Instructions to use jwl50-stanford/gdn-grow-initf64-h256-20k-r1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jwl50-stanford/gdn-grow-initf64-h256-20k-r1 with Transformers:

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

pipe = pipeline("text-generation", model="jwl50-stanford/gdn-grow-initf64-h256-20k-r1")

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("jwl50-stanford/gdn-grow-initf64-h256-20k-r1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jwl50-stanford/gdn-grow-initf64-h256-20k-r1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jwl50-stanford/gdn-grow-initf64-h256-20k-r1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jwl50-stanford/gdn-grow-initf64-h256-20k-r1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/jwl50-stanford/gdn-grow-initf64-h256-20k-r1

SGLang

How to use jwl50-stanford/gdn-grow-initf64-h256-20k-r1 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 "jwl50-stanford/gdn-grow-initf64-h256-20k-r1" \
    --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": "jwl50-stanford/gdn-grow-initf64-h256-20k-r1",
		"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 "jwl50-stanford/gdn-grow-initf64-h256-20k-r1" \
        --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": "jwl50-stanford/gdn-grow-initf64-h256-20k-r1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use jwl50-stanford/gdn-grow-initf64-h256-20k-r1 with Docker Model Runner:
```
docker model run hf.co/jwl50-stanford/gdn-grow-initf64-h256-20k-r1
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

gdn-grow-initf64-h256-20k-r1

This bundle was generated from an mlp-mixer training checkpoint on 2026-05-16 21:09:31 UTC.

Training Run

W&B run: hazy-research/mlp-mixer/m05d14y26-gdn-grow-initf64-h256-20k-r1
W&B state: finished
W&B URL: https://wandb.ai/hazy-research/mlp-mixer/runs/m05d14y26-gdn-grow-initf64-h256-20k-r1
Hydra experiment: mlp_mixer/gdn_paper_120m_growing_initf64
Source checkpoint: /pscratch/sd/j/jwl50/online-mlps/artifacts/checkpoints/mlp-mixer-pile/m05d14y26-gdn-grow-initf64-h256-20k-r1/step_20000.ckpt
Original config: training_config.yaml

The original Lightning checkpoint is included at lightning/step_20000.ckpt.

Metrics

Metric	Value
`trainer/global_step`	`19999`
`train/ppl_step`	`6.517290941504979`
`train/ppl_epoch`	`12.71058381595014`
`train/loss_step`	`2.395888328552246`
`train/loss_epoch`	`2.54245662689209`
`train/num-tokens_step`	`10485760000`
`val/ppl`	`10.258418173375262`
`val/loss`	`2.3280985355377197`
`val/num-tokens`	`7667089408`

Architecture

Parameters: 173650464
Trainable parameters: 173650464
Hidden size: 768
Layers: 12
Heads: 12
Config head_dim: 128
Effective Q/K head dim: 256
Effective V head dim: 64
V expansion: 0.25
ShortConv: True
Growing Q/K mask: True
Growing max Q/K feature dim: 256
Initial active Q/K features: 64
Expansion tokens/features: 85 / 8

Loading

This is a custom mlp-mixer / FLA architecture, so use a compatible checkout of the training repo and install its dependencies before loading.

from train.models.modeling_gated_deltanet import GatedDeltaNetForCausalLM

model = GatedDeltaNetForCausalLM.from_pretrained("ORG/REPO")
model.eval()

For the repo's eval suite, point hf_path at ORG/REPO and use hf_checkpoint_file if you uploaded the Lightning checkpoint instead of loading the save_pretrained weights directly.

Downloads last month: -

Safetensors

Model size

0.2B params

Tensor type

F32