Instructions to use microsoft/bitnet-b1.58-2B-4T with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use microsoft/bitnet-b1.58-2B-4T with Transformers:

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

pipe = pipeline("text-generation", model="microsoft/bitnet-b1.58-2B-4T", trust_remote_code=True)
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("microsoft/bitnet-b1.58-2B-4T", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("microsoft/bitnet-b1.58-2B-4T", trust_remote_code=True, 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

vLLM

How to use microsoft/bitnet-b1.58-2B-4T with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "microsoft/bitnet-b1.58-2B-4T"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/bitnet-b1.58-2B-4T",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/microsoft/bitnet-b1.58-2B-4T

SGLang

How to use microsoft/bitnet-b1.58-2B-4T 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 "microsoft/bitnet-b1.58-2B-4T" \
    --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": "microsoft/bitnet-b1.58-2B-4T",
		"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 "microsoft/bitnet-b1.58-2B-4T" \
        --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": "microsoft/bitnet-b1.58-2B-4T",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use microsoft/bitnet-b1.58-2B-4T with Docker Model Runner:
```
docker model run hf.co/microsoft/bitnet-b1.58-2B-4T
```

Packed trits vs the bf16 master weights, and bf16 vs f32 scales (data inside)

#48

by dsbusiness20206 - opened 17 days ago

Discussion

dsbusiness20206

17 days ago

We compared this packed checkpoint with microsoft/bitnet-b1.58-2B-4T-bf16 and microsoft/bitnet-b1.58-2B-4T-gguf at their current revisions, using independent decoders. Two observations:

Same trits as the GGUF, same scale at lower precision. In layer 0 q_proj and down_proj (24.2M weights) the trits here and in the I2_S GGUF are identical. In all 210 ternary tensors, weight_scale here is the GGUF's f32 scale rounded to bf16 (for layer 0 down_proj: 2.15625 vs 2.1631613), so the two differ by 0.14% at the median and 0.36% at most.
These trits cannot be recomputed from the bf16 master weights. transformers WeightQuant (float32 mean) applied to the bf16 weights gives different trits for 1.22% of layer 0 q_proj and 0.57% of layer 0 down_proj weights. In each tensor all of them have the one bf16 value nearest the threshold, |w| = 0.5 × weight_scale, and this checkpoint maps that value sometimes to ±1 and sometimes to 0 (79,719 of 163,223 weights are ±1 in q_proj). That fits trits computed from higher-precision master weights, of which the bf16 repository is a rounded copy.

Is the f32 scale in the GGUF the reference, and should the bf16 repository (online mode) be expected to reproduce these trits?

Details, revisions and a standalone reproduction (Python standard library only): https://github.com/microsoft/BitNet/issues/632

Write-up and data: https://github.com/dmitrii-f-t27/trinity-memory/blob/master/docs/ternary-check.md

Dmitrii Fedorov, Trinity (GitHub: dmitrii-f-t27)

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment