helm/ulysses
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Limit cookbook spacing change to advanced tab

Browse source
This commit is contained in:
pewdiepie-archdaemon 2026-06-27 21:05:07 +00:00
parent 7af3f15288
commit 28974ae787
2 changed files with 3 additions and 26 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
static/js/cookbookServe.js
View file

@ -513,10 +513,8 @@ function _estimateLlamaContextFit(model, fields, modelCtxMax, modelWeightsGb = 0
}
const raw = Math.floor(freeForKv / kvGbPerToken);
const rounded = Math.max(1024, Math.floor(raw / 1024) * 1024);
const gpuEstimateCtx = Math.min(modelMax, rounded);
let ctx = gpuEstimateCtx;
let ctx = Math.min(modelMax, rounded);
let reasonSuffix = '';
let unifiedSpill = false;
if (isUnifiedMode) {
// Unified memory is not just "GPU math with a slightly bigger VRAM number".
// llama.cpp can spill into system RAM, so a conservative pure-VRAM KV
@ -529,7 +527,6 @@ function _estimateLlamaContextFit(model, fields, modelCtxMax, modelWeightsGb = 0
const unifiedCtx = Math.min(modelMax, unifiedCap);
if (unifiedCtx > ctx) {
ctx = unifiedCtx;
unifiedSpill = true;
reasonSuffix = '; unified can spill into system RAM, slower than pure GPU';
}
const gpuUsableGb = Math.max(1, totalVramGb - Math.max(1.0, selectedCount * 0.6));
@ -548,8 +545,6 @@ function _estimateLlamaContextFit(model, fields, modelCtxMax, modelWeightsGb = 0
ctx,
modelGb,
kvGbPerToken,
gpuEstimateCtx,
unifiedSpill,
reason: `~${ctx.toLocaleString()} tokens fits llama.cpp KV (${freeForKv.toFixed(1)}G free ${isUnifiedMode ? 'unified' : 'VRAM'}${reasonSuffix})`,
};
}
@ -1660,31 +1655,22 @@ function _rerenderCachedModels() {
const ggufGb = _selectedGgufSizeGb(m, f.gguf_file);
fit = _estimateLlamaContextFit(m, f, panel._modelCtxMax, ggufGb || panel._modelWeightsGb, panel._fitSystem, panel._contextProfileData);
} else {
if (_ctxAutoNote) {
_ctxAutoNote.textContent = '';
_ctxAutoNote.classList.remove('gpu-fit', 'unified-spill');
}
if (_ctxAutoNote) _ctxAutoNote.textContent = '';
return;
}
if (!fit) {
if (_ctxAutoNote) {
_ctxAutoNote.textContent = '';
_ctxAutoNote.classList.remove('gpu-fit', 'unified-spill');
}
if (_ctxAutoNote) _ctxAutoNote.textContent = '';
return;
}
if ((fit.needsHardwareScan || fit.needsModelSize) && !fit.ctx) {
if (_ctxAutoNote) {
_ctxAutoNote.textContent = fit.reason;
_ctxAutoNote.title = fit.reason;
_ctxAutoNote.classList.remove('gpu-fit', 'unified-spill');
}
return;
}
if (_ctxAutoNote) {
_ctxAutoNote.textContent = `Auto ${fit.ctx.toLocaleString()} · ${fit.reason}`;
_ctxAutoNote.classList.toggle('gpu-fit', !!fit.ctx && !fit.unifiedSpill);
_ctxAutoNote.classList.toggle('unified-spill', !!fit.unifiedSpill);
const _llamaMemoryLabel = String(f.llama_mode || '').toLowerCase() === 'unified' || f.unified_mem
? 'unified system memory'
: 'selected GPU memory';
@ -1727,7 +1713,6 @@ function _rerenderCachedModels() {
if (_ctxAutoNote && data?.model_probe_error && (!ctxMax || !weightsGb)) {
_ctxAutoNote.textContent = data.model_probe_error;
_ctxAutoNote.title = data.model_probe_error;
_ctxAutoNote.classList.remove('gpu-fit', 'unified-spill');
}
return { ctxMax, weightsGb, data };
}

8
static/style.css
View file

@ -20203,14 +20203,6 @@ body.gallery-selecting .gallery-dl-btn,
overflow: hidden;
text-overflow: ellipsis;
}
.hwfit-auto-ctx-note.gpu-fit {
color: var(--green, var(--color-success, #50fa7b));
opacity: 0.82;
}
.hwfit-auto-ctx-note.unified-spill {
color: var(--yellow, var(--color-warning, #f0ad4e));
opacity: 0.9;
}
.hwfit-serve-row label select,
.hwfit-serve-row label input {
display: block;