File tree Expand file tree Collapse file tree
Expand file tree Collapse file tree Original file line number Diff line number Diff line change 11# TIDE -- Token-Informed Depth Execution
22
3+ <p align =" center " >
4+ <img src =" assets/tide-diagram.svg " alt =" TIDE: Per-token early exit for transformer inference " width =" 100% " />
5+ </p >
6+
37** Make any LLM faster by skipping layers tokens don't need.**
48
59TIDE learns which tokens are "easy" (converge early) and which are "hard" (need all layers).
610Easy tokens exit early. Hard tokens go deep. No model retraining. No architecture changes.
711Drop it onto any HuggingFace model in 3 lines.
812
9- ```
10- Standard LLM TIDE LLM
11- ========== ========
12-
13- "The cat sat" "The cat sat"
14- | | | | | |
15- [ Layer 1 Layer 1 Layer 1 ] [ Layer 1 Layer 1 Layer 1 ]
16- | | | | | |
17- [ Layer 2 Layer 2 Layer 2 ] [ Layer 2 Layer 2 Layer 2 ]
18- | | | | | |
19- [ Layer 3 Layer 3 Layer 3 ] [ Layer 3 Layer 3 Layer 3 ]
20- | | | | | |----> converged! exit.
21- [ Layer 4 Layer 4 Layer 4 ] [ Layer 4 Layer 4 ] |
22- | | | | | |
23- ... ... ... ... ... |
24- | | | | | |
25- [ Layer N Layer N Layer N ] [ Layer N Layer N ] |
26- | | | | | |
27- logits logits logits logits logits logits
28-
29- Every token runs every layer. Easy tokens exit early.
30- N layers x 3 tokens = 3N ops. Fewer ops. Same quality.
31- ```
32-
3313## Install
3414
3515``` bash
You can’t perform that action at this time.
0 commit comments