sync from abhijitramesh/webgpu-bench@def1735599

data/combined.json

@@ -1,366 +0,0 @@
-{
-  "meta": {
-    "machines": [
-      {
-        "slug": "apple-metal-3-4gb-darwin",
-        "cpus": "apple metal-3",
-        "platform": "darwin",
-        "arch": "arm",
-        "totalMemoryGB": 4,
-        "submittedAt": "2026-04-25T21:31:12.111Z",
-        "resultCount": 7,
-        "passCount": 7,
-        "llamaCppCommit": null,
-        "llamaCppDescribe": null,
-        "submitters": [
-          {
-            "name": "abhijitramesh",
-            "hubId": "5fe0b7319126717e4c356efb",
-            "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg",
-            "count": 3,
-            "latestAt": "2026-04-25T21:15:00.076Z"
-          },
-          {
-            "name": "abhijitramesh",
-            "hubId": null,
-            "avatarUrl": null,
-            "count": 1,
-            "latestAt": "2026-04-25T04:50:36.488Z"
-          }
-        ]
-      },
-      {
-        "slug": "apple-metal-3-8gb-darwin",
-        "cpus": "apple metal-3",
-        "platform": "darwin",
-        "arch": "arm",
-        "totalMemoryGB": 8,
-        "submittedAt": "2026-04-25T21:31:12.111Z",
-        "resultCount": 2,
-        "passCount": 2,
-        "llamaCppCommit": null,
-        "llamaCppDescribe": null,
-        "submitters": [
-          {
-            "name": "abhijitramesh",
-            "hubId": "5fe0b7319126717e4c356efb",
-            "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg",
-            "count": 1,
-            "latestAt": "2026-04-25T18:38:18.454Z"
-          }
-        ]
-      }
-    ],
-    "models": [
-      "Llama-3.2-1B-Instruct",
-      "Qwen3-0.6B",
-      "Qwen3.5-2B",
-      "gemma-3-270m-it"
-    ],
-    "browsers": [
-      "chromium-147"
-    ],
-    "generatedAt": "2026-04-25T21:31:12.161Z"
-  },
-  "results": [
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T03:44:12.854Z",
-      "browser": "chromium-147",
-      "model": "Llama-3.2-1B-Instruct",
-      "repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
-      "variant": "Q2_K",
-      "filename": "Llama-3.2-1B-Instruct-Q2_K.gguf",
-      "sizeMB": 553.96,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 108503.70000004768,
-      "prefill_tok_s": 59.07,
-      "decode_tok_s": 40.57,
-      "n_p_eval": 31,
-      "t_p_eval_ms": 526,
-      "n_eval": 127,
-      "t_eval_ms": 3125.7,
-      "consistency_rate": 0.9688,
-      "consistency_first_disagree": 43,
-      "cpu_baseline_prefill_tok_s": 3.93,
-      "cpu_baseline_decode_tok_s": 3.55,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": null,
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T03:47:08.023Z",
-      "browser": "chromium-147",
-      "model": "Llama-3.2-1B-Instruct",
-      "repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
-      "variant": "Q4_K_M",
-      "filename": "Llama-3.2-1B-Instruct-Q4_K_M.gguf",
-      "sizeMB": 770.28,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 174666.89999997616,
-      "prefill_tok_s": 83.51,
-      "decode_tok_s": 46.15,
-      "n_p_eval": 31,
-      "t_p_eval_ms": 350.5,
-      "n_eval": 127,
-      "t_eval_ms": 2820.9,
-      "consistency_rate": 1,
-      "consistency_first_disagree": -1,
-      "cpu_baseline_prefill_tok_s": 9.84,
-      "cpu_baseline_decode_tok_s": 6.95,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": null,
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T03:51:41.928Z",
-      "browser": "chromium-147",
-      "model": "Llama-3.2-1B-Instruct",
-      "repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
-      "variant": "Q8_0",
-      "filename": "Llama-3.2-1B-Instruct-Q8_0.gguf",
-      "sizeMB": 1259.88,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 273256.60000002384,
-      "prefill_tok_s": 110.2,
-      "decode_tok_s": 41.82,
-      "n_p_eval": 31,
-      "t_p_eval_ms": 267.1,
-      "n_eval": 127,
-      "t_eval_ms": 3016.4,
-      "consistency_rate": 0.9922,
-      "consistency_first_disagree": 109,
-      "cpu_baseline_prefill_tok_s": 8.98,
-      "cpu_baseline_decode_tok_s": 6.61,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": null,
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T04:50:36.488Z",
-      "browser": "chromium-147",
-      "model": "Qwen3.5-2B",
-      "repo": "unsloth/Qwen3.5-2B-GGUF",
-      "variant": "IQ4_XS",
-      "filename": "Qwen3.5-2B-IQ4_XS.gguf",
-      "sizeMB": 1118.66,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 137132.30000001192,
-      "prefill_tok_s": 61.42,
-      "decode_tok_s": 28.23,
-      "n_p_eval": 30,
-      "t_p_eval_ms": 1043.3,
-      "n_eval": 127,
-      "t_eval_ms": 4552.3,
-      "consistency_rate": 0.9922,
-      "consistency_first_disagree": 33,
-      "cpu_baseline_prefill_tok_s": 2.6,
-      "cpu_baseline_decode_tok_s": 2.21,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": {
-        "name": "abhijitramesh",
-        "hubId": null,
-        "avatarUrl": null
-      },
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T19:43:30.850Z",
-      "browser": "chromium-147",
-      "model": "gemma-3-270m-it",
-      "repo": "unsloth/gemma-3-270m-it-GGUF",
-      "variant": "UD-IQ2_XXS",
-      "filename": "gemma-3-270m-it-UD-IQ2_XXS.gguf",
-      "sizeMB": 171.76,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 44796.59999996424,
-      "prefill_tok_s": 169.85,
-      "decode_tok_s": 106.01,
-      "n_p_eval": 27,
-      "t_p_eval_ms": 470.3,
-      "n_eval": 127,
-      "t_eval_ms": 1194.8,
-      "consistency_rate": 0.9688,
-      "consistency_first_disagree": 23,
-      "cpu_baseline_prefill_tok_s": 23.36,
-      "cpu_baseline_decode_tok_s": 11.69,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": {
-        "name": "abhijitramesh",
-        "hubId": "5fe0b7319126717e4c356efb",
-        "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg"
-      },
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T19:43:30.850Z",
-      "browser": "chromium-147",
-      "model": "gemma-3-270m-it",
-      "repo": "unsloth/gemma-3-270m-it-GGUF",
-      "variant": "UD-IQ2_XXS",
-      "filename": "gemma-3-270m-it-UD-IQ2_XXS.gguf",
-      "sizeMB": 171.76,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 44796.59999996424,
-      "prefill_tok_s": 169.85,
-      "decode_tok_s": 106.01,
-      "n_p_eval": 27,
-      "t_p_eval_ms": 470.3,
-      "n_eval": 127,
-      "t_eval_ms": 1194.8,
-      "consistency_rate": 0.9688,
-      "consistency_first_disagree": 23,
-      "cpu_baseline_prefill_tok_s": 23.36,
-      "cpu_baseline_decode_tok_s": 11.69,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": {
-        "name": "abhijitramesh",
-        "hubId": "5fe0b7319126717e4c356efb",
-        "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg"
-      },
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-4gb-darwin",
-      "timestamp": "2026-04-25T21:15:00.076Z",
-      "browser": "chromium-147",
-      "model": "Qwen3-0.6B",
-      "repo": "unsloth/Qwen3-0.6B-GGUF",
-      "variant": "Q4_1",
-      "filename": "Qwen3-0.6B-Q4_1.gguf",
-      "sizeMB": 390.14,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 77587.89999997616,
-      "prefill_tok_s": 157.43,
-      "decode_tok_s": 73.45,
-      "n_p_eval": 30,
-      "t_p_eval_ms": 221.3,
-      "n_eval": 127,
-      "t_eval_ms": 1728.4,
-      "consistency_rate": 0.9688,
-      "consistency_first_disagree": 32,
-      "cpu_baseline_prefill_tok_s": 6.21,
-      "cpu_baseline_decode_tok_s": 5.23,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": {
-        "name": "abhijitramesh",
-        "hubId": "5fe0b7319126717e4c356efb",
-        "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg"
-      },
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-8gb-darwin",
-      "timestamp": "2026-04-24T07:21:54.498Z",
-      "browser": "chromium-147",
-      "model": "Llama-3.2-1B-Instruct",
-      "repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
-      "variant": "Q4_1",
-      "filename": "Llama-3.2-1B-Instruct-Q4_1.gguf",
-      "sizeMB": 793.21,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 121113.60000002384,
-      "prefill_tok_s": 141.19,
-      "decode_tok_s": 58.63,
-      "n_p_eval": 31,
-      "t_p_eval_ms": 410.4,
-      "n_eval": 127,
-      "t_eval_ms": 2188.4,
-      "consistency_rate": 0.9922,
-      "consistency_first_disagree": 100,
-      "cpu_baseline_prefill_tok_s": 3.04,
-      "cpu_baseline_decode_tok_s": 2.62,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": null,
-      "iterations": 5
-    },
-    {
-      "machineSlug": "apple-metal-3-8gb-darwin",
-      "timestamp": "2026-04-25T18:38:18.454Z",
-      "browser": "chromium-147",
-      "model": "Qwen3-0.6B",
-      "repo": "unsloth/Qwen3-0.6B-GGUF",
-      "variant": "BF16",
-      "filename": "Qwen3-0.6B-BF16.gguf",
-      "sizeMB": 1142.68,
-      "status": "done",
-      "error": null,
-      "buildType": "jspi",
-      "webgpuAvailable": true,
-      "nGpuLayers": 999,
-      "wallTimeMs": 562839.5,
-      "prefill_tok_s": 2.92,
-      "decode_tok_s": 1.84,
-      "n_p_eval": 30,
-      "t_p_eval_ms": 10280.8,
-      "n_eval": 127,
-      "t_eval_ms": 67368.9,
-      "consistency_rate": 1,
-      "consistency_first_disagree": -1,
-      "cpu_baseline_prefill_tok_s": 3.14,
-      "cpu_baseline_decode_tok_s": 2.24,
-      "llamaCppCommit": null,
-      "llamaCppDescribe": null,
-      "dawnTag": null,
-      "submittedBy": {
-        "name": "abhijitramesh",
-        "hubId": "5fe0b7319126717e4c356efb",
-        "avatarUrl": "https://cdn-avatars.huggingface.co/v1/production/uploads/5fe0b7319126717e4c356efb/iAaiUJpDmv1nBN3_ZNGVU.jpeg"
-      },
-      "iterations": 5
-    }
-  ]
-}

js/data.js

@@ -1,57 +1,71 @@
-import { fetchRecentRuns } from './dataset.js';
+import { fetchAllRuns } from './dataset.js';
 import { HF_DATASET_REPO } from './run/config.js';
 
+// In-memory cache for the current page session.
 let cachedData = null;
+// sessionStorage cache so a refresh-within-a-minute doesn't re-fetch the
+// entire dataset. Short TTL — submissions land continuously and the
+// dashboard is the surface where we actually want freshness.
+const SESSION_CACHE_KEY = 'webgpu-bench:dashboard-data';
+const SESSION_CACHE_TTL_MS = 60 * 1000;
 
 export async function loadData() {
   if (cachedData) return cachedData;
 
-  // Static baseline — pre-built by CI from the dataset on the last code push.
-  const baselineResp = await fetch('data/combined.json');
-  const baseline = await baselineResp.json();
+  const fromSession = readSessionCache();
+  if (fromSession) {
+    cachedData = fromSession;
+    return cachedData;
+  }
 
-  // Live delta — pull recent submissions from the HF dataset that landed
-  // since baseline.meta.generatedAt. Best-effort: a CORS / network / parse
-  // failure here just leaves the dashboard on the static baseline rather
-  // than blocking the whole page.
+  // Single source of truth: the HF dataset repo. No static baseline. A new
+  // dashboard with zero submissions shows an empty state until something is
+  // submitted.
+  const empty = makeEmptyDataset();
   try {
-    const since = baseline.meta?.generatedAt;
-    const { records, machines, fileCount } = await fetchRecentRuns(HF_DATASET_REPO, since);
+    const { records, machines, fileCount } = await fetchAllRuns(HF_DATASET_REPO);
     if (fileCount > 0) {
-      mergeLiveDelta(baseline, records, machines);
+      mergeRecords(empty, records, machines);
     }
+    cachedData = empty;
+    writeSessionCache(cachedData);
   } catch (err) {
-    console.warn(`Live dataset sync skipped: ${err.message}`);
+    console.warn(`Live dataset load failed: ${err.message}`);
+    cachedData = empty;
   }
-
-  cachedData = baseline;
   return cachedData;
 }
 
-/* Append fresh records into the baseline and recompute the meta lookups so
-   the dashboard's filters/charts see them as first-class data. Records that
-   already exist in the baseline (deduped by a composite key) are dropped —
-   if CI ran recently, the static combined.json already contains them. */
-function mergeLiveDelta(baseline, records, machines) {
-  const baselineKeys = new Set(baseline.results.map(keyOf));
-  const fresh = records.filter(r => !baselineKeys.has(keyOf(r)));
-  if (fresh.length === 0) return;
+function makeEmptyDataset() {
+  return {
+    meta: {
+      machines: [],
+      models: [],
+      browsers: [],
+      generatedAt: new Date().toISOString(),
+    },
+    results: [],
+  };
+}
+
+/* Append records into an empty payload and recompute the meta lookups. Same
+   shape the old combined.json had, so all downstream consumers (charts,
+   tables, machine cards) work unchanged. */
+function mergeRecords(payload, records, machines) {
+  if (records.length === 0) return;
 
-  baseline.results.push(...fresh);
+  payload.results.push(...records);
 
-  // meta.models / meta.browsers union.
-  const modelsSet = new Set(baseline.meta.models || []);
-  const browsersSet = new Set(baseline.meta.browsers || []);
-  for (const r of fresh) {
+  const modelsSet = new Set(payload.meta.models || []);
+  const browsersSet = new Set(payload.meta.browsers || []);
+  for (const r of records) {
     if (r.model) modelsSet.add(r.model);
     if (r.browser) browsersSet.add(r.browser);
   }
-  baseline.meta.models = [...modelsSet].sort();
-  baseline.meta.browsers = [...browsersSet].sort();
+  payload.meta.models = [...modelsSet].sort();
+  payload.meta.browsers = [...browsersSet].sort();
 
-  // meta.machines: keep existing entries, add any new slugs from the live
-  // delta, then recompute resultCount/passCount across baseline+fresh.
-  const machineMap = new Map((baseline.meta.machines || []).map(m => [m.slug, m]));
+  const machineMap = new Map((payload.meta.machines || []).map(m => [m.slug, m]));
   for (const m of machines) {
     if (!machineMap.has(m.slug)) machineMap.set(m.slug, { ...m });
   }
@@ -59,12 +73,12 @@ function mergeLiveDelta(baseline, records, machines) {
     m.resultCount = 0;
     m.passCount = 0;
   }
+
   // Per-machine submitter aggregation — counts contributions and tracks the
-  // most recent submission so the dashboard can render a stacked-avatar row
-  // sorted by activity. Mirrors scripts/build-site.js logic so live and
-  // baseline data stay byte-equivalent.
+  // most-recent submission so the machine card can render a stacked-avatar
+  // row sorted by activity.
   const submitterAccumulator = new Map(); // slug → Map(key → {profile, count, latestAt})
-  for (const r of baseline.results) {
+  for (const r of payload.results) {
     const m = machineMap.get(r.machineSlug);
     if (!m) continue;
     m.resultCount += 1;
@@ -92,8 +106,26 @@ function mergeLiveDelta(baseline, records, machines) {
       .map(({ profile, count, latestAt }) => ({ ...profile, count, latestAt }))
       .sort((a, b) => b.count - a.count || (b.latestAt || '').localeCompare(a.latestAt || ''));
   }
-  baseline.meta.machines = [...machineMap.values()];
-  baseline.meta.generatedAt = new Date().toISOString();
+  payload.meta.machines = [...machineMap.values()];
+  payload.meta.generatedAt = new Date().toISOString();
+}
+
+function readSessionCache() {
+  try {
+    const raw = sessionStorage.getItem(SESSION_CACHE_KEY);
+    if (!raw) return null;
+    const { ts, data } = JSON.parse(raw);
+    if (typeof ts !== 'number' || (Date.now() - ts) > SESSION_CACHE_TTL_MS) return null;
+    return data;
+  } catch {
+    return null;
+  }
+}
+
+function writeSessionCache(data) {
+  try {
+    sessionStorage.setItem(SESSION_CACHE_KEY, JSON.stringify({ ts: Date.now(), data }));
+  } catch { /* quota or disabled */ }
 }
 
 /* Reduce a flat result set down to one canonical row per
@@ -123,13 +155,6 @@ export function selectBestResults(records) {
   return [...bestByCell.values()];
 }
 
-/* Composite key for dedupe. The submit pipeline writes one record per
-   (machine, browser, model, variant, timestamp) tuple, so this is unique
-   in practice. */
-function keyOf(r) {
-  return `${r.machineSlug}|${r.browser}|${r.model}|${r.variant}|${r.timestamp}`;
-}
-
 export function filterResults(results, filters) {
   return results.filter(r => {
     if (filters.machine && filters.machine !== 'all' && r.machineSlug !== filters.machine) return false;

js/dataset.js

@@ -20,9 +20,12 @@ const HF = 'https://huggingface.co';
 // dedupe out anyway.
 const CLOCK_SKEW_MS = 10 * 60 * 1000;
 
-// Cap on parallel/total file fetches per dashboard load. 50 covers ~weeks
-// of submissions before CI rebuilds; beyond that we silently truncate.
-const MAX_FETCH = 50;
+// Cap on parallel/total file fetches per dashboard load. The dashboard now
+// pulls the entire dataset live (no static baseline), so this cap is the
+// upper bound on how many run files the page will download at once. 1000
+// is conservative — actual bench submissions are typically ≤ 1 KB each so
+// the bandwidth ceiling is well under a megabyte even at the cap.
+const MAX_FETCH = 1000;
 
 /* Fetch the runs/ tree from the dataset. Returns the file entries that
    look newer than `sinceISO` (with a clock-skew buffer applied). On any
@@ -61,12 +64,20 @@ async function fetchRunFile(datasetRepo, filePath) {
   return resp.json();
 }
 
+/* List the dataset tree and download every file in `runs/`. Pure-live
+   variant of fetchRecentRuns — no cutoff, returns the entire dataset.
+   Caller is responsible for rate-limiting/caching. */
+export async function fetchAllRuns(datasetRepo) {
+  return fetchRunsBatch(datasetRepo, await listRecentRunFiles(datasetRepo, null));
+}
+
 /* List the dataset tree and download every file that's newer than the
-   baseline's generatedAt. Returns flattened records ready to append to
-   combined.results, plus the raw machine info per file so the merge can
-   keep meta.machines accurate. */
+   baseline's generatedAt. Kept for callers that still want a delta view. */
 export async function fetchRecentRuns(datasetRepo, sinceISO) {
-  const files = await listRecentRunFiles(datasetRepo, sinceISO);
+  return fetchRunsBatch(datasetRepo, await listRecentRunFiles(datasetRepo, sinceISO));
+}
+
+async function fetchRunsBatch(datasetRepo, files) {
   if (files.length === 0) return { records: [], machines: [], fileCount: 0 };
 
   const records = [];