95 lines
3.2 KiB
Markdown
95 lines
3.2 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
tags:
|
|
- gpt2
|
|
- scaling-study
|
|
- benchmarking
|
|
- banterhearts
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
license: mit
|
|
---
|
|
|
|
# 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](https://github.com/Sahil170595/Banterhearts) 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
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("Crusadersk/gpt2-25m")
|
|
tokenizer = AutoTokenizer.from_pretrained("Crusadersk/gpt2-25m")
|
|
|
|
inputs = tokenizer("Hello", return_tensors="pt")
|
|
outputs = model.generate(**inputs, max_new_tokens=32, do_sample=False)
|
|
print(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
|
|
|
|
```bibtex
|
|
@misc{banterhearts2026gpt225m,
|
|
title = {Custom GPT-2 Scaling Checkpoint (25M) for Inference Benchmarking Research},
|
|
author = {Kadadekar, Sahil},
|
|
year = {2026},
|
|
url = {https://huggingface.co/Crusadersk/gpt2-25m},
|
|
note = {Part of the Banterhearts research program. NeurIPS 2026 submission.}
|
|
}
|
|
```
|
|
|
|
## Acknowledgments
|
|
|
|
This work is part of a 40-TR research program on consumer LLM deployment safety, conducted independently as pre-doctoral research. Full program details at [github.com/Sahil170595/Banterhearts](https://github.com/Sahil170595/Banterhearts).
|