deformers 0.2.0

Reverse engineer transformer models.

Deformers

Experiments with modular neural networks by patching pre-trained LLMs.

Original Model

All experiments use the open-source model qwen/qwen3.5-9b.

Relevant configuration:

• hidden size: 4096
• number of layers: 32
• vocabulary size: 248320
• embedding size: 4096
• positional encoding: rotary
• embedding weights are not tied with the output head

The embedding layer and output head each contain approximately:

4096 × 248320 ~ 1.02B parameters

Patching Layers

Composite Embedding (Prefix Patch)

Frozen components

• tokenization (Qwen BPE tokenizer)
• transformer trunk (all the hidden transformer layers)
• positional encoding
• output head

Replaced component

The original token embedding Embedding(V=248320, D=4096) is replaced with a composite embedding layer with:

• a group dimension of 32
• an input dimension of 256 (byte values)
• an embedding dimension of 128
• for a total of 256 * 128 = 32768 parameters

So that a tensor of shape (B, 32 * S) is processed into (B, S, 32 * 128) = (B, S, 4096).

Then a regular transformer block maps these composite embeddings with the original emebddings of the Qwen model.

Input representation

Input text is tokenized using the original Qwen tokenizer.

Each token string is encoded as UTF-8 bytes.

Shorter tokens are then padded into a block of 32 bytes and the longer tokens are truncated.

Patch Training

The training was performed on a multilingual corpora with a custom loss:

$$L_{k} = || H_{k, patch}(x) − H_{k, qwen}(x) ||^{2}$$

Where:

• $k$ is the depth inside the original Qwen 3.5 model
• $H_{k, qwen}$ is the hidden state at depth $k$ in the original model
• $H_{k, patch}$ is the hiden state obtained when replacing the embedding layer

Hierarchical Softmax Head

Frozen components

• tokenizer
• embedding layer
• positional encoding
• transformer trunk (all the hidden layers)

Replaced component

Original output head Linear(4096 => 248320) is replaced by a hierarchical softmax tree.

Tokens are organized in a binary tree with a depth of 18 ~ log_2(248320)

Each token corresponds to a unique path from root to leaf.

Patch Training

Here a simple cross entropy loss is enough.

License

Licensed under the aGPLv3.