# E21v: KV cache type vs model quality (KL divergence vs f16 KV), Qwen3.5-4B, WikiText-2 test
# Edit the paths below for your machine.
$ErrorActionPreference='Continue'
$log='e21v_chain.log'
function L($m){ "$(Get-Date -Format o)  $m" | Out-File -FilePath $log -Append -Encoding utf8 }
L "=== e21v start ==="
$out='e21v'; New-Item -ItemType Directory -Force $out | Out-Null
$bin='llama.cpp\build-cuda\bin\llama-perplexity.exe'   # llama.cpp with upstream PR 29827
$m4='Qwen3.5-4B-Q4_K_M.gguf'; $txt='wiki_test.txt'     # WikiText-2 raw test split
foreach($ctx in @(16384)){
 $base="$out\kld_base_c$ctx.bin"
 & $bin -m $m4 -f $txt -c $ctx --chunks 8 -ngl 99 -fa on -ctk f16 -ctv f16 -b 2048 -ub 512 --kl-divergence-base $base *> "$out\f16_c$ctx.log"
 L "f16 base c$ctx exit $LASTEXITCODE"
 foreach($kv in @(@('q8_0','q8_0'),@('q5_0','q5_0'),@('q4_1','q4_1'),@('iq4_nl','iq4_nl'),@('q4_0','q4_0'),@('q8_0','q4_0'),@('q4_0','q8_0'),@('f16','f16'))){
  $tag="$($kv[0])_$($kv[1])_c$ctx"
  & $bin -m $m4 -f $txt -c $ctx --chunks 8 -ngl 99 -fa on -ctk $kv[0] -ctv $kv[1] -b 2048 -ub 512 --kl-divergence-base $base --kl-divergence *> "$out\$tag.log"
  $sum=(Select-String -Path "$out\$tag.log" -Pattern 'Mean PPL|Mean    KLD|Same top|RMS' | % { $_.Line.Trim() }) -join ' | '
  L "result $tag exit $LASTEXITCODE :: $sum"
 }
}
L "=== E21vDONE ==="