distilgpt2

Repackaging of distilbert/distilgpt2 with the model.safetensors keys stripped of the transformer. prefix, so it loads directly via mlx-lm on Apple silicon. The weights are bit-for-bit identical; only the names changed.

Usage

mlx-lm (Apple silicon)

from mlx_lm import load, generate

model, tokenizer = load("gabfssilva/distilgpt2")
print(generate(model, tokenizer, "Once upon a time", max_tokens=50))

PyTorch (CUDA / MPS / CPU / probably ROCm)

Also loads cleanly via transformers — from_pretrained tolerates the missing transformer. prefix, so the same weights run on CUDA, Apple Metal (MPS), or CPU without any extra step:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("gabfssilva/distilgpt2", device_map="cuda") # or "mps", "cpu", "auto"
tokenizer = AutoTokenizer.from_pretrained("gabfssilva/distilgpt2")

device_map requires pip install accelerate.

Source

• Base model: distilbert/distilgpt2 (Apache 2.0)
• Difference: keys renamed from transformer.h.* to h.* to match the sanitize() in mlx_lm/models/gpt2.py.

Base model

distilbert/distilgpt2

(1503)