HackerReposA FIELD GUIDE TO OPEN SOURCE
ProSponsor the shelf♧ 0
← Back to top all-time
cenconq25
delta-compress-llm
C++2026–2026
cenconq25 / delta-compress-llm

delta-compress-llm

Proof of concept: Exploiting temporal coherence in LLM inference-- delta encoding for KV cache compression and weight-skip prediction. Achieves F16-quality KV cache at Q4_0 compression ratios with zero perplexity loss on llama.cpp.

2Distinct people
2Mentions
0Explicit recommenders
51GitHub stars
REPOSITORY AT A GLANCE

What to know before exploring

What it is

Proof of concept: Exploiting temporal coherence in LLM inference-- delta encoding for KV cache compression and weight-skip prediction. Achieves F16-quality KV cache at Q4_0 compression ratios with zero perplexity loss on llama.cpp.

GitHub

Where it is used

No documented product use has been verified yet.

What people are saying

No recommendation phrase was found in the indexed comments.

No classified warning is attached to this page. Read the discussions before treating silence as approval.

Also in these discussions

No other repository is repeatedly linked in the same discussions.

This week on Hacker News

Nov 7–Nov 14 UTC · Direct repository links and separately labeled related project stories. Indexed through 11/14/2026, 5:04:57 PM UTC.

No direct repository links were found in the indexed HN discussions over the past seven days. Older discussions are available below.

Mentions over time

HN mentions by hour2 dated mentions

Observed mentions: 2026-03-22 23:27:32 UTC to 2026-03-23 10:05:28 UTC.

012026-03-22 23:00:00 UTC to 2026-03-23 00:00:00 UTC (exclusive): 1 mentions; 1 submissions, 0 commentsMar 22, 23:002026-03-23 00:00:00 UTC to 2026-03-23 01:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 01:00:00 UTC to 2026-03-23 02:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 02:00:00 UTC to 2026-03-23 03:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 03:00:00 UTC to 2026-03-23 04:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 04:00:00 UTC to 2026-03-23 05:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 commentsMar 23, 04:002026-03-23 05:00:00 UTC to 2026-03-23 06:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 06:00:00 UTC to 2026-03-23 07:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 07:00:00 UTC to 2026-03-23 08:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 08:00:00 UTC to 2026-03-23 09:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 09:00:00 UTC to 2026-03-23 10:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 10:00:00 UTC to 2026-03-23 11:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 commentsMar 23, 10:00

Counts come from the repository’s dated source evidence. Index cutoff: 2026-11-02 08:16:59 UTC.

View mention counts and date ranges
UTC ranges include their start and exclude their end. Select a period to filter the source evidence.
PeriodStart (UTC)End (exclusive, UTC)MentionsSubmissionsComments
Mar 22, 23:002026-03-22 23:00:00 UTC2026-03-23 00:00:00 UTC110
Mar 23, 00:002026-03-23 00:00:00 UTC2026-03-23 01:00:00 UTC000
Mar 23, 01:002026-03-23 01:00:00 UTC2026-03-23 02:00:00 UTC000
Mar 23, 02:002026-03-23 02:00:00 UTC2026-03-23 03:00:00 UTC000
Mar 23, 03:002026-03-23 03:00:00 UTC2026-03-23 04:00:00 UTC000
Mar 23, 04:002026-03-23 04:00:00 UTC2026-03-23 05:00:00 UTC000
Mar 23, 05:002026-03-23 05:00:00 UTC2026-03-23 06:00:00 UTC000
Mar 23, 06:002026-03-23 06:00:00 UTC2026-03-23 07:00:00 UTC000
Mar 23, 07:002026-03-23 07:00:00 UTC2026-03-23 08:00:00 UTC000
Mar 23, 08:002026-03-23 08:00:00 UTC2026-03-23 09:00:00 UTC000
Mar 23, 09:002026-03-23 09:00:00 UTC2026-03-23 10:00:00 UTC000
Mar 23, 10:002026-03-23 10:00:00 UTC2026-03-23 11:00:00 UTC101

What Hacker News says

2 sources
"video compression" by analogy only, what this claims to actually do is delta encode the values in each token from the previous token. Interesting idea, but the results seem almost suspicious? even accounting for the extra bits used to store the 16-bit start value for each block - ~5% for k=64 The code does funky things, like the encoder updates the reference value for each encoded token, using the non-quantized value! [1] But the decoder just ignored all that. [2] how can this work? [1] https://github.com/cenconq25/delta-compress-llm/commit/f185f... [2] https://github.com/cenconq25/delta-compress-llm/commit/f185f...
In Apply video compression on KV cache to 10,000x less error at Q4 quant

MAKE YOURSELF AT HOME

A few shortcuts

/
Search the collection
?
Open this guide
Esc
Close dialogs and navigation

Your saved repositories and interests stay in this browser.