Poistettu 1.5B Q4 -vaihtoehto: GGUF dequantisointi liian hidas WASM:ssa
1.5B Q4_K_M: ~33s/token (0.03 tok/s) — käyttökelvoton 0.5B F32: ~2.5s/token (0.4 tok/s) — käyttökelpoinen kpn load lataa nyt suoraan 0.5B:n ilman valintalistaa. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -2226,35 +2226,14 @@ Write the corrected code.`;
|
||||
}
|
||||
|
||||
if (sub === 'load') {
|
||||
const arg = parts[2];
|
||||
const btn = document.getElementById('agent-compute-btn');
|
||||
// Mallikatalogista valinta numerolla tai nimellä
|
||||
const loadModels = [
|
||||
{ id: '1', key: '05b', name: 'Qwen2.5-Coder:0.5B', size: '~990 MB', coderSize: '05b' },
|
||||
{ id: '2', key: '3b', name: 'Qwen2.5-Coder:1.5B Q4', size: '~1 GB', coderSize: '3b' },
|
||||
];
|
||||
if (!arg) {
|
||||
// Näytetään lista
|
||||
termLog(' Ladattavat mallit:', '#c9d1d9');
|
||||
for (const m of loadModels) {
|
||||
const active = (btn?.dataset.state === 'ready' && coderSize === m.coderSize) ? ' <span style="color:#3fb950">✓ ladattu</span>' : '';
|
||||
termLog(` <span style="color:#58a6ff">${m.id}</span> ${m.name} <span style="color:#8b949e">(${m.size})</span>${active}`);
|
||||
}
|
||||
termLog(' Käyttö: kpn load <numero>', '#8b949e');
|
||||
if (btn?.dataset.state === 'ready') {
|
||||
termLog(' ✓ Qwen2.5-Coder:0.5B on jo ladattu ja valmis', '#3fb950');
|
||||
return;
|
||||
}
|
||||
const selected = loadModels.find(m => m.id === arg || m.key === arg || m.coderSize === arg);
|
||||
if (!selected) {
|
||||
termLog(` Tuntematon malli "${esc(arg)}". Kokeile: kpn load`, '#f85149');
|
||||
return;
|
||||
}
|
||||
if (btn?.dataset.state === 'ready' && coderSize === selected.coderSize) {
|
||||
termLog(` ✓ ${selected.name} on jo ladattu ja valmis`, '#3fb950');
|
||||
return;
|
||||
}
|
||||
coderSize = selected.coderSize;
|
||||
coderSize = '05b';
|
||||
localStorage.setItem('kpn-coder-size', coderSize);
|
||||
termLog(` Alustetaan ${selected.name} (${selected.size})...`, '#d29922');
|
||||
termLog(' Alustetaan Qwen2.5-Coder:0.5B (~990 MB)...', '#d29922');
|
||||
if (btn) btn.click();
|
||||
else ensureCoderNode();
|
||||
return;
|
||||
|
||||
Reference in New Issue
Block a user