Instructions to use zerolimites/BitNet-TinyDialogues-1.58b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use zerolimites/BitNet-TinyDialogues-1.58b with Transformers:

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

pipe = pipeline("text-generation", model="zerolimites/BitNet-TinyDialogues-1.58b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("zerolimites/BitNet-TinyDialogues-1.58b")
model = AutoModelForCausalLM.from_pretrained("zerolimites/BitNet-TinyDialogues-1.58b", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use zerolimites/BitNet-TinyDialogues-1.58b with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0
# Run inference directly in the terminal:
llama cli -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0
# Run inference directly in the terminal:
llama cli -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

Use Docker

docker model run hf.co/zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

Llama new
LM Studio
Jan

vLLM

How to use zerolimites/BitNet-TinyDialogues-1.58b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "zerolimites/BitNet-TinyDialogues-1.58b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zerolimites/BitNet-TinyDialogues-1.58b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

SGLang

How to use zerolimites/BitNet-TinyDialogues-1.58b 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 "zerolimites/BitNet-TinyDialogues-1.58b" \
    --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": "zerolimites/BitNet-TinyDialogues-1.58b",
		"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 "zerolimites/BitNet-TinyDialogues-1.58b" \
        --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": "zerolimites/BitNet-TinyDialogues-1.58b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use zerolimites/BitNet-TinyDialogues-1.58b with Ollama:
```
ollama run hf.co/zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0
```
Unsloth Desktop
Docker Model Runner
How to use zerolimites/BitNet-TinyDialogues-1.58b with Docker Model Runner:
```
docker model run hf.co/zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0
```

Lemonade

How to use zerolimites/BitNet-TinyDialogues-1.58b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull zerolimites/BitNet-TinyDialogues-1.58b:TQ1_0

Run and chat with the model

lemonade run user.BitNet-TinyDialogues-1.58b-TQ1_0

List all available models

lemonade list

Atomic Chat

BitNet-TinyDialogues-1.58b / README.md

zerolimites

BitNet b1.58 TinyDialogues: trained checkpoint, GGUF, scripts, patches

655f931 verified 1 day ago

preview code

Raw History Blame Contribute Delete

2.56 kB

---
license: mit
language:
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- bitnet
- bitnet-b1.58
- ternary
- gguf
- cpu
- tinydialogues
datasets:
- styfeng/TinyDialogues
metrics:
- perplexity
---
# BitNet TinyDialogues 1.58b

Modelo experimental **BitNet b1.58-style** treinado do zero em CPU com
[`styfeng/TinyDialogues`](https://huggingface.co/datasets/styfeng/TinyDialogues).

## O que é

- Transformer causal pequeno com `BitLinear`: pesos ternários `{-1, 0, +1}` no forward (1,58 bit/peso).
- Ativações quantizadas em 8 bits por token durante o treino (W1.58A8).
- FFN com `ReLU²`, RoPE, RMSNorm/subLN, embeddings tied.
- GGUF: `TQ1_0` e `TQ2_0` (ternários). O arquivo também contém embeddings/normas em maior precisão.

> “1-bit BitNet” aqui significa a família BitNet b1.58 ternária, não 1 bit matemático por peso.

## Arquitetura

| Campo | Valor |
|---|---|
| Parâmetros | ~2,89 M |
| Camadas | 4 |
| Hidden | 256 |
| FFN | 512 |
| Heads | 4 (KV 2) |
| Vocab | 2048 BPE SentencePiece |
| Contexto | 128 tokens |
| Treino | 8000 steps, batch 8, seq 128, CPU |
| Melhor val loss | 2.5838 (step 7000) |
| Perplexidade val | ~13.25 |

## Arquivos

- `gguf/bitnet-tinydialogues-TQ1_0.gguf` — principal, menor (~1,1 MiB).
- `gguf/bitnet-tinydialogues-TQ2_0.gguf` — alternativa ternária (~1,2 MiB).
- `gguf/bitnet-tinydialogues-f32.gguf` — referência F32 (~11 MiB).
- `model.safetensors`, `config.json`, `tokenizer.model`, `tokenizer.json` — checkpoint HF.
- `scripts/` — preparação, treino, exportação e demo.
- `patches/llama.cpp-bitnet-fixes.patch` — patch ReLU² + mapeamento de tensores.
- `Dockerfile` — Space Docker/Gradio para CPU (requer PRO ou hardware pago).

## Uso rápido com llama.cpp

O runtime precisa usar `ReLU²` no FFN do BitNet. O llama.cpp upstream ainda pode usar SiLU hardcoded;
aplique `patches/llama.cpp-bitnet-fixes.patch` ou use o `Dockerfile`.

```bash
llama-cli -m gguf/bitnet-tinydialogues-TQ1_0.gguf \
  -p "user: can we play? assistant:" -n 64 --temp 0.7 -t 2 -r "user:"
```

Exemplo real (TQ1_0, greedy):
```text
user: can we play? assistant: yes, but we need to clean up. we need to clean up.
```
## HF demo (ternário)
```bash
python scripts/demo_hf.py --model . --prompt "user: can we play? assistant:" --tokens 40
```
## Limitações
Modelo muito pequeno, treinado apenas em diálogos infantis em inglês. Serve para demonstrar
BitNet/GGUF/CPU; não é um assistente geral nem tem conhecimento confiável.
## Licença
MIT. Dataset TinyDialogues: MIT.

	---
	license: mit
	language:
	- en
	pipeline_tag: text-generation
	library_name: transformers
	tags:
	- bitnet
	- bitnet-b1.58
	- ternary
	- gguf
	- cpu
	- tinydialogues
	datasets:
	- styfeng/TinyDialogues
	metrics:
	- perplexity
	---

	# BitNet TinyDialogues 1.58b

	Modelo experimental BitNet b1.58-style treinado do zero em CPU com
	[`styfeng/TinyDialogues`](https://huggingface.co/datasets/styfeng/TinyDialogues).

	## O que é

	- Transformer causal pequeno com `BitLinear`: pesos ternários `{-1, 0, +1}` no forward (1,58 bit/peso).
	- Ativações quantizadas em 8 bits por token durante o treino (W1.58A8).
	- FFN com `ReLU²`, RoPE, RMSNorm/subLN, embeddings tied.
	- GGUF: `TQ1_0` e `TQ2_0` (ternários). O arquivo também contém embeddings/normas em maior precisão.

	> “1-bit BitNet” aqui significa a família BitNet b1.58 ternária, não 1 bit matemático por peso.

	## Arquitetura

	\| Campo \| Valor \|
	\|---\|---\|
	\| Parâmetros \| ~2,89 M \|
	\| Camadas \| 4 \|
	\| Hidden \| 256 \|
	\| FFN \| 512 \|
	\| Heads \| 4 (KV 2) \|
	\| Vocab \| 2048 BPE SentencePiece \|
	\| Contexto \| 128 tokens \|
	\| Treino \| 8000 steps, batch 8, seq 128, CPU \|
	\| Melhor val loss \| 2.5838 (step 7000) \|
	\| Perplexidade val \| ~13.25 \|

	## Arquivos

	- `gguf/bitnet-tinydialogues-TQ1_0.gguf` — principal, menor (~1,1 MiB).
	- `gguf/bitnet-tinydialogues-TQ2_0.gguf` — alternativa ternária (~1,2 MiB).
	- `gguf/bitnet-tinydialogues-f32.gguf` — referência F32 (~11 MiB).
	- `model.safetensors`, `config.json`, `tokenizer.model`, `tokenizer.json` — checkpoint HF.
	- `scripts/` — preparação, treino, exportação e demo.
	- `patches/llama.cpp-bitnet-fixes.patch` — patch ReLU² + mapeamento de tensores.
	- `Dockerfile` — Space Docker/Gradio para CPU (requer PRO ou hardware pago).

	## Uso rápido com llama.cpp

	O runtime precisa usar `ReLU²` no FFN do BitNet. O llama.cpp upstream ainda pode usar SiLU hardcoded;
	aplique `patches/llama.cpp-bitnet-fixes.patch` ou use o `Dockerfile`.

	```bash
	llama-cli -m gguf/bitnet-tinydialogues-TQ1_0.gguf \
	-p "user: can we play? assistant:" -n 64 --temp 0.7 -t 2 -r "user:"
	```

	Exemplo real (TQ1_0, greedy):

	```text
	user: can we play? assistant: yes, but we need to clean up. we need to clean up.
	```

	## HF demo (ternário)

	```bash
	python scripts/demo_hf.py --model . --prompt "user: can we play? assistant:" --tokens 40
	```

	## Limitações

	Modelo muito pequeno, treinado apenas em diálogos infantis em inglês. Serve para demonstrar
	BitNet/GGUF/CPU; não é um assistente geral nem tem conhecimento confiável.

	## Licença

	MIT. Dataset TinyDialogues: MIT.