Instructions to use Dibachain/Diba-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Dibachain/Diba-Base with Transformers:

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

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

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Dibachain/Diba-Base", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Dibachain/Diba-Base with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/Dibachain/Diba-Base

SGLang

How to use Dibachain/Diba-Base 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 "Dibachain/Diba-Base" \
    --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": "Dibachain/Diba-Base",
		"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 "Dibachain/Diba-Base" \
        --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": "Dibachain/Diba-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Dibachain/Diba-Base with Docker Model Runner:
```
docker model run hf.co/Dibachain/Diba-Base
```

CPU inference throughput at 8k context and recurrent hybrid scaling

by AndrewThompson1233 - opened 1 day ago

Discussion

AndrewThompson1233

1 day ago

Hi Dibachain team,

Impressive work shipping a Persian-first 4B foundation model optimized for offline CPU deployment via GGUF and llama.cpp.

Given your focus on local CPU execution with 8,192 context lengths:
Token generation on CPUs is strictly memory-bandwidth bound (typically 40-80 GB/s on DDR4/DDR5 channels).
In standard full-attention or dense GQA backbones, the KV cache footprint at 8k context causes significant memory bus stalls, leading to noticeable generation slowdowns as context fills up.

In an open architecture project called Maba (101M reference model: https://huggingface.co/AndrewThompson1233/maba-v1-architecture), we address this compute profile using a hybrid recurrent design:
75% linear recurrent layers (GDN-2) paired with 25% grouped-query attention.
Linear recurrent layers maintain a fixed O(1) state per layer regardless of sequence length.
On CPU architectures, this cuts memory bandwidth pressure by roughly 70% at 8k context, keeping decoding speeds flat instead of degrading over long documents or tool-calling traces.

Additionally, if your Persian tokenizer uses an expanded vocabulary to maintain high fertility on ZWNJ and Arabic script, factorizing the embedding table through a low-rank bottleneck recovers significant parameter budget to keep the GGUF footprint within tight L3/RAM constraints.

What attention mechanism does the custom Diba architecture use under the hood, and how flat does your CPU token generation speed remain from 1k out to 8k context?

Best,
Andrew

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