Skip to content

Instantly share code, notes, and snippets.

@mverrilli
mverrilli / kv-ppl-results.md
Created April 29, 2026 01:39
KV Cache Quantization — WikiText-2 PPL sweep (llama3.2:3b, llama3.1:8b, qwen2.5:7b, qwen3.5:9b, gemma4:27b) on Tesla P40

KV Cache Quantization — WikiText-2 Perplexity Results

Setup: WikiText-2 test set (1,256,449 bytes), ctx=512, Tesla P40 (cc 6.1, 24 GB VRAM). Binary: dist/bin/perplexity from ollama-bugfix branch turboquant-qjl. Teacher-forced PPL: for each token position i, measures −log P(tokenᵢ | token₀…ᵢ₋₁), PPL = exp(mean NLL). Same methodology as KVQuant, KIVI, TurboQuant papers.


1. Per-model results

@mverrilli
mverrilli / dsdt.patch
Created February 5, 2020 08:57 — forked from mailhost/dsdt.patch
Dell Precision 5510 1.10.0 DSDT fixes
--- dsdt.orig 2019-06-10 16:09:40.945676250 +0200
+++ dsdt.dsl 2019-06-10 16:28:15.274723137 +0200
@@ -3905,11 +3905,13 @@
If (LEqual (PM6H, One))
{
CreateBitField (BUF0, \_SB.PCI0._Y0C._RW, ECRW) // _RW_: Read-Write Status
- Store (Zero, ECRW (If (PM0H)
- {
- CreateDWordField (BUF0, \_SB.PCI0._Y0D._LEN, F0LN) // _LEN: Length
- Store (Zero, F0LN) /* \_SB_.PCI0._CRS.F0LN */