Skip to content

eval unique conc #120

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
adibarra wants to merge 5 commits into from
Closed

eval unique conc #120

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
92a891e
fix: widen eval_results unique constraint to include isl, osl, conc
adibarra Mar 29, 2026
7d42782
fix: add p90 and p99.9 percentile metrics to known keys
adibarra Mar 29, 2026
8eb8461
fix: filter evals by exact date and dedup by hw-framework-spec-precision
adibarra Mar 29, 2026
7f7b0dc
fix: update grid line extents on re-render when chart dimensions change
adibarra Mar 29, 2026
04575bf
feat: horizontal eval bar chart with labels on y-axis and dynamic height
adibarra Mar 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
72 changes: 34 additions & 38 deletions packages/app/src/components/evaluation/EvaluationContext.tsx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,13 +167,13 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
const dbModelKey = DISPLAY_MODEL_TO_DB[selectedModel];
if (!dbModelKey) return [];

const filteredData = rawData
// Map all rows up to selected date
const allData = rawData
.filter((item) => {
const itemDate = item.date;
return (
item.task === selectedBenchmark &&
item.model === dbModelKey &&
itemDate <= selectedRunDate
item.date <= selectedRunDate
);
})
.map((item): EvaluationChartData | null => {
Expand All @@ -187,9 +187,6 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
) as keyof typeof HARDWARE_CONFIG;
if (hwKey === 'unknown') return null;

const itemDate = item.date;
const itemDateTime = item.timestamp ?? '';

return {
configId: item.config_id,
hwKey,
Expand All @@ -199,8 +196,8 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
model: item.model,
benchmark: item.task,
specDecode: item.spec_method,
date: itemDate,
datetime: itemDateTime,
date: item.date,
datetime: item.timestamp ?? '',
precision: item.precision,
framework: item.framework,
tp: item.decode_tp,
Expand All @@ -211,39 +208,38 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
})
.filter((item): item is EvaluationChartData => item !== null);

// Group by config_id and keep most recent per TP+conc combination
const groupMap = new Map<number, EvaluationChartData[]>();
filteredData.forEach((item) => {
if (!groupMap.has(item.configId)) groupMap.set(item.configId, []);
groupMap.get(item.configId)!.push(item);
});
// Group by hw-framework-specmethod-precision. Find the latest date for each
// group (up to selectedRunDate), then keep ALL rows from that date for the
// group. This means if parallelism (conc/tp/ep) changes between runs, we
// only show the latest run's parallelism — old combos don't leak in.
const groupKeyFn = (item: EvaluationChartData) =>
`${item.hwKey}_${item.framework}_${item.specDecode}_${item.precision}`;

const latestDateForGroup = new Map<string, string>();
for (const item of allData) {
const key = groupKeyFn(item);
const existing = latestDateForGroup.get(key);
if (!existing || item.date > existing) latestDateForGroup.set(key, item.date);
}

const result: EvaluationChartData[] = [];
groupMap.forEach((groupItems) => {
// Dedup by TP+conc, keeping most recent date for each combination
const dedupMap = new Map<string, EvaluationChartData>();
groupItems.forEach((item) => {
const key = `${item.tp}_${item.conc}`;
const existing = dedupMap.get(key);
if (!existing || item.date > existing.date) dedupMap.set(key, item);
});

dedupMap.forEach((item) => {
const hwConfig = HARDWARE_CONFIG[item.hwKey];
const hwLabel = String(hwConfig?.label || item.hwKey);
result.push({
...item,
configLabel: buildConfigLabel(
hwLabel,
item.framework,
item.specDecode,
item.precision,
item.conc,
item.tp,
),
});
for (const item of allData) {
const key = groupKeyFn(item);
if (item.date !== latestDateForGroup.get(key)) continue;
const hwConfig = HARDWARE_CONFIG[item.hwKey];
const hwLabel = String(hwConfig?.label || item.hwKey);
result.push({
...item,
configLabel: buildConfigLabel(
hwLabel,
item.framework,
item.specDecode,
item.precision,
item.conc,
item.tp,
),
});
});
}

return result.sort((a, b) => a.configLabel.localeCompare(b.configLabel));
}, [rawData, selectedBenchmark, selectedModel, selectedRunDate]);
Expand Down
Loading
Loading