Instructions to use rpDungeon/marvin-cpt-12b-adapter with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use rpDungeon/marvin-cpt-12b-adapter with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("/workspace/models/gemma-4-12B-it")
model = PeftModel.from_pretrained(base_model, "rpDungeon/marvin-cpt-12b-adapter")

Transformers

How to use rpDungeon/marvin-cpt-12b-adapter with Transformers:

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

pipe = pipeline("text-generation", model="rpDungeon/marvin-cpt-12b-adapter")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("rpDungeon/marvin-cpt-12b-adapter", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use rpDungeon/marvin-cpt-12b-adapter with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "rpDungeon/marvin-cpt-12b-adapter"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rpDungeon/marvin-cpt-12b-adapter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/rpDungeon/marvin-cpt-12b-adapter

SGLang

How to use rpDungeon/marvin-cpt-12b-adapter 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 "rpDungeon/marvin-cpt-12b-adapter" \
    --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": "rpDungeon/marvin-cpt-12b-adapter",
		"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 "rpDungeon/marvin-cpt-12b-adapter" \
        --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": "rpDungeon/marvin-cpt-12b-adapter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use rpDungeon/marvin-cpt-12b-adapter with Docker Model Runner:
```
docker model run hf.co/rpDungeon/marvin-cpt-12b-adapter
```

marvin-cpt-12b-adapter

Masked continued-pretraining (CPT) LoRA adapter for gemma-4-12B-it.

Summary

Type: PEFT LoRA adapter (r=32, α=64, bf16)
Base: gemma-4-12B-it (internal instruct base)
Objective: masked CPT on the marvin corpus (private, non-redistributable training data — not included here)
Adapter size: ~500 MB (adapter_model.safetensors, 656 tensors)

Training

1 epoch, max sequence length 32768 (split-not-truncate, packed with per-doc isolation)
Optimizer: paged AdamW 8-bit, lr 1e-5, constant schedule w/ warmup-ratio 0.05, seed 42
Effective batch 4 (mbs 1 × grad-accum 2 × 2-GPU DDP) → 270 optimizer steps
Instruct-subspace masking ON (r256 subspace + per-layer importance + logistic scale mask)
CCE (cut cross-entropy) loss; doc-isolation enforced via per-doc-reset position_ids (flash varlen)
Final train loss ≈ 2.7–3.0

Hardware / runtime

2× RTX PRO 6000 Blackwell (sm_120), torch 2.11.0+cu128, transformers 5.10.x, peft 0.19.1
Sliding-window layers routed to a Blackwell flash-attn build; global layers on SDPA (~2.4× vs all-SDPA)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("<gemma-4-12B-it path>", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "rpDungeon/marvin-cpt-12b-adapter")
tok = AutoTokenizer.from_pretrained("rpDungeon/marvin-cpt-12b-adapter")

A chat_template.jinja and the tokenizer are bundled with the adapter.

Framework versions

PEFT 0.19.1

Downloads last month: 1