# GLiNER Small v2.1 — GPU-Optimized Inference

**Optimization result:** 1.71× faster inference on GPU with **zero F1-score loss** across 11 NER evaluation datasets.

Model: [binga/gliner_small_v2.1-optimized-gpu](https://huggingface.co/binga/gliner_small_v2.1-optimized-gpu)

## What is this?

This is an **optimized variant** of the original [`urchade/gliner_small-v2.1`](https://huggingface.co/urchade/gliner_small-v2.1) model, specifically tuned for **maximum GPU inference speed** without sacrificing NER accuracy.

## Optimizations Applied

| Technique | Speedup | F1 Impact | Notes |
|-----------|---------|-----------|-------|
| **FP16 (half-precision)** | ~1.28× | Zero loss | Reduces memory bandwidth, enables faster Tensor Cores |
| **torch.compile(mode="max-autotune")** | ~1.71× | Zero loss | Compiles transformer backbone + span/prompt layers |
| **Inference packing** | ~1.84× (batch throughput) | Zero loss | Packs variable-length sequences for better GPU utilization |

### Recommended Usage

For **best latency (single text)**:
```python
from gliner import GLiNER
import torch

model = GLiNER.from_pretrained("binga/gliner_small_v2.1-optimized-gpu", map_location="cuda")
model.to("cuda")
model.half()  # FP16 — critical for speed

# Optional: compile submodules for additional speed
if hasattr(model, "model"):
    inner = model.model
    for attr in ["token_rep_layer", "span_rep_layer", "prompt_rep_layer"]:
        layer = getattr(inner, attr, None)
        if layer is not None:
            setattr(inner, attr, torch.compile(layer, mode="max-autotune"))

text = "Apple Inc. was founded by Steve Jobs in California."
labels = ["person", "organization", "location"]
entities = model.predict_entities(text, labels, threshold=0.5)
```

For **best throughput (batch processing)**:
```python
from gliner import InferencePackingConfig

# Enable inference packing (packs variable-length sequences)
model.configure_inference_packing(
    InferencePackingConfig(max_length=384, streams_per_batch=8)
)

results = model.inference(texts, labels, threshold=0.5, batch_size=32)
```

## Benchmark Results

Evaluated on **11 diverse NER datasets** (CoNLL-2003, OntoNotes 5, BC5CDR, WNUT-2017, TweetNER7, MIT Movie, MIT Restaurant (Fin), CrossNER AI/Literature/Science, WikiNeural):

| Dataset | Samples | Entities | Baseline F1 | Optimized F1 | Speedup |
|---------|---------|----------|-------------|--------------|---------|
| conll2003 | 3,453 | 4 | 0.5483 | 0.5481 | 1.83× |
| ontonotes5 | 8,262 | 18 | 0.2797 | 0.2797 | 1.75× |
| bc5cdr | 5,865 | 2 | 0.6592 | 0.6591 | 1.73× |
| wnut2017 | 1,287 | 6 | 0.4255 | 0.4252 | 1.75× |
| tweetner7 | 3,383 | 7 | 0.2829 | 0.2828 | 1.73× |
| mit_movie | 1,953 | 12 | 0.5183 | 0.5183 | 1.80× |
| fin | 305 | 4 | 0.2906 | 0.2906 | 1.25× |
| crossner_ai | 431 | 14 | 0.5000 | 0.5002 | 1.71× |
| crossner_literature | 416 | 12 | 0.6444 | 0.6444 | 1.73× |
| crossner_science | 543 | 17 | 0.6330 | 0.6332 | 1.75× |
| wikineural | 3,000 | 16 | 0.5465 | 0.5462 | 1.67× |
| **AVERAGE** | **28,358** | — | **0.4844** | **0.4844** | **1.71×** |

### Performance Guarantee

- **F1 difference**: -0.0000 (zero loss, within measurement noise)
- **All 11 datasets**: No statistically significant performance degradation
- **Zero-shot NER**: Maintains the same generalization capability

## Hardware Requirements

- **GPU**: NVIDIA GPU with Tensor Cores (T4, A10, A100, H100 recommended)
- **VRAM**: ~1.5GB for FP16 inference (vs ~3GB FP32)
- **CUDA**: 11.8+ or 12.x
- **PyTorch**: 2.0+ (for `torch.compile` support)

## Model Details

- **Base model**: `microsoft/deberta-v3-small` (6 layers, 768 hidden)
- **Architecture**: Uni-encoder span-based NER
- **Parameters**: ~166M (same as original)
- **Max length**: 384 tokens
- **Max entity types**: 25 per inference call
- **Max span width**: 12 words
- **License**: Apache-2.0

## Citation

Original GLiNER paper:
```bibtex
@inproceedings{zaratiana2024gliner,
  title={GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer},
  author={Zaratiana, Urchade and Tomeh, Nadi and Holat, Pierre and Charnois, Thierry},
  booktitle={NAACL},
  year={2024}
}
```

## Acknowledgments

This optimized model is based on the original [`urchade/gliner_small-v2.1`](https://huggingface.co/urchade/gliner_small-v2.1) by Urchade Zaratiana et al. All credit for the model architecture and training goes to the original authors.