dropbox/hqq

• highreadme#1

  Reposition the README's opening paragraph to highlight LLM focus and key differentiators

  原因：

  当前

  ## Half-Quadratic Quantization (HQQ)
  This repository contains the official implementation of Half-Quadratic Quantization (<b>HQQ</b>) presented in our articles: 
  * HQQ: https://dropbox.github.io/hqq_blog/
  * HQQ+: https://dropbox.github.io/1bit_blog/
  
  ### What is HQQ?
  <b>HQQ</b> is a fast and accurate model quantizer that skips the need for calibration data. Quantize the largest models, without calibration data, in just a few minutes at most 🚀.

  复制粘贴的修复

  复制

  ## Half-Quadratic Quantization (HQQ)
  This repository contains the official implementation of Half-Quadratic Quantization (<b>HQQ</b>), a fast and accurate post-training quantization method for large language models (LLMs) that uniquely skips the need for calibration data. Quantize the largest models, without calibration data, in just a few minutes at most 🚀. HQQ supports 8,4,3,2,1 bits and is compatible with PEFT training and `torch.compile` for faster inference and training. Learn more in our articles: 
  * HQQ: https://dropbox.github.io/hqq_blog/
  * HQQ+: https://dropbox.github.io/1bit_blog/

  待办完成忽略
• mediumtopics#2

  Add more specific topics to improve keyword matching for LLM quantization

  原因：

  当前

  llm, machine-learning, quantization

  复制粘贴的修复

  复制

  llm, machine-learning, quantization, post-training-quantization, low-bit-quantization, llm-inference, peft-compatible

  待办完成忽略
• lowreadme#3

  Move the 'Why use HQQ' comparison out of the FAQ details tag

  原因：

  当前

  <details>
    <summary>FAQ </summary>
   <b> Why should I use HQQ instead of other quantization methods? </b><br>
  <ul>
  <li> HQQ is very fast to quantize models.</li>
  <li> It supports 8,4,3,2,1 bits.</li>
  <li> You can use it on any model (LLMs, Vision, etc.).</li>
  <li> The dequantization step is a linear operation, this means that HQQ is compatbile with various optimized CUDA/Triton kernels.</li>
  <li> HQQ is compatible with peft training.</li>
  <li> We try to make HQQ fully compatible `torch.compile` for faster inference and training.</li>
  </ul>
    
    <b>What is the quality of the quantized models? </b><br>

  复制粘贴的修复

  复制

  ### Why use HQQ instead of other quantization methods?
  <ul>
  <li> HQQ is very fast to quantize models.</li>
  <li> It supports 8,4,3,2,1 bits.</li>
  <li> You can use it on any model (LLMs, Vision, etc.).</li>
  <li> The dequantization step is a linear operation, this means that HQQ is compatbile with various optimized CUDA/Triton kernels.</li>
  <li> HQQ is compatible with peft training.</li>
  <li> We try to make HQQ fully compatible `torch.compile` for faster inference and training.</li>
  </ul>
  
  <details>
    <summary>FAQ </summary>
    <b>What is the quality of the quantized models? </b><br>

  待办完成忽略

本次扫描解析到的品类 GEO 通道：google/gemini-2.5-flash, deepseek/deepseek-v4-flash

• 品类问题

  How to quickly quantize large language models without needing extensive calibration data?

  你：未被推荐
  AI 推荐顺序：

  1. GPTQ
  2. AutoGPTQ
  3. optimum
  4. AWQ
  5. LLM.int8()
  6. SmoothQuant
  7. OFT
  8. ZeroQuant

  AI 推荐了 8 个替代方案，却始终没点名 dropbox/hqq。这就是要补上的差距。

  查看 AI 完整回答
• 品类问题

  Looking for an LLM quantization library supporting low bit-widths and fine-tuning compatibility.

  你：未被推荐
  AI 推荐顺序：

  1. AWQ
  2. GPTQ
  3. bitsandbytes
  4. Hugging Face Optimum
  5. NVIDIA TensorRT-LLM

  AI 推荐了 5 个替代方案，却始终没点名 dropbox/hqq。这就是要补上的差距。

  查看 AI 完整回答

• Metadata completeness

  pass

• README presence

  pass

• Compared to common alternatives in this category, what is the core differentiator of dropbox/hqq?

  pass
  ✓AI 明确点名了 dropbox/hqq

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？

• If a team adopts dropbox/hqq in production, what risks or prerequisites should they evaluate first?

  pass
  ✓AI 明确点名了 dropbox/hqq

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？

• In one sentence, what problem does the repo dropbox/hqq solve, and who is the primary audience?

  pass
  ✓AI 明确点名了 dropbox/hqq

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？