Instructions to use tevfikk/granite-docling-258M-bitnet with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tevfikk/granite-docling-258M-bitnet with Transformers:

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

pipe = pipeline("text-generation", model="tevfikk/granite-docling-258M-bitnet", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("tevfikk/granite-docling-258M-bitnet", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tevfikk/granite-docling-258M-bitnet with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tevfikk/granite-docling-258M-bitnet"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tevfikk/granite-docling-258M-bitnet",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tevfikk/granite-docling-258M-bitnet

SGLang

How to use tevfikk/granite-docling-258M-bitnet 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 "tevfikk/granite-docling-258M-bitnet" \
    --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": "tevfikk/granite-docling-258M-bitnet",
		"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 "tevfikk/granite-docling-258M-bitnet" \
        --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": "tevfikk/granite-docling-258M-bitnet",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tevfikk/granite-docling-258M-bitnet with Docker Model Runner:
```
docker model run hf.co/tevfikk/granite-docling-258M-bitnet
```

Granite-Docling-258M-BitNet-1.58b 🚀

This repository hosts the first-ever BitNet b1.58 ternary (-1, 0, 1) quantized edition of IBM's ibm-granite/granite-docling-258M.

Trained via Knowledge Distillation with Straight-Through Estimators (STE) on the Edge-Intelligence framework, this model replaces 210 matrix multiplication operations with multiplication-free ternary addition and subtraction, reducing parameter memory footprint by ~3.7x while achieving 93.7% logit fidelity to the original FP32 teacher model.

📊 Benchmark Highlights

Metric	Original FP32	BitNet b1.58 (This Model)	Improvement
Model Size	530 MB	143.85 MB	3.68x Compression
Weight Precision	16-bit Float	1.58-bit Ternary (-1, 0, 1)	10.1x weight density
Logit Cosine Similarity	1.0000	0.9374 (%93.74)	High semantic fidelity
Top-1 Token Agreement	`<doctag>` (100327)	`<doctag>` (100327)	100% Match
C99 Inference Throughput (AVX2)	~40 tok/s	156.6 tok/s	~3.9x Speedup
Latency per Token (C99)	25.0 ms	6.38 ms	Ultra-responsive

🛠️ Usage via Hugging Face (`transformers`)

You can load and run this model directly using trust_remote_code=True:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tevfikk/granite-docling-258M-bitnet"

# 1. Load Tokenizer & Model
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True, 
    torch_dtype=torch.float32
)

# 2. Format Input with Docling Chat Template
prompt = "<|start_of_role|>user<|end_of_role|>Convert this document to markdown.<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>"
inputs = tokenizer(prompt, return_tensors="pt")

# 3. Generate DocTags
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=32, do_sample=False)

print(tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:]))
# Output: <doctag><section_header_level_1>...

⚡ Edge Binary (.eifm)

Included in this repository is granite_docling_bitnet_dense.eifm (143 MB), a zero-heap, single-file binary container format optimized for multiplication-free embedded C99 inference on microcontrollers, Raspberry Pi, and edge accelerators using EIF-Runtime:

# Clone the open-source C99 runtime & build
git clone https://github.com/tevfik/eif-runtime.git
cd eif-runtime && cmake -B build && cmake --build build

# Run at 156+ tokens/sec without Python or CUDA!
./build/eif-run granite_docling_bitnet_dense.eifm "Invoice Summary" 64

📜 Citation & Credits

Base Model: IBM Granite Docling 258M
BitNet 1.58-bit Architecture: The Era of 1-bit LLMs (Microsoft Research)
Distillation & Quantization: Edge-Intelligence Research

Downloads last month: -

Safetensors

Model size

0.2B params

Tensor type

BF16

Model tree for tevfikk/granite-docling-258M-bitnet

Base model

ibm-granite/granite-docling-258M

Finetuned

(11)

this model

Paper for tevfikk/granite-docling-258M-bitnet

The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

Paper • 2402.17764 • Published Feb 27, 2024 • 630