GPT-2 25M
Custom-trained GPT-2 checkpoint with deliberate depth-width configuration for inference benchmarking research.
Created as part of the
Banterhearts research program investigating benchmarking integrity for local LLM inference.
| |
|---|
| Architecture | GPT2LMHeadModel (MHA) |
| Parameters | 25M |
| Config | n_embd=384, n_head=2, n_layer=3, n_inner=1536 |
| Context length | 1,024 tokens |
| Precision | FP32 |
| Model size | 96 MB |
| Vocab size | 50,257 |
Purpose
Small-scale MHA baseline for depth-width trade-off analysis.
These checkpoints are not general-purpose language models. They are deliberately sized scaling-study artifacts designed to isolate the effect of model depth vs width on GPU inference latency. The key finding: in the small-model GPU regime, layer depth (not parameter count) dominates latency, producing inversions where a 5M-parameter model can be 3.6x slower than a 25M-parameter model.
Source Technical Reports
Used in: TR117, TR126, TR147
| TR | Role |
|---|
| TR117 | Original cross-backend benchmark matrix (7 backends, 4 model groups) |
| TR126 | Linux/Triton compiler validation with phase-separated measurement |
| TR147 | Second-regime portability validation on RTX 6000 Ada |
Design Rationale
The GPT-2 family (25M, 50M, 100M) uses a 2x3 factorial design:
| Model | n_embd | n_layer | n_inner | Params | Design role |
|---|
| gpt2-25m | 384 | 3 | 1,536 | 25M | Shallow, narrow |
| gpt2-50m | 512 | 8 | 2,048 | 50M | Deep, medium width |
| gpt2-100m | 768 | 8 | 3,072 | 100M | Deep, wide |
All models use 2 attention heads (MHA, not GQA) to isolate architecture effects from attention-group structure. Dropout is set to 0.0 for deterministic inference measurement.
Usage
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model = AutoModelForCausalLM.from_pretrained("Crusadersk/gpt2-25m")
4tokenizer = AutoTokenizer.from_pretrained("Crusadersk/gpt2-25m")
5
6inputs = tokenizer("Hello", return_tensors="pt")
7outputs = model.generate(**inputs, max_new_tokens=32, do_sample=False)
8print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Compatibility
| Framework | Supported |
|---|
| Transformers | Yes |
| torch.compile (Inductor) | Yes |
| Ollama | No (not GGUF format) |
| vLLM | Yes |
Citation
1@misc{banterhearts2026gpt225m,
2 title = {Custom GPT-2 Scaling Checkpoint (25M) for Inference Benchmarking Research},
3 author = {Kadadekar, Sahil},
4 year = {2026},
5 url = {https://huggingface.co/Crusadersk/gpt2-25m},
6 note = {Part of the Banterhearts research program. NeurIPS 2026 submission.}
7}
Acknowledgments
This work is part of the Chimera/Banterhearts technical-report program on deployment-time LLM behavior, quantization, refusal robustness, batching effects, and inference-stack reliability. Canonical public archive:
Chimeraforge Reports; source context:
github.com/Sahil170595/Banterhearts.
