Research Papers 论文研究 7h ago Updated 3h ago 更新于 3小时前 45

Post-Training Ternarization of Qwen3-4B Capability, Effective Bit Budget, Storage Compression, and Deployment Qwen3-4B的后训练三值化:能力、有效位预算、存储压缩与部署

Post-training ternarization of Qwen3-4B achieves 1.641 effective bits per weight, compressing model size from 8.29 GiB to 3.96 GiB while retaining 81.62% of parameters Task capability degrades unevenly: BoolQ retains 84.6% of teacher performance, while ARC-Challenge drops to 43.8%, with overall accuracy falling from 64.5% to 54.7% Perplexity increases significantly across benchmarks (WikiText-2: 13.639→18.748, PTB: 24.700→31.992, C4: 19.81→28.966), indicating notable quality loss A preliminary T 对Qwen3-4B模型进行后训练三元化,有效位预算降至1.641 bits/weight,覆盖81.62%参数 模型能力显著下降:10项基准测试准确率从64.5%降至54.7%,任务退化不均(BoolQ保留84.6%,ARC-Challenge仅43.8%) 存储压缩效果明显:模型体积从8.29 GiB压缩至3.96 GiB,困惑度基本不变 推理性能未改善:Triton GEMV微基准比FP16 cuBLAS慢4.6倍,压缩本身不带来推理加速

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Post-training ternarization of Qwen3-4B achieves 1.641 effective bits per weight, compressing model size from 8.29 GiB to 3.96 GiB while retaining 81.62% of parameters
  • Task capability degrades unevenly: BoolQ retains 84.6% of teacher performance, while ARC-Challenge drops to 43.8%, with overall accuracy falling from 64.5% to 54.7%
  • Perplexity increases significantly across benchmarks (WikiText-2: 13.639→18.748, PTB: 24.700→31.992, C4: 19.81→28.966), indicating notable quality loss
  • A preliminary Triton GEMV microbenchmark shows 4.6x slower inference compared to FP16 cuBLAS, meaning compression alone does not yield faster inference
  • The study demonstrates that nominal bit labels (e.g., "1.58-bit") are insufficient to capture effective bit accounting, capability retention, and runtime behavior

Why It Matters

This research provides a critical reality check for the ultra-low-bit quantization community by demonstrating that aggressive ternarization of a 4B-parameter instruction-tuned model incurs substantial and uneven capability degradation. For practitioners considering extreme compression for edge deployment, the findings highlight that storage savings come at the cost of both performance and inference speed, challenging the assumption that quantization alone can enable efficient deployment without complementary optimizations.

Technical Details

  • Methodology: End-to-end post-training conversion using KOTMS rotation, E2M-ATQ ternarization, and GPTQ-style error compensation from TWLA; weight-only quantization with activations retained at 16-bit precision
  • Effective bit accounting: 1.641 effective bits per quantized linear weight, with 81.62% of model parameters targeted for ternarization
  • Benchmark results: Evaluated across ten scored capability comparisons and three perplexity datasets (WikiText-2, PTB, C4); degradation is highly task-dependent, with reasoning-heavy benchmarks suffering more than factual recall tasks
  • Packing optimization: A subsequent packing run preserving ternary planes and scales reduced model size by ~52% (8.29 GiB → 3.96 GiB) with negligible perplexity impact; a third-party lossy packing attempt was excluded
  • Deployment benchmarking: Preliminary Triton GEMV microbenchmark shows 4.6x slowdown versus FP16 cuBLAS on tested shapes; end-to-end task accuracy and generation throughput for the packed artifact remain unbenchmarked

Industry Insight

  • The uneven degradation pattern suggests that ultra-low-bit quantization may be viable for narrow, fact-retrieval-heavy applications but remains risky for reasoning-intensive tasks; practitioners should benchmark per-task before deploying ternarized models in production
  • Compression-only strategies are insufficient for inference acceleration at extreme bit rates; hardware-aware kernel optimization (e.g., custom Triton kernels) must advance in parallel with quantization research to realize end-to-end efficiency gains
  • The gap between the packed artifact's storage metrics and its unbenchmarked runtime performance underscores the need for standardized end-to-end evaluation pipelines that report both compression ratios and actual deployment throughput

TL;DR

  • 对Qwen3-4B模型进行后训练三元化,有效位预算降至1.641 bits/weight,覆盖81.62%参数
  • 模型能力显著下降:10项基准测试准确率从64.5%降至54.7%,任务退化不均(BoolQ保留84.6%,ARC-Challenge仅43.8%)
  • 存储压缩效果明显:模型体积从8.29 GiB压缩至3.96 GiB,困惑度基本不变
  • 推理性能未改善:Triton GEMV微基准比FP16 cuBLAS慢4.6倍,压缩本身不带来推理加速

为什么值得看

该研究揭示了超低比特量化中"名义位宽"与实际存储表示、能力保留和运行时行为之间的差距,为从业者提供了量化压缩与推理性能权衡的实证参考。

技术解析

  • 采用KOTMS旋转+E2M-ATQ三元化+GPTQ风格误差补偿(来自TWLA)的端到端后训练转换流程,激活值保持16-bit精度
  • 有效位计算显示1.641 bits/weight,但仅覆盖81.62%参数,名义"1.58-bit"标签无法完整描述实际存储表示
  • 困惑度评估:WikiText-2从13.639升至18.748,PTB从24.700升至31.992,C4从19.831升至28.966
  • 打包后存储从8.29 GiB降至3.96 GiB,但第三方打包尝试为有损且被排除

行业启示

  • 超低比特量化(<2 bits)在存储压缩上可行,但需警惕能力退化不均问题,不同任务对量化敏感度差异显著
  • 压缩与推理加速并非正相关,需配套专用硬件/算子优化才能实现端到端性能收益
  • 量化论文应报告有效位预算、参数覆盖率和任务级退化分布,而非仅依赖名义位宽标签

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Quantization 量化 LLM 大模型 Deployment 部署 Inference 推理 Research 科学研究