Skip to content

Conversation

@Joao-L-S-Almeida
Copy link
Collaborator

@Joao-L-S-Almeida Joao-L-S-Almeida commented Apr 3, 2025

  • Introducing the MoELayer.
  • A partial support for scalar models (classification).
  • An example for EuroSat.
  • Some basic tests.

To be merged in version 1.2

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
…the transformer layer input. How to automaticalluy estimate it ?

Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
@Joao-L-S-Almeida Joao-L-S-Almeida self-assigned this Apr 3, 2025
@Joao-L-S-Almeida Joao-L-S-Almeida linked an issue Apr 3, 2025 that may be closed by this pull request
@Joao-L-S-Almeida Joao-L-S-Almeida changed the title [WiP] MoE for LoRA (MoLoRA) [WiP] Attempting to use MoE for LoRA (MoLoRA) Apr 3, 2025
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
@Joao-L-S-Almeida Joao-L-S-Almeida changed the title [WiP] Attempting to use MoE for LoRA (MoLoRA) [WiP] Introducing MoELayer Apr 25, 2025
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: Joao Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
@Joao-L-S-Almeida Joao-L-S-Almeida changed the title [WiP] Introducing MoELayer [WiP] Introducing MoELayer and basic usage for the head network. May 12, 2025
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>

Completig tests for IdentityDecoder

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>

A method to unsqueeze the gating weights when necessary

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>

MoE head layer for PixelWise models

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>

Tests for PixelWise Models + MoE head layer

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
Signed-off-by: João Lucas de Sousa Almeida <[email protected]>

Annotations for MoELayer

Signed-off-by: João Lucas de Sousa Almeida <[email protected]>
@Joao-L-S-Almeida Joao-L-S-Almeida changed the title [WiP] Introducing MoELayer and basic usage for the head network. Introducing MoELayer and basic usage for the head network. May 12, 2025
@Joao-L-S-Almeida
Copy link
Collaborator Author

@paolo-fraccaro @blumenstiel @romeokienzler Is it something we want to have for 1.1 ?

@Joao-L-S-Almeida Joao-L-S-Almeida added 1.2 and removed 1.1 labels May 28, 2025
@Joao-L-S-Almeida Joao-L-S-Almeida removed the request for review from paolofraccaro August 26, 2025 12:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

MoE layers combined with LoRA (MoLoRA)

2 participants