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 Production tier of Qwen 2.5 7B Instruct.

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("dystrio/Qwen2.5-7B-Instruct-sculpt-production", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("dystrio/Qwen2.5-7B-Instruct-sculpt-production")

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 Qwen 2.5 7B Instruct on A100 80GB, bf16:

Model	PPL	PPL Ratio	Weights (GB)	Chat Prefill TPS	RAG TTFT p95 (ms)	Decode TPS
Baseline	12.4633	1.0	14.185191	11510.6	117.869	71.1
sculpt-default	12.334	0.9896	12.964976	12352.7	110.714	72.7
sculpt-production	21.9239	1.7591	10.596324	14700.3	95.291	73.5
sculpt-throughput	23.2366	1.8644	9.950328	15386.6	91.914	73.3

Key Metrics (this model)

Metric	Value
Weights memory	10.596324 GB (25% smaller)
PPL ratio	1.7591
Chat prefill TPS	14700.3 (+28%)
RAG TTFT p95	95.291 ms (-19%)
Decode TPS	73.5 (flat)
Parameters	5.69B

All Sculpt Tiers

Tier	HuggingFace	Size	PPL Ratio	Use Case
default	dystrio/Qwen2.5-7B-Instruct-sculpt-default	12.964976 GB	0.9896	Zero-regret: quality preserved, smaller footprint
production	dystrio/Qwen2.5-7B-Instruct-sculpt-production 👈 this model	10.596324 GB	1.7591	Practical savings with modest quality tradeoff
throughput	dystrio/Qwen2.5-7B-Instruct-sculpt-throughput	9.950328 GB	1.8644	Maximum usable compression for speed/edge

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 on A100-80GB, bf16, zero-shot.

Benchmark	Baseline	This Model	Delta
ARC-Challenge	0.5282	0.3754	-0.1528
HellaSwag	0.6204	0.4462	-0.1742
MMLU	0.7176	0.4781	-0.2395
TruthfulQA MC2	0.6475	0.4458	-0.2017