Instructions to use Efe2898/caba-memory-v2-100m-tr with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Efe2898/caba-memory-v2-100m-tr with Transformers:

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

pipe = pipeline("text-generation", model="Efe2898/caba-memory-v2-100m-tr", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Efe2898/caba-memory-v2-100m-tr", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Efe2898/caba-memory-v2-100m-tr with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Efe2898/caba-memory-v2-100m-tr"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Efe2898/caba-memory-v2-100m-tr",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Efe2898/caba-memory-v2-100m-tr

SGLang

How to use Efe2898/caba-memory-v2-100m-tr 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 "Efe2898/caba-memory-v2-100m-tr" \
    --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": "Efe2898/caba-memory-v2-100m-tr",
		"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 "Efe2898/caba-memory-v2-100m-tr" \
        --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": "Efe2898/caba-memory-v2-100m-tr",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Efe2898/caba-memory-v2-100m-tr with Docker Model Runner:
```
docker model run hf.co/Efe2898/caba-memory-v2-100m-tr
```

ÇABA Selective Convolution + EMA Memory v1

This is an experimental causal language model. It combines a train-only sparse token-transition graph, a full-support tied global softmax, a K=8 causal depthwise convolution, and a selective multi-timescale EMA memory with constant-size cache. It is experimental and does not claim validated language quality.

Architecture: caba-selective-conv-ema-v1-100m-tr-warmstart
Trainable parameters: 100,298,529
Tokenizer: vngrs-ai/Kumru-2B-Base at 55711ea224e4bf5d4e11a4baf79130ae73785ece
Tokenized data: Efe2898/tokenized at 6cb993aa63064cd89b511bb164e8c6e5512902ed
Training source weights: {"tr": 1.0}
Sequence length: 256
Graph K: 32; held-out candidate coverage: 0.50739
Run signature: ae00f5c778efe97519b5cf62c7342abafc46043c0b1dfcf78b24b7a47287680e

Loading

from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "Efe2898/caba-memory-v2-100m-tr"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

The model code and training scripts are included in this repository. Review configuration_caba.py, modeling_caba.py, and training/ before enabling trust_remote_code=True for any repository.

Inference cache: past_key_values stores (conv_state, ema_state) per layer. Both states are constant-size with respect to prompt length. Batched generation requires left padding so the final prompt position is a real token. The notebook checks full-prefix/incremental equality, the EMA recurrence, left-padding behavior, and the real Transformers generation path before training.

Downloads last month: 308

Safetensors

Model size

0.1B params

Tensor type

I64

F32

Collection including Efe2898/caba-memory-v2-100m-tr

CABA

Collection

All training checkpoints and training datasets for CABA models. • 2 items • Updated 1 day ago