Instructions to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("ByteDance/Ouro-2.6B-Thinking")
model = PeftModel.from_pretrained(base_model, "LatentMT/LatentMT-2.6B-eng-latn-kas-deva")

Transformers

How to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva with Transformers:

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

pipe = pipeline("text-generation", model="LatentMT/LatentMT-2.6B-eng-latn-kas-deva")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("LatentMT/LatentMT-2.6B-eng-latn-kas-deva", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LatentMT/LatentMT-2.6B-eng-latn-kas-deva"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LatentMT/LatentMT-2.6B-eng-latn-kas-deva",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/LatentMT/LatentMT-2.6B-eng-latn-kas-deva

SGLang

How to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva 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 "LatentMT/LatentMT-2.6B-eng-latn-kas-deva" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LatentMT/LatentMT-2.6B-eng-latn-kas-deva",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "LatentMT/LatentMT-2.6B-eng-latn-kas-deva" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LatentMT/LatentMT-2.6B-eng-latn-kas-deva",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use LatentMT/LatentMT-2.6B-eng-latn-kas-deva with Docker Model Runner:
```
docker model run hf.co/LatentMT/LatentMT-2.6B-eng-latn-kas-deva
```

LatentMT-2.6B-eng-latn-kas-deva

This repository includes the LoRA adapter checkpoint for eng_Latn-kas_Deva from the paper LatentMT: Machine Translation with Latent Reasoning.

It reflects the paper's trained latent-reasoning setting, where additional recurrent steps are spent inside hidden states rather than exposed as generated chain-of-thought tokens.

The repository makes this efficient translation setup directly reusable through the included adapter weights and metadata.

Checkpoint Information

Language pair: eng_Latn-kas_Deva
Recurrent depth: 4

Only adapter release files are included in this repository: adapter_config.json, adapter_model.safetensors or adapter_model.bin, and README.md.

Environment

The relevant dependency requirement specifiers are:

torch==2.7.1
transformers==4.56.2
datasets>=2.14.0
peft>=0.10.0
bitsandbytes>=0.41.0

Loading

from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from peft import PeftConfig, PeftModel

base_model_id = "ByteDance/Ouro-2.6B-Thinking"
adapter_id = "LatentMT/LatentMT-2.6B-eng-latn-kas-deva"
total_ut_steps = 4

peft_config = PeftConfig.from_pretrained(adapter_id)
base_model_id = peft_config.base_model_name_or_path or base_model_id

config = AutoConfig.from_pretrained(
    base_model_id,
    trust_remote_code=True,
)
config.total_ut_steps = total_ut_steps

tokenizer = AutoTokenizer.from_pretrained(
    base_model_id,
    trust_remote_code=True,
)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    config=config,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()
model.config.use_cache = True
if getattr(model, "generation_config", None) is not None:
    model.generation_config.use_cache = True

Intended Use

This is for machine translation research.

Licenses

This repo is made open-source under the Apache License 2.0.

This adapter was trained for ByteDance/Ouro-2.6B-Thinking.

The base model was published by under Apache 2.0 (https://huggingface.co/ByteDance/Ouro-2.6B-Thinking).

Downloads last month: 7

Model tree for LatentMT/LatentMT-2.6B-eng-latn-kas-deva

Base model

ByteDance/Ouro-2.6B-Thinking

Adapter

(65)

this model