What it is. A runtime wrapper that keeps GPT-2’s fused attention kernels and swaps only QKV/PROJ linears to low-rank factors.
No weights are stored here: on load it fetches openai-community/gpt2 and patches it at runtime. Because this uses trust_remote_code, HF Inference API/Serverless is not supported; run in your own environment or notebook.

Headline results (A100, FP16, greedy): ~1.8–2.0x tokens/sec at ~0.8% perplexity delta on WikiText-2 (raw).

Quick start (pin a safe revision)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch as th

repo = "jackal79/gpt2-ibce-lowrank-192"
rev  = "a71015d1333757bedaaa2437333cd44415076218"  # pinned revision

tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True, revision=rev)
mdl = AutoModelForCausalLM.from_pretrained(
    repo, trust_remote_code=True, revision=rev, dtype=th.float16
).cuda().eval()

x = tok("Hello from IBCE low-rank!", return_tensors="pt").to("cuda")
y = mdl.generate(**x, max_new_tokens=64, do_sample=False)
print(tok.decode(y[0], skip_special_tokens=True))

## Licensing
This repository is released under an **Evaluation License** for non-production, non-hosted use.
Commercial production or hosted use requires a separate license.
Contact **jack@thestardrive.com**.

The full license text is in [`LICENSE.txt`](https://huggingface.co/jackal79/gpt2-ibce-lowrank-192/blob/main/LICENSE.txt).