初始化项目，由ModelHub XC社区提供模型

Model: dystrio/Qwen2.5-3B-Instruct-sculpt-throughput Source: Original Platform
2026-06-20 17:28:20 +08:00
commit c37eee5948
8 changed files with 358 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,150 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+language:
+  - en
+base_model: Qwen/Qwen2.5-3B-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 (Qwen2.5-3B-Instruct Throughput)
+    results:
+      - task:
+          type: text-generation
+        dataset:
+          name: WikiText-103 (validation)
+          type: wikitext
+        metrics:
+          - name: perplexity
+            type: perplexity
+            value: 22.8847
+          - name: ppl_ratio
+            type: ppl_ratio
+            value: 1.6342
+---
+
+# dystrio/Qwen2.5-3B-Instruct-sculpt-throughput
+
+> **19% smaller, +16% faster prefill, 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 **Throughput** tier of [Qwen2.5 3B Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct).
+
+## Quick Start
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("dystrio/Qwen2.5-3B-Instruct-sculpt-throughput", torch_dtype="bfloat16", device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("dystrio/Qwen2.5-3B-Instruct-sculpt-throughput")
+
+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 [Qwen2.5 3B Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) on A100 80GB, bf16:
+
+| Model | PPL | PPL Ratio | Weights (GB) | Chat Prefill TPS | RAG TTFT p95 (ms) | Decode TPS |
+|-------|-----|-----------|-------------|------------------|-------------------|------------|
+| **Baseline** | 14.0033 | 1.0 | 5.748009 | 22079.6 | 75.483 | 59.4 |
+| **sculpt-default** | 15.5137 | 1.1079 | 5.220665 | 23360.8 | 73.544 | 59.0 |
+| **sculpt-production** | 18.9373 | 1.3523 | 4.956993 | 24367.4 | 69.025 | 59.0 |
+| **sculpt-throughput** | 22.8847 | 1.6342 | 4.640587 | 25556.6 | 68.084 | 59.0 |
+| **sculpt-experimental** | 31.3266 | 2.2371 | 4.165977 | 26731.9 | 66.499 | 59.5 |
+
+### Key Metrics (this model)
+
+| Metric | Value |
+|--------|-------|
+| **Weights memory** | 4.640587 GB (19% smaller) |
+| **PPL ratio** | 1.6342 |
+| **Chat prefill TPS** | 25556.6 (+16%) |
+| **RAG TTFT p95** | 68.084 ms (-10%) |
+| **Decode TPS** | 59.0 (flat) |
+| **Parameters** | 2.49B |
+
+## All Sculpt Tiers
+
+| Tier | HuggingFace | Size | PPL Ratio | Use Case |
+|------|-------------|------|-----------|----------|
+| default | [dystrio/Qwen2.5-3B-Instruct-sculpt-default](https://huggingface.co/dystrio/Qwen2.5-3B-Instruct-sculpt-default) | 5.220665 GB | 1.1079 | Zero-regret: quality preserved, smaller footprint |
+| production | [dystrio/Qwen2.5-3B-Instruct-sculpt-production](https://huggingface.co/dystrio/Qwen2.5-3B-Instruct-sculpt-production) | 4.956993 GB | 1.3523 | Practical savings with modest quality tradeoff |
+| throughput | [dystrio/Qwen2.5-3B-Instruct-sculpt-throughput](https://huggingface.co/dystrio/Qwen2.5-3B-Instruct-sculpt-throughput) 👈 **this model** | 4.640587 GB | 1.6342 | Maximum usable compression for speed/edge |
+| experimental | [dystrio/Qwen2.5-3B-Instruct-sculpt-experimental](https://huggingface.co/dystrio/Qwen2.5-3B-Instruct-sculpt-experimental) | 4.165977 GB | 2.2371 | 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 | 0.4573 | 0.3294 | -0.1279 |
+| HellaSwag | 0.5635 | 0.4388 | -0.1247 |
+| MMLU | 0.6545 | 0.4838 | -0.1707 |
+| TruthfulQA MC2 | 0.5874 | 0.5045 | -0.0829 |