Post
604
Some interesting architectural choices made in Llama 4 models -- were these key to the 10M context? Possibly 🤔
🔍 Takeaways:
🧩 Interleaved Attention without position encoding
- LLaMA 4 removes explicit positional encoding in some attention layers to boost performance on longer contexts.
- The principles here could be similar to the residual connections to facilitate attention to early tokens without positional decay.
⚖️ Scaled Softmax to increase attention at inference time
- The max attention value (output of softmax) decreases as context size increases.
- Llama 4 incorporates a context-size dependent temperature in the softmax function to modify the slope of softmax, allowing the model to focus better on relevant tokens.
- Done only at inference time -- guessing it was more a choice after some observation on eval datasets.
What did you think of these choices?
🔍 Takeaways:
🧩 Interleaved Attention without position encoding
- LLaMA 4 removes explicit positional encoding in some attention layers to boost performance on longer contexts.
- The principles here could be similar to the residual connections to facilitate attention to early tokens without positional decay.
⚖️ Scaled Softmax to increase attention at inference time
- The max attention value (output of softmax) decreases as context size increases.
- Llama 4 incorporates a context-size dependent temperature in the softmax function to modify the slope of softmax, allowing the model to focus better on relevant tokens.
- Done only at inference time -- guessing it was more a choice after some observation on eval datasets.
What did you think of these choices?