* Add the flux autoencoder.
* Add the encoder down-blocks.
* Upsampling in the decoder.
* Sketch the flow matching model.
* More flux model.
* Add some of the positional embeddings.
* Add the rope embeddings.
* Add the sampling functions.
* Add the flux example.
* Fix the T5 bits.
* Proper T5 tokenizer.
* Clip encoder path fix.
* Get the clip embeddings.
* No configurable weights in layer norm.
* More weights related fixes.
* Yet another shape fix.
* DType fix.
* Fix a couple more shape issues.
* DType fixes.
* Fix the latent dims.
* Fix more shape issues.
* Autoencoder fixes.
* Get some generations out.
* Bugfix.
* T5 padding.
* Clippy fix.
* Add the decode only mode.
* Fix.
* More fixes.
* Finally get some generations to work.
* Add readme.
* bert attention mask
* Allow for using None as a mask.
* Revert part of the changes so that the proper default mask applies.
* Cosmetic change.
* Another cosmetic tweak.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* Add Llama 3.1 rope
* Clippy
* Format
* Clippy
* Add support for multiple eos tokens:
* Untagged either
* Remove either dep and fix settings.json
* Make the max positional embeddings configurable
* onnx: fix pad, unsqueeze
both implementations have off-by-one errors:
- Pad 'reflect' cycle for eg `dim==3` is `[0,1,2,1]` which has length of
4 (or `dim*2 - 2`) not 5 (current code `dim*2 - 1`)
- Unsqueeze(-1) for tensor with `dim==3` should be 3 (ie `dim+index+1`)
not 2 (ie currently `dim+index`)
in addition, Pad is incorrectly calculating the starting padding.
If we want to pad out 2 elements to the start, and we have this cycle
of indices of length 6, then we should skip 4 elements, but currently
we skip 2. A more visual representation of what's going on is below:
```
pad_start: 2
data: [a,b,c,d]
indices: [0, 1, 2, 3, 2, 1, 0, 1, 2, 3, 2, 1, 0, ..] // zigzag between 0..4
actual: skip [ c d| c b a b]
expected: ~ skip ~ [ c b| a b c d]
```
The values between `[` and `|` are padding and the values between
`|` and `]` in the example should match the original data being padded.
* Fix clippy lints.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* Add gradient test for conv_transpose2d with stride of 2.
* Swap dilation and stride in ConvTranspose2D backpropagation.
Without this, a shape mismatch occurs with a stride of 2 and dilation of 1.
* Add further tests of the ConvTranspose2D gradient.
Values calculated with torch, minor numerical errors adjusted and commented.
