`包裹，长文用`
`/`
`划分小节（本文较短，可能不需要或仅用一个小节标题，或者直接段落）。 3. 列表用`
- `，引用用`
`，代码用`
`，重要句子可用``。
4. 保留原文所有信息，不得增删改内容。
5. 不要输出``/``/``等文档级标签。
* 输入文本：一段关于MINE (Memory-efficient INference Engine) 的学术/技术新闻摘要。
2. 分析输入文本结构：
* "We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as \texttt{BF16} and under additional assumptions." -> 第一段，介绍背景和提出的方法。
* "Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments." -> 第二段，解释问题背景（单流解码的内存瓶颈）。
* "At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference." -> 第三段，介绍核心技术。
* "It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost." -> 第四段，技术细节。
* "We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency." -> 第五段，优化方法。
* "MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget." -> 第六段，系统集成。
* "On Llama~3.1~8B~Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX~4060~Ti and an RTX~5090, at effective bitrates of 2.24 and 3.43~bits per weight." -> 第七段，实验结果。
* "MINE is available as open source." -> 第八段，开源声明。
3. 应用排版规则：
* *段落划分*：原文是一整段，或者可以理解为多个逻辑段落。为了阅读体验，我可以将其拆分为几个逻辑段落，或者保留原来的长段落但用``包裹。考虑到学术摘要通常作为一个大段落，但我可以按语义拆成几个句子组合的段落，或者直接用一个`
`。不过规则说“每个自然段落用`
`包裹”，原文是一整段。我最好按语义拆分成几个自然段，或者如果原文没有换行，就保持一个`
`。为了美观，我将其拆分为几个逻辑相关的段落。
* Para 1: 介绍 (We present... assumptions.)
* Para 2: 动机 (Single-stream... deployments.)
* Para 3: 方法与核心 (At the core... decoding cost.)
* Para 4: 优化与集成 (We auto-tune... memory budget.)
* Para 5: 实验结果 (On Llama... per weight.)
* Para 6: 开源 (MINE is available as open source.)
* *特殊标记*：
* `\texttt{BF16}` -> `BF16`
* `MINE`, `dtANS-LLM`, `vLLM` 等可以用 `` 强调。
* `2.52$\times$ and 1.48$\times$` -> 可以保留LaTeX或者转为纯文本/HTML，由于规则说“不得增删改内容”，最好保留原样或稍微HTML化数学符号（如`×`），但原文本里是`$\times$`。为安全起见，保留`$\times$`或者转义为HTML实体，这里保留原文字符串最稳妥：`2.52$\times$`。
* `Llama~3.1~8B~Instruct` -> 这里的 `~` 是LaTeX的不换行空格，可以转为 ` ` 或普通空格。为了严格遵守“不得增删改内容”，我直接保留原文的字符，或者把 `~` 替换为 HTML 实体 ` `。这里用 ` ` 更符合排版工程师的身份。
* `RTX~4060~Ti` -> `RTX 4060 Ti`
* `RTX~5090` -> `RTX 5090`
* *小节标题*：原文没有明确的标题，我可以加一个 `
` 来概括，或者不加。规则说“长文用/
划分小节”，这篇不算特别长，但分成几个逻辑块加小标题会更好看。或者直接输出段落。我不加文本内容不存在的标题，以免“增改内容”。那就只用 `
`。 * 让我们细化HTML结构： ```html
We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as BF16 and under additional assumptions.
Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments.
At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference. It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost.
We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency. MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget.
On Llama 3.1 8B Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX 4060 Ti and an RTX 5090, at effective bitrates of 2.24 and 3.43 bits per weight.
MINE is available as open source.
```
* 检查约束条件：
1. 只输出HTML片段本身，不要Markdown围栏、不要解释。 (Check)
2. 每个自然段落用``包裹，长文用`
`，代码用``，重要句子可用``。 4. 保留原文所有信息，不得增删改内容。 5. 不要输出``/``/``等文档级标签。 * 输入文本：一段关于MINE (Memory-efficient INference Engine) 的学术/技术新闻摘要。 2. 分析输入文本结构： * "We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as \texttt{BF16} and under additional assumptions." -> 第一段，介绍背景和提出的方法。 * "Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments." -> 第二段，解释问题背景（单流解码的内存瓶颈）。 * "At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference." -> 第三段，介绍核心技术。 * "It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost." -> 第四段，技术细节。 * "We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency." -> 第五段，优化方法。 * "MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget." -> 第六段，系统集成。 * "On Llama~3.1~8B~Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX~4060~Ti and an RTX~5090, at effective bitrates of 2.24 and 3.43~bits per weight." -> 第七段，实验结果。 * "MINE is available as open source." -> 第八段，开源声明。 3. 应用排版规则： * *段落划分*：原文是一整段，或者可以理解为多个逻辑段落。为了阅读体验，我可以将其拆分为几个逻辑段落，或者保留原来的长段落但用``包裹。考虑到学术摘要通常作为一个大段落，但我可以按语义拆成几个句子组合的段落，或者直接用一个`
`。不过规则说“每个自然段落用`
`包裹”，原文是一整段。我最好按语义拆分成几个自然段，或者如果原文没有换行，就保持一个`
`。为了美观，我将其拆分为几个逻辑相关的段落。 * Para 1: 介绍 (We present... assumptions.) * Para 2: 动机 (Single-stream... deployments.) * Para 3: 方法与核心 (At the core... decoding cost.) * Para 4: 优化与集成 (We auto-tune... memory budget.) * Para 5: 实验结果 (On Llama... per weight.) * Para 6: 开源 (MINE is available as open source.) * *特殊标记*： * `\texttt{BF16}` -> `
BF16` * `MINE`, `dtANS-LLM`, `vLLM` 等可以用 `` 强调。 * `2.52$\times$ and 1.48$\times$` -> 可以保留LaTeX或者转为纯文本/HTML，由于规则说“不得增删改内容”，最好保留原样或稍微HTML化数学符号（如`×`），但原文本里是`$\times$`。为安全起见，保留`$\times$`或者转义为HTML实体，这里保留原文字符串最稳妥：`2.52$\times$`。 * `Llama~3.1~8B~Instruct` -> 这里的 `~` 是LaTeX的不换行空格，可以转为 ` ` 或普通空格。为了严格遵守“不得增删改内容”，我直接保留原文的字符，或者把 `~` 替换为 HTML 实体 ` `。这里用 ` ` 更符合排版工程师的身份。 * `RTX~4060~Ti` -> `RTX 4060 Ti` * `RTX~5090` -> `RTX 5090` * *小节标题*：原文没有明确的标题，我可以加一个 `` 来概括，或者不加。规则说“长文用
/
划分小节”，这篇不算特别长，但分成几个逻辑块加小标题会更好看。或者直接输出段落。我不加文本内容不存在的标题，以免“增改内容”。那就只用 `
`。 * 让我们细化HTML结构： ```html
We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as
BF16and under additional assumptions.Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments.
At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference. It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost.
We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency. MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget.
On Llama 3.1 8B Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX 4060 Ti and an RTX 5090, at effective bitrates of 2.24 and 3.43 bits per weight.
MINE is available as open source.
``` * 检查约束条件： 1. 只输出HTML片段本身，不要Markdown围栏、不要解释。 (Check) 2. 每个自然段落用``包裹，长文用`
`/`
`划分小节。 (没有长文小节，直接用`
`) 3. 列表内容用`
- `，引用用`
`，代码用`
`，重要句子可用``。 (用到了``和``。原文里的`\texttt{BF