Instructions to use ml-ryanlee/moe-32L-d512-1e18-a100 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ml-ryanlee/moe-32L-d512-1e18-a100 with Transformers:

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

pipe = pipeline("text-generation", model="ml-ryanlee/moe-32L-d512-1e18-a100", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ml-ryanlee/moe-32L-d512-1e18-a100", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ml-ryanlee/moe-32L-d512-1e18-a100 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ml-ryanlee/moe-32L-d512-1e18-a100"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ml-ryanlee/moe-32L-d512-1e18-a100",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/ml-ryanlee/moe-32L-d512-1e18-a100

SGLang

How to use ml-ryanlee/moe-32L-d512-1e18-a100 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 "ml-ryanlee/moe-32L-d512-1e18-a100" \
    --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": "ml-ryanlee/moe-32L-d512-1e18-a100",
		"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 "ml-ryanlee/moe-32L-d512-1e18-a100" \
        --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": "ml-ryanlee/moe-32L-d512-1e18-a100",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use ml-ryanlee/moe-32L-d512-1e18-a100 with Docker Model Runner:
```
docker model run hf.co/ml-ryanlee/moe-32L-d512-1e18-a100
```

moe-32L-d512-1e18-a100

32-layer compute-optimal checkpoint for Sparse Layers are Critical to Scaling Looped Language Models (arXiv:2605.09165).


architecture	`moe`
d_model	512
effective layers	32
experts	8 total, 2 active
compute budget	1e18 FLOPs
training hardware	A100-80GB
training steps	65,960
parameters (stored)	361,972,736
peak LR	0.005
batch size	16
muP width_ratio	2.0 (d_base=256)
final val loss	n/a

This width is the architecture's own measured A100 minimum on the 1e18 isoFLOP sweep. Architectures optimise at different widths; at fixed compute a wider model simply trains on fewer tokens (C = 6·N_act·D), so this is the best that architecture does with the budget. Do not compare these against the B200 repos (base-32L-d512-1e18, looped-16x2-d640-1e18, looped-moe-16x2-d512-1e18) — same widths, different hardware.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
    "ml-ryanlee/moe-32L-d512-1e18-a100", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("gpt2")

Pass max_length=1024 when evaluating — the RoPE buffer is sized to the training context.

Downloads last month: 53

Safetensors

Model size

0.4B params

Tensor type

F32

Paper for ml-ryanlee/moe-32L-d512-1e18-a100

Sparse Layers are Critical to Scaling Looped Language Models

Paper • 2605.09165 • Published May 9