Instructions to use gszauer/Gab100M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use gszauer/Gab100M with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="gszauer/Gab100M", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("gszauer/Gab100M", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use gszauer/Gab100M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "gszauer/Gab100M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "gszauer/Gab100M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/gszauer/Gab100M

SGLang

How to use gszauer/Gab100M with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "gszauer/Gab100M" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "gszauer/Gab100M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "gszauer/Gab100M" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "gszauer/Gab100M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use gszauer/Gab100M with Docker Model Runner:
```
docker model run hf.co/gszauer/Gab100M
```

Gab 100M

Try it in your browser: giftofgab.chat

Gab is a ~100 million parameter language model trained from scratch on a Mac Studio. It is a decoder-only transformer with pre-normalization (RMSNorm), rotary position embeddings, exact-GeLU feed-forward blocks with a hidden bias, causal self-attention, and tied input/output embeddings. The canonical model definition is reference.js (included in this repo) — a dependency-free JavaScript implementation the whole training suite is matched against.

It speaks English, holds short conversations, and supports an optional <think> reasoning trace inside assistant turns.

Setting	Value
Parameters	99,753,216
Transformer blocks	13
Feature dimension	768
Attention heads	12
Head dimension	64
MLP hidden dimension	3072 (4x, with a per-neuron bias)
Context length	2048
Vocabulary size	10,000
RoPE theta	10,000 (interleaved pairs)
RMSNorm epsilon	1e-5
Precision	float32

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "gszauer/Gab100M"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

messages = [{"role": "user", "content": "Hello, how are you?"}]
input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    # enable_thinking=True,  # seed the reply with a <think> trace
    return_tensors="pt",
)
output = model.generate(input_ids, max_new_tokens=128)
print(tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True))

The Transformers port is a simple, faithful implementation that recomputes the full sequence each generation step (no KV cache) — fine for a 100M model. The browser runtime and the training suite's inference.py both implement KV-cached generation.

Files

File	What it is
`model.safetensors`	float32 weights with HF-style tensor names (per-head attention matrices fused, see below)
`gab100m.weights`	the native format: a headerless flat float32 buffer in `reference.js` `parameters()` order, loadable via `deserializeFromArrayBuffer`
`vocab.json`	the tokenizer: `{reserved, merges}`
`reference.js`	the canonical implementation — tensor library, autograd, tokenizer, model, KV cache, and AdamW trainer in dependency-free JavaScript
`modeling_gift_of_gab.py` etc.	the Transformers port (`trust_remote_code`)

Architecture notes

Attention is trained per head. Each of the 12 heads owns its own Wq, Wk, Wv of shape [768, 64] and its own output projection Wo of shape [64, 768]. Heads are never concatenated — each head projects its result back up to 768 on its own and the per-head results are summed:

Attention(x) = sum_i( softmax(causal(q_i @ k_i.T / sqrt(64))) @ v_i @ Wo_i )

Summing per-head outputs is algebraically identical to a concat plus one big [768, 768] output projection, so model.safetensors ships fused [768, 768] tensors and the port runs the standard formulation — float32 results match the reference bit for bit. The unfused per-head matrices are preserved in gab100m.weights.

RoPE uses interleaved pairs. Adjacent channels (2i, 2i+1) rotate together with freq_i = 1/10000^(2i/64) — unlike Llama-style split-half RoPE. There is no position embedding table; position enters only through RoPE.

The MLP has a hidden bias and no gate. MLP(x) = GeLU(x @ Wup + b) @ Wdown with exact (erf) GeLU.

The unembedding is tied. logits = RMSNorm(h) @ E.T — no separate output head.

The tokenizer is byte-level BPE with sequential merge application. Ids 0–255 are raw bytes; every other id is a merge of two earlier ids, learned in order. Encoding applies every merge rule in sequence over the UTF-8 bytes. Reserved tokens (<|user|>, <|assistant|>, <|end|>, <|endoftext|>, <think>, </think>) are chained merges taught before training, so they encode to a single id and are atomic.

Chat format

No system role. Turns are concatenated directly:

<|user|>QUESTION<|end|><|assistant|>ANSWER<|end|>

Thinking is optional and lives inside the assistant turn:

<|user|>QUESTION<|end|><|assistant|><think>PRIVATE_REASONING</think>ANSWER<|end|>

The chat template supports enable_thinking=True to seed the reply with <think>.

Training

Pre-trained on 33,882,142,110 tokens (Gab100MPretrain plus Cosmopedia v2), then fine-tuned into a chat model on 1,016,529,229 tokens (Gab100MFinetune plus SmolTalk), learning only from assistant replies.

AdamW with linear warmup and cosine decay: peak learning rate 3e-4 (pre-train) / 3e-5 (fine-tune), betas 0.9/0.999, weight decay 0.01 (decoupled), global-norm gradient clip 1.0. Trained with MLX on a Mac Studio.

Limitations

A compact experimental model trained on a desktop machine. It can chat in English and follow simple instructions, but it is dumb by LLM standards — expect factual errors and quirks.

gszauer
/

Gab100M