Instructions to use evalstate/tiny-gpt-memorization-8m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use evalstate/tiny-gpt-memorization-8m with Transformers:

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

pipe = pipeline("text-generation", model="evalstate/tiny-gpt-memorization-8m")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("evalstate/tiny-gpt-memorization-8m")
model = AutoModelForCausalLM.from_pretrained("evalstate/tiny-gpt-memorization-8m", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use evalstate/tiny-gpt-memorization-8m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "evalstate/tiny-gpt-memorization-8m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "evalstate/tiny-gpt-memorization-8m",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/evalstate/tiny-gpt-memorization-8m

SGLang

How to use evalstate/tiny-gpt-memorization-8m 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 "evalstate/tiny-gpt-memorization-8m" \
    --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": "evalstate/tiny-gpt-memorization-8m",
		"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 "evalstate/tiny-gpt-memorization-8m" \
        --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": "evalstate/tiny-gpt-memorization-8m",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use evalstate/tiny-gpt-memorization-8m with Docker Model Runner:
```
docker model run hf.co/evalstate/tiny-gpt-memorization-8m
```

tiny-gpt-memorization-8m / README.md

evalstate HF Staff

upload 8m near-capacity checkpoint

6f7b07b verified 14 days ago

preview code

Raw

History Blame Contribute Delete

1.31 kB

	---
	language: en
	library_name: transformers
	tags:
	- memorization
	- capacity
	- gpt2
	- random-data
	license: mit
	---
	# Tiny GPT memorization checkpoint (8m, near-capacity / saturation boundary)

	From an exploratory tiny-scale replication of
	[How much do language models memorize?](https://huggingface.co/papers/2505.24832).

	- Architecture: GPT-2 (transformers), trained from scratch.
	- Parameters: 7,916,160 (non-embedding 7,098,624).
	- Vocab: 2048 data tokens (uniform random) + BOS = 2049 model vocab.
	- Sequence length: 64 (paper S=64).
	- Dataset: 44000 sequences, 2,816,000 data tokens,
	dataset entropy 30,976,000 bits (30.976 Mbits).
	- Trained 15698 steps, AdamW, bfloat16, lr 0.002, batch 512.
	- Result: train loss 5.8075 bits/tok, held loss 16.6023,
	memorized 14,622,163 bits = 1.847 bits/parameter.

	This is the near-capacity (saturation-boundary) run for this model size.
	Below- and above-capacity checkpoints for the same architecture are published as
	`state.pt` files in the results dataset `evalstate/tiny-memorization-results`.

	Load with:
	```python
	from transformers import GPT2LMHeadModel
	model = GPT2LMHeadModel.from_pretrained("evalstate/tiny-gpt-memorization-8m")
	```

	Findings are scoped as an exploratory tiny-scale check (three architectures),
	NOT a universal scaling law.