40 227 217

Aymeric Roucher

m-ric

http://aymeric-roucher.github.io

AI & ML interests

MLE at Hugging Face 🤗 LLMs, Agents, RAG, Multimodal.

Articles

Organizations

Posts 49

Post

1049

💥 𝐋-𝐌𝐮𝐥: 𝐀𝐝𝐝𝐢𝐭𝐢𝐨𝐧-𝐎𝐧𝐥𝐲 𝐌𝐮𝐥𝐭𝐢𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐜𝐚𝐧 𝐬𝐥𝐚𝐬𝐡 𝐜𝐨𝐦𝐩𝐮𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐜𝐨𝐬𝐭𝐬 𝐛𝐲 𝟖𝟎%!

Microsoft researchers dropped a groundbreaking technique that could slash the energy use in transformer computations : their novel "linear-complexity multiplication" (L-Mul) algorithm approximates floating-point multiplication using energy-efficient integer addition instead of costly multiplications.

💡 Quick reminder on how floats are coded on 8 bits (FP8):
In the e4m3 FP8 standard, you encode a number as:
Sign (1 bit) | Exponent (4 bits) | Mantissa (3 bits)
Example: 0 (positive) | 1000 (8) | 101 (1/2 + 1/8 = 0.625)
Calculation: you add one to the mantissa, and multiply it by 2 power (the exponent - a bias term which is 7 for e4m3):

➡️ You get (1 + 0.625) × 2^(8-7) = 3.25

Now back to the paper. 𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀:

⚡️ Multiplication is extremely energy-intensive compared to addition. For 32-bit operations, multiplication (3.7 pJ) uses 37x more energy than addition (0.1 pJ)!

🧮 Traditional floating-point multiplication go like (noting xm the mantissa and xe the exponent): Mul(x,y) = (1 + xm) · 2^xe · (1 + ym) · 2^ye = (1 + xm + ym + xm · ym) · 2^(xe+ye)

💡 L-Mul cleverly approximates this as: L-Mul(x,y) = (1 + xm + ym + 2^-l(m)) · 2^(xe+ye), eliminating the costly xm · ym term

🔧 l(m) term is adaptively set based on mantissa size for optimal accuracy

📊 Benchmarks on the Llama-3.1-8B-Instruct model show L-Mul preserves precision across various NLP tasks, with performance nearly identical to full BFloat16 precision

💬 Authors claim: "We can achieve the same model inference performance while reducing the energy cost of attention computations by 80%."

This breakthrough is still theoretical and would need implementation on dedicated hardware to confirm real-world gains, but it’s a really exciting path for more sustainable AI! 🌱

Read the paper here 👉 Addition is All You Need for Energy-efficient Language Models (2410.00907)

Post

2869

📜 𝐎𝐥𝐝-𝐬𝐜𝐡𝐨𝐨𝐥 𝐑𝐍𝐍𝐬 𝐜𝐚𝐧 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐫𝐢𝐯𝐚𝐥 𝐟𝐚𝐧𝐜𝐲 𝐭𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫𝐬!

Researchers from Mila and Borealis AI just have shown that simplified versions of good old Recurrent Neural Networks (RNNs) can match the performance of today's transformers.

They took a fresh look at LSTMs (from 1997!) and GRUs (from 2014). They stripped these models down to their bare essentials, creating "minLSTM" and "minGRU". The key changes:
❶ Removed dependencies on previous hidden states in the gates
❷ Dropped the tanh that had been added to restrict output range in order to avoid vanishing gradients
❸ Ensured outputs are time-independent in scale (not sure I understood that well either, don't worry)

⚡️ As a result, you can use a “parallel scan” algorithm to train these new, minimal RNNs, in parallel, taking 88% more memory but also making them 200x faster than their traditional counterparts for long sequences

🔥 The results are mind-blowing! Performance-wise, they go toe-to-toe with Transformers or Mamba.

And for Language Modeling, they need 2.5x fewer training steps than Transformers to reach the same performance! 🚀

🤔 Why does this matter?

By showing there are simpler models with similar performance to transformers, this challenges the narrative that we need advanced architectures for better performance!

💬 François Chollet wrote in a tweet about this paper:

“The fact that there are many recent architectures coming from different directions that roughly match Transformers is proof that architectures aren't fundamentally important in the curve-fitting paradigm (aka deep learning)”

“Curve-fitting is about embedding a dataset on a curve. The critical factor is the dataset, not the specific hard-coded bells and whistles that constrain the curve's shape.”

It’s the Bitter lesson by Rich Sutton striking again: don’t need fancy thinking architectures, just scale up your model and data!

Read the paper 👉 Were RNNs All We Needed? (2410.01201)

View all posts