---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
  - en
base_model: meta-llama/Llama-3.1-8B-Instruct
tags:
  - dystrio
  - sculpt
  - pruned
  - compressed
  - efficient
  - dense
  - runtime-agnostic
  - no-custom-kernels
  - hf-drop-in
  - drop-in-replacement
  - smaller
  - faster
  - llama
datasets:
  - wikitext
model-index:
  - name: Dystrio Sculpt (Llama-3.1-8B-Instruct Default)
    results:
      - task:
          type: text-generation
        dataset:
          name: WikiText-103 (validation)
          type: wikitext
        metrics:
          - name: perplexity
            type: perplexity
            value: 14.7778
          - name: ppl_ratio
            type: ppl_ratio
            value: 1.0641
---

# dystrio/Llama-3.1-8B-Instruct-sculpt-default

> **10% smaller, quality preserved (1.0641x PPL), drop-in replacement. No custom kernels. No runtime changes.**

Dystrio Sculpt structurally compresses transformer models, producing dense models that load with standard `transformers` — no custom code, no new ops, no deployment friction.

This is the **Default** tier of [Llama 3.1 8B Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct).

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("dystrio/Llama-3.1-8B-Instruct-sculpt-default", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("dystrio/Llama-3.1-8B-Instruct-sculpt-default")

inputs = tokenizer("The future of AI inference is", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Benchmark Results

All tiers compiled from [Llama 3.1 8B Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) on A100 80GB, bf16:

| Model | PPL | PPL Ratio | Weights (GB) | Chat Prefill TPS | RAG TTFT p95 (ms) | Decode TPS |
|-------|-----|-----------|-------------|------------------|-------------------|------------|
| **Baseline** | 13.8879 | 1.0 | 14.957527 | 10570.4 | 126.745 | 66.8 |
| **sculpt-default** | 14.7778 | 1.0641 | 13.457527 | 11418.6 | 116.957 | 65.5 |
| **sculpt-production** | 21.9236 | 1.5786 | 11.863777 | 12760.5 | 112.529 | 66.7 |
| **sculpt-throughput** | 27.7463 | 1.9979 | 11.020027 | 13408.6 | 104.086 | 67.5 |
| **sculpt-experimental** | 29.3853 | 2.1159 | 10.832527 | 13483.3 | 103.432 | 67.4 |

### Key Metrics (this model)

| Metric | Value |
|--------|-------|
| **Weights memory** | 13.457527 GB (10% smaller) |
| **PPL ratio** | 1.0641 |
| **Chat prefill TPS** | 11418.6 (+8%) |
| **RAG TTFT p95** | 116.957 ms (-8%) |
| **Decode TPS** | 65.5 (flat) |
| **Parameters** | 7.22B |

## All Sculpt Tiers

| Tier | HuggingFace | Size | PPL Ratio | Use Case |
|------|-------------|------|-----------|----------|
| default | [dystrio/Llama-3.1-8B-Instruct-sculpt-default](https://huggingface.co/dystrio/Llama-3.1-8B-Instruct-sculpt-default) 👈 **this model** | 13.457527 GB | 1.0641 | Zero-regret: quality preserved, smaller footprint |
| production | [dystrio/Llama-3.1-8B-Instruct-sculpt-production](https://huggingface.co/dystrio/Llama-3.1-8B-Instruct-sculpt-production) | 11.863777 GB | 1.5786 | Practical savings with modest quality tradeoff |
| throughput | [dystrio/Llama-3.1-8B-Instruct-sculpt-throughput](https://huggingface.co/dystrio/Llama-3.1-8B-Instruct-sculpt-throughput) | 11.020027 GB | 1.9979 | Maximum usable compression for speed/edge |
| experimental | [dystrio/Llama-3.1-8B-Instruct-sculpt-experimental](https://huggingface.co/dystrio/Llama-3.1-8B-Instruct-sculpt-experimental) | 10.832527 GB | 2.1159 | Boundary exploration, maximum structural compression |

## What is Dystrio Sculpt?

Dystrio Sculpt compiles transformer models into smaller, faster variants. Output models:

- Are **dense** (not sparse) — standard architecture, fewer parameters
- Load with **standard HuggingFace Transformers** — no custom code needed
- Require **no custom kernels** and **no runtime changes**
- Work as a one-step compile before deployment
- Stack with quantization (AWQ, GPTQ, GGUF) for compound savings

## Compatibility

- ✅ HuggingFace Transformers
- ✅ vLLM
- ✅ TGI (Text Generation Inference)
- ✅ llama.cpp / GGUF conversion
- ✅ AWQ / GPTQ quantization
- ✅ Any framework that loads standard safetensors

## Benchmark Environment

- **GPU**: NVIDIA A100-SXM4-80GB
- **dtype**: bf16
- **Torch**: 2.10.0+cu128
- **Transformers**: 5.3.0
- **Deterministic**: True
- Single-GPU, standard HuggingFace Transformers, no custom kernels.

## Metric Definitions

- **PPL ratio**: WikiText-103 perplexity relative to baseline. <1.0 = quality improved.
- **Prefill TPS**: Tokens per second during prompt encoding (higher = faster).
- **TTFT p95**: Time to first token at 95th percentile (lower = faster).
- **Decode TPS**: Tokens per second during generation (higher = faster).
- **Weights (GB)**: Model parameter memory (deterministic, runtime-independent).

## Citation

```bibtex
@misc{dystrio_sculpt_2026,
  title={Dystrio Sculpt: Structural Compilation for Transformer LLMs},
  author={Dystrio},
  year={2026},
  url={https://huggingface.co/dystrio}
}
```

## Downstream Benchmarks (lm-eval)

Evaluated with [lm-eval-harness](https://github.com/EleutherAI/lm-evaluation-harness) on A100-80GB, bf16, zero-shot.

| Benchmark | Baseline | This Model | Delta |
|-----------|:--------:|:----------:|:-----:|
| ARC-Challenge | 0.5358 | 0.4283 | -0.1075 |
| HellaSwag | 0.5977 | 0.5416 | -0.0561 |
| MMLU | 0.6844 | 0.5590 | -0.1254 |
| TruthfulQA MC2 | 0.5456 | 0.4824 | -0.0632 |