`, divide long texts with `
`/`
`. 3. Use `
- ` for lists, `
` for quotes, `
` for code, `` for important sentences.
4. Retain all information from the original text, no adding, deleting, or altering content.
5. No ``/``/`` tags.
Looking at the input text:
"Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random. We trace this to the gate that scales each retrieved row. A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code. GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest. Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models."
This text is exactly one long paragraph. There is no list, no code, no blockquotes. I cannot invent any content. The text must remain exactly as it is.
Let's analyze the content to inject `` tags without altering words.
First sentence:
"Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points."
I can bold "conditions removal's cost".
`Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points.
`
Second sentence:
"Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random."
Bold the main finding.
`Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
`
Third sentence:
"We trace this to the gate that scales each retrieved row."
Keep it plain or bold part. Let's bold "gate".
`We trace this to the gate that scales each retrieved row.
`
Fourth sentence:
"A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code."
Bold key characteristics.
`A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code.
`
Fifth sentence:
"GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest."
Bold the core mechanism.
`GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest.
`
Sixth sentence:
"Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models."
Bold the metrics.
`Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models.
`
Let me check if there are any other rules I missed.
"每个自然段落用包裹，长文用
/
划分小节" - I'll add `
Research on Conditional Memory Layers and Quantization
` at the very beginning just to define the section, as it's a long text about research. Wait, "不得增删改内容" (must not add, delete, or alter content). Adding a title might be considered adding content. To be absolutely safe, I should NOT add `` or `
` since the text is only one paragraph and does not contain sub-headings. I will just wrap it in `
` tags as I did above.
Wait, is there any punctuation error in my HTML tags placement? Let's verify.
`
Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points.
`
Looks perfect. The `` tags wrap the exact text without modifying it.
Let me review the second sentence.
`Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
`
Wait, "refining" - I accidentally used a ligature or something? No, I should copy exactly "refining".
`Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
`
Yes, good catch.
Third sentence:
`We trace this to the gate that scales each retrieved row.
`
Fourth sentence:
`A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code.
`
Fifth sentence:
`GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest.
`
Sixth
` for quotes, `` for code, `` for important sentences. 4. Retain all information from the original text, no adding, deleting, or altering content. 5. No ``/``/`` tags. Looking at the input text: "Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random. We trace this to the gate that scales each retrieved row. A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code. GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest. Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models." This text is exactly one long paragraph. There is no list, no code, no blockquotes. I cannot invent any content. The text must remain exactly as it is. Let's analyze the content to inject `` tags without altering words. First sentence: "Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points." I can bold "conditions removal's cost". `Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points.
` Second sentence: "Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random." Bold the main finding. `Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
` Third sentence: "We trace this to the gate that scales each retrieved row." Keep it plain or bold part. Let's bold "gate". `We trace this to the gate that scales each retrieved row.
` Fourth sentence: "A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code." Bold key characteristics. `A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code.
` Fifth sentence: "GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest." Bold the core mechanism. `GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest.
` Sixth sentence: "Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models." Bold the metrics. `Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models.
` Let me check if there are any other rules I missed. "每个自然段落用包裹，长文用
/
划分小节" - I'll add `
Research on Conditional Memory Layers and Quantization
` at the very beginning just to define the section, as it's a long text about research. Wait, "不得增删改内容" (must not add, delete, or alter content). Adding a title might be considered adding content. To be absolutely safe, I should NOT add `` or `
` since the text is only one paragraph and does not contain sub-headings. I will just wrap it in `
` tags as I did above. Wait, is there any punctuation error in my HTML tags placement? Let's verify. `
Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points.
` Looks perfect. The `` tags wrap the exact text without modifying it. Let me review the second sentence. `Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
` Wait, "refining" - I accidentally used a ligature or something? No, I should copy exactly "refining". `Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random.
` Yes, good catch. Third sentence: `We trace this to the gate that scales each retrieved row.
` Fourth sentence: `A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code.
` Fifth sentence: `GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest.
` Sixth