomwolfĀ
authored
a
paper
3 months ago
lvwerraĀ
authored
a
paper
3 months ago
Post
3898
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 thereāll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST š¤
science
Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe thereāll be a shiny new release to talk about?
Thursday 4th September, 8AM-11AM PST š¤
Post
714
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
dark-mode
š¤
1
11
#82 opened 11 months ago
by
serhany
tfrereĀ
updated
a
Space
5 months ago
thomwolfĀ
updated
a
Space
5 months ago
order button
#118 opened 5 months ago
by
lvwerra
order button
#118 opened 5 months ago
by
lvwerra
lvwerraĀ
updated
a
Space
5 months ago
julien-cĀ
updated
a
Space
5 months ago
Update README.md
#1 opened 5 months ago
by
lvwerra
Update README.md
#2 opened 5 months ago
by
lvwerra
Update README.md
#3 opened 5 months ago
by
lvwerra
julien-cĀ
updated
a
dataset
5 months ago
julien-cĀ
published
a
dataset
5 months ago
Post
4795
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