Files
ModelHub XC 502f007389 初始化项目,由ModelHub XC社区提供模型
Model: dystrio/Qwen2.5-7B-Instruct-sculpt-default
Source: Original Platform
2026-06-20 17:42:18 +08:00

149 lines
5.2 KiB
Markdown

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
- en
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
- dystrio
- sculpt
- pruned
- compressed
- efficient
- dense
- runtime-agnostic
- no-custom-kernels
- hf-drop-in
- drop-in-replacement
- smaller
- faster
- qwen
datasets:
- wikitext
model-index:
- name: Dystrio Sculpt (Qwen2.5-7B-Instruct Default)
results:
- task:
type: text-generation
dataset:
name: WikiText-103 (validation)
type: wikitext
metrics:
- name: perplexity
type: perplexity
value: 12.334
- name: ppl_ratio
type: ppl_ratio
value: 0.9896
---
# dystrio/Qwen2.5-7B-Instruct-sculpt-default
> **9% smaller, quality improved (0.9896x 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 [Qwen 2.5 7B Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct).
## Quick Start
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("dystrio/Qwen2.5-7B-Instruct-sculpt-default", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("dystrio/Qwen2.5-7B-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 [Qwen 2.5 7B Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-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** | 12.964976 GB (9% smaller) |
| **PPL ratio** | 0.9896 |
| **Chat prefill TPS** | 12352.7 (+7%) |
| **RAG TTFT p95** | 110.714 ms (-6%) |
| **Decode TPS** | 72.7 (flat) |
| **Parameters** | 6.96B |
## All Sculpt Tiers
| Tier | HuggingFace | Size | PPL Ratio | Use Case |
|------|-------------|------|-----------|----------|
| default | [dystrio/Qwen2.5-7B-Instruct-sculpt-default](https://huggingface.co/dystrio/Qwen2.5-7B-Instruct-sculpt-default) 👈 **this model** | 12.964976 GB | 0.9896 | Zero-regret: quality preserved, smaller footprint |
| production | [dystrio/Qwen2.5-7B-Instruct-sculpt-production](https://huggingface.co/dystrio/Qwen2.5-7B-Instruct-sculpt-production) | 10.596324 GB | 1.7591 | Practical savings with modest quality tradeoff |
| throughput | [dystrio/Qwen2.5-7B-Instruct-sculpt-throughput](https://huggingface.co/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](https://github.com/EleutherAI/lm-evaluation-harness) on A100-80GB, bf16, zero-shot.
| Benchmark | Baseline | This Model | Delta |
|-----------|:--------:|:----------:|:-----:|
| ARC-Challenge | 0.5282 | 0.4676 | -0.0606 |
| HellaSwag | 0.6204 | 0.5650 | -0.0554 |
| MMLU | 0.7176 | 0.6506 | -0.0670 |
| TruthfulQA MC2 | 0.6475 | 0.5457 | -0.1018 |