kiku-agent 1.3.2 → 1.3.4

package/src/data/awardsData.js

@@ -1,146 +1,146 @@
-export const awardsData = [
-  {
-    id: "award-2024-weibo",
-    year: 2024,
-    titleTh: "นักแสดงหญิงยอดนิยมแห่งปี (Popular Actress of the Year)",
-    titleCn: "年度人气女演员",
-    ceremony: "Weibo TV & Internet Video Summit 2024",
-    work: "In Blossom (ฮวาเจียนลิ่ง ในม่านเมฆ)",
-    category: "Acting"
-  },
-  {
-    id: "award-2023-weibo-night",
-    year: 2023,
-    titleTh: "รางวัลบุคคลผู้ทรงอิทธิพลแห่งปี (Influential Figure of the Year)",
-    titleCn: "年度影响力艺人",
-    ceremony: "Weibo Night 2023",
-    work: "Beauty of Resilience (ตำนานเลือดฟีนิกซ์)",
-    category: "Honor"
-  },
-  {
-    id: "award-2022-erc",
-    year: 2022,
-    titleTh: "นักร้องยอดนิยมแห่งปี (Popular Singer of the Year)",
-    titleCn: "年度最佳人气歌手",
-    ceremony: "29th ERC Chinese Top Ten Awards (第29届东方风云榜)",
-    work: "Single '0.2s' & Solo Music Portfolio",
-    category: "Music"
-  },
-  {
-    id: "award-2021-tencent",
-    year: 2021,
-    titleTh: "นักแสดงนำหญิงซีรีส์ยอดเยี่ยมแห่งปี (Outstanding Actress)",
-    titleCn: "年度卓越女演员",
-    ceremony: "Tencent Video Starlight Awards 2021",
-    work: "Rebirth for You (มู่หนานจือ)",
-    category: "Acting"
-  },
-  {
-    id: "award-2020-erc-stage",
-    year: 2020,
-    titleTh: "การแสดงบนเวทียอดเยี่ยมแห่งปี (Best Stage Interpretation)",
-    titleCn: "最佳舞台演绎",
-    ceremony: "27th ERC Chinese Top Ten Awards (第27届东方风云榜)",
-    work: "Love Emergency Report (恋爱告急)",
-    category: "Music"
-  },
-  {
-    id: "award-2020-iqiyi-breakthrough",
-    year: 2020,
-    titleTh: "นักแสดงหญิงผู้ก้าวกระโดดยอดเยี่ยม (Breakthrough Actress)",
-    titleCn: "年度飞跃女演员",
-    ceremony: "iQIYI Scream Night 2020",
-    work: "The Blooms at Ruyi Pavilion & In a Class of Her Own",
-    category: "Acting"
-  },
-  {
-    id: "award-2019-china-canada",
-    year: 2019,
-    titleTh: "นักแสดงนำหญิงยอดเยี่ยม (Best Actress - Web Series)",
-    titleCn: "最佳网剧女主角",
-    ceremony: "China Canada Television Festival Award 2019 (中加电视节)",
-    work: "Legend of Yunxi (ตำนานหยุนซี)",
-    category: "Acting"
-  },
-  {
-    id: "award-2019-erc-trend",
-    year: 2019,
-    titleTh: "นักร้องผู้นำเทรนด์แห่งปี (Trend Singer Award)",
-    titleCn: "年度风尚歌手",
-    ceremony: "26th ERC Chinese Top Ten Awards (第26届东方风云榜)",
-    work: "Solo Music & OST Works",
-    category: "Music"
-  },
-  {
-    id: "award-2019-iqiyi-drama-actress",
-    year: 2019,
-    titleTh: "นักแสดงหญิงยอดนิยมด้านการแสดง (Drama Actress of the Year)",
-    titleCn: "年度戏剧人气女艺人",
-    ceremony: "iQIYI Scream Night 2019",
-    work: "The Legend of White Snake (ตำนานนางพญางูขาว)",
-    category: "Acting"
-  },
-  {
-    id: "award-2018-weibo-improved",
-    year: 2018,
-    titleTh: "ศิลปินผู้มีพัฒนาการก้าวกระโดดแห่งปี (Improved Artist of the Year)",
-    titleCn: "年度进取艺人",
-    ceremony: "Weibo Awards Ceremony 2018",
-    work: "Legend of Yunxi & Solo Career Launch",
-    category: "Honor"
-  },
-  {
-    id: "award-2017-may4th-medal",
-    year: 2017,
-    titleTh: "เหรียญรางวัลเกียรติยศเยาวชน 4 พฤษภาคม (May 4th Youth Medal)",
-    titleCn: "五四优秀青年勋章",
-    ceremony: "Chinese Communist Youth League (共青团中央)",
-    work: "เยาวชนตัวอย่างผู้สร้างคุณประโยชน์ทางศิลปวัฒนธรรม",
-    category: "State Honor"
-  },
-  {
-    id: "award-2017-erc-breakthrough",
-    year: 2017,
-    titleTh: "ศิลปินผู้มีพัฒนาการโดดเด่นยอดเยี่ยม (Best Breakthrough Artist)",
-    titleCn: "最佳飞跃艺人",
-    ceremony: "24th ERC Chinese Top Ten Awards (第24届东方风云榜)",
-    work: "EP Everyday & Music Solo Works",
-    category: "Music"
-  },
-  {
-    id: "award-2017-migu-top10",
-    year: 2017,
-    titleTh: "1 ใน 10 เพลงยอดเยี่ยมแห่งปี (Top 10 Songs of the Year)",
-    titleCn: "年度十大金曲",
-    ceremony: "11th Migu Music Awards (第11届音乐盛典咪咕汇)",
-    work: "Think of You (想你了)",
-    category: "Music"
-  },
-  {
-    id: "award-2017-snh48-champion",
-    year: 2017,
-    titleTh: "ชนะเลิศอันดับ 1 การเลือกตั้งทั่วไป SNH48 ครั้งที่ 4 (แชมป์ 2 สมัยซ้อน)",
-    titleCn: "SNH48第四届偶像年度人气总决选冠军 (晋升明星殿堂)",
-    ceremony: "SNH48 4th General Election",
-    work: "บรรจุเข้าสู่ SNH48 Star Hall of Fame (วิหารแห่งดาว)",
-    category: "Hall of Fame"
-  },
-  {
-    id: "award-2016-tencent-new-actress",
-    year: 2016,
-    titleTh: "นักแสดงโทรทัศน์หน้าใหม่ยอดเยี่ยมแห่งปี (New Television Actress)",
-    titleCn: "年度新锐电视剧女演员",
-    ceremony: "Tencent Video Star Awards 2016 (腾讯视频星光大赏)",
-    work: "Novoland: The Castle in the Sky (九州·天空城)",
-    category: "Acting"
-  },
-  {
-    id: "award-2016-snh48-champion",
-    year: 2016,
-    titleTh: "ชนะเลิศอันดับ 1 การเลือกตั้งทั่วไป SNH48 ครั้งที่ 3 (คะแนนเกิน 230,000)",
-    titleCn: "SNH48第三届偶像年度人气总决选冠军",
-    ceremony: "SNH48 3rd General Election",
-    work: "เซ็นเตอร์เพลง Princess's Cloak (公主披风)",
-    category: "Idol & Acting"
-  }
-];
+export const awardsData = [
+  {
+    id: "award-2024-weibo",
+    year: 2024,
+    titleTh: "นักแสดงหญิงยอดนิยมแห่งปี (Popular Actress of the Year)",
+    titleCn: "年度人气女演员",
+    ceremony: "Weibo TV & Internet Video Summit 2024",
+    work: "In Blossom (ฮวาเจียนลิ่ง ในม่านเมฆ)",
+    category: "Acting"
+  },
+  {
+    id: "award-2023-weibo-night",
+    year: 2023,
+    titleTh: "รางวัลบุคคลผู้ทรงอิทธิพลแห่งปี (Influential Figure of the Year)",
+    titleCn: "年度影响力艺人",
+    ceremony: "Weibo Night 2023",
+    work: "Beauty of Resilience (ตำนานเลือดฟีนิกซ์)",
+    category: "Honor"
+  },
+  {
+    id: "award-2022-erc",
+    year: 2022,
+    titleTh: "นักร้องยอดนิยมแห่งปี (Popular Singer of the Year)",
+    titleCn: "年度最佳人气歌手",
+    ceremony: "29th ERC Chinese Top Ten Awards (第29届东方风云榜)",
+    work: "Single '0.2s' & Solo Music Portfolio",
+    category: "Music"
+  },
+  {
+    id: "award-2021-tencent",
+    year: 2021,
+    titleTh: "นักแสดงนำหญิงซีรีส์ยอดเยี่ยมแห่งปี (Outstanding Actress)",
+    titleCn: "年度卓越女演员",
+    ceremony: "Tencent Video Starlight Awards 2021",
+    work: "Rebirth for You (มู่หนานจือ)",
+    category: "Acting"
+  },
+  {
+    id: "award-2020-erc-stage",
+    year: 2020,
+    titleTh: "การแสดงบนเวทียอดเยี่ยมแห่งปี (Best Stage Interpretation)",
+    titleCn: "最佳舞台演绎",
+    ceremony: "27th ERC Chinese Top Ten Awards (第27届东方风云榜)",
+    work: "Love Emergency Report (恋爱告急)",
+    category: "Music"
+  },
+  {
+    id: "award-2020-iqiyi-breakthrough",
+    year: 2020,
+    titleTh: "นักแสดงหญิงผู้ก้าวกระโดดยอดเยี่ยม (Breakthrough Actress)",
+    titleCn: "年度飞跃女演员",
+    ceremony: "iQIYI Scream Night 2020",
+    work: "The Blooms at Ruyi Pavilion & In a Class of Her Own",
+    category: "Acting"
+  },
+  {
+    id: "award-2019-china-canada",
+    year: 2019,
+    titleTh: "นักแสดงนำหญิงยอดเยี่ยม (Best Actress - Web Series)",
+    titleCn: "最佳网剧女主角",
+    ceremony: "China Canada Television Festival Award 2019 (中加电视节)",
+    work: "Legend of Yunxi (ตำนานหยุนซี)",
+    category: "Acting"
+  },
+  {
+    id: "award-2019-erc-trend",
+    year: 2019,
+    titleTh: "นักร้องผู้นำเทรนด์แห่งปี (Trend Singer Award)",
+    titleCn: "年度风尚歌手",
+    ceremony: "26th ERC Chinese Top Ten Awards (第26届东方风云榜)",
+    work: "Solo Music & OST Works",
+    category: "Music"
+  },
+  {
+    id: "award-2019-iqiyi-drama-actress",
+    year: 2019,
+    titleTh: "นักแสดงหญิงยอดนิยมด้านการแสดง (Drama Actress of the Year)",
+    titleCn: "年度戏剧人气女艺人",
+    ceremony: "iQIYI Scream Night 2019",
+    work: "The Legend of White Snake (ตำนานนางพญางูขาว)",
+    category: "Acting"
+  },
+  {
+    id: "award-2018-weibo-improved",
+    year: 2018,
+    titleTh: "ศิลปินผู้มีพัฒนาการก้าวกระโดดแห่งปี (Improved Artist of the Year)",
+    titleCn: "年度进取艺人",
+    ceremony: "Weibo Awards Ceremony 2018",
+    work: "Legend of Yunxi & Solo Career Launch",
+    category: "Honor"
+  },
+  {
+    id: "award-2017-may4th-medal",
+    year: 2017,
+    titleTh: "เหรียญรางวัลเกียรติยศเยาวชน 4 พฤษภาคม (May 4th Youth Medal)",
+    titleCn: "五四优秀青年勋章",
+    ceremony: "Chinese Communist Youth League (共青团中央)",
+    work: "เยาวชนตัวอย่างผู้สร้างคุณประโยชน์ทางศิลปวัฒนธรรม",
+    category: "State Honor"
+  },
+  {
+    id: "award-2017-erc-breakthrough",
+    year: 2017,
+    titleTh: "ศิลปินผู้มีพัฒนาการโดดเด่นยอดเยี่ยม (Best Breakthrough Artist)",
+    titleCn: "最佳飞跃艺人",
+    ceremony: "24th ERC Chinese Top Ten Awards (第24届东方风云榜)",
+    work: "EP Everyday & Music Solo Works",
+    category: "Music"
+  },
+  {
+    id: "award-2017-migu-top10",
+    year: 2017,
+    titleTh: "1 ใน 10 เพลงยอดเยี่ยมแห่งปี (Top 10 Songs of the Year)",
+    titleCn: "年度十大金曲",
+    ceremony: "11th Migu Music Awards (第11届音乐盛典咪咕汇)",
+    work: "Think of You (想你了)",
+    category: "Music"
+  },
+  {
+    id: "award-2017-snh48-champion",
+    year: 2017,
+    titleTh: "ชนะเลิศอันดับ 1 การเลือกตั้งทั่วไป SNH48 ครั้งที่ 4 (แชมป์ 2 สมัยซ้อน)",
+    titleCn: "SNH48第四届偶像年度人气总决选冠军 (晋升明星殿堂)",
+    ceremony: "SNH48 4th General Election",
+    work: "บรรจุเข้าสู่ SNH48 Star Hall of Fame (วิหารแห่งดาว)",
+    category: "Hall of Fame"
+  },
+  {
+    id: "award-2016-tencent-new-actress",
+    year: 2016,
+    titleTh: "นักแสดงโทรทัศน์หน้าใหม่ยอดเยี่ยมแห่งปี (New Television Actress)",
+    titleCn: "年度新锐电视剧女演员",
+    ceremony: "Tencent Video Star Awards 2016 (腾讯视频星光大赏)",
+    work: "Novoland: The Castle in the Sky (九州·天空城)",
+    category: "Acting"
+  },
+  {
+    id: "award-2016-snh48-champion",
+    year: 2016,
+    titleTh: "ชนะเลิศอันดับ 1 การเลือกตั้งทั่วไป SNH48 ครั้งที่ 3 (คะแนนเกิน 230,000)",
+    titleCn: "SNH48第三届偶像年度人气总决选冠军",
+    ceremony: "SNH48 3rd General Election",
+    work: "เซ็นเตอร์เพลง Princess's Cloak (公主披风)",
+    category: "Idol & Acting"
+  }
+];

package/src/data/musicVideoData.js

@@ -59,12 +59,12 @@ export const musicVideoData = [
     titlePinyin: "Lu Zhu Zai Chen Xi Qian Xiao San",
     releaseYear: 2026,
     category: "OST",
-    dramaName: "Veil of Shadows (《月鳞绮纪》) — เพลงลูกทรงของลู่หวู่อี",
-    aliases: ["Dewdrops", "Dewdrops Dissipate Before Dawn", "Dew Before Dawn", "露珠在晨曦前消散", "หยาดน้ำค้างสลายก่อนรุ่งอรุณ", "หยาดน้ำค้าง", "D-vEPdasbLo"],
-    keywords: ["dewdrops", "veil of shadows", "lu wuyi", "月鳞绮纪", "露珠", "เพลงลูกทรง"],
+    dramaName: "Veil of Shadows (《月鳞绮纪》) / In Blossom (《花间令》)",
+    aliases: ["Dewdrops", "Dewdrops Dissipate Before Dawn", "Dew Before Dawn", "In Blossom", "花间令", "露水", "露珠在晨曦前消散", "หยาดน้ำค้างสลายก่อนรุ่งอรุณ", "หยาดน้ำค้าง", "D-vEPdasbLo"],
+    keywords: ["dewdrops", "in blossom", "veil of shadows", "lu wuyi", "花间令", "月鳞绮纪", "露珠", "露水", "เพลงลูกทรง"],
     youtubeId: "D-vEPdasbLo",
     duration: "04:42",
-    description: "เพลงลูกทรงของลู่หวู่อีจากซีรีส์ Veil of Shadows เวอร์ชันทางการเผยแพร่โดย Tencent Music",
+    description: "เพลงลูกทรงของลู่หวู่อีจากซีรีส์ Veil of Shadows / In Blossom เวอร์ชันทางการเผยแพร่โดย Tencent Music",
     isFeatured: true
   },
   {
@@ -283,6 +283,8 @@ export const musicVideoData = [
     releaseYear: 2019,
     category: "Single",
     dramaName: "Dance Pop EP",
+    aliases: ["Heartbeat Alert", "Love Emergency", "Love Emergency Report", "恋爱告急", "แจ้งเตือนหัวใจฉุกเฉิน", "Heartbeat", "5MLPdCmnlmE"],
+    keywords: ["heartbeat alert", "love emergency", "恋爱告急", "แจ้งเตือนหัวใจฉุกเฉิน"],
     youtubeId: "5MLPdCmnlmE",
     duration: "03:35",
     description: "เพลงแดนซ์ป็อปไวรัลสุดสดใส คว้ารางวัล Best Stage Interpretation",

package/src/lib/vram/estimator.js

@@ -115,31 +115,96 @@ export function formatGb(bytes, digits = 2) {
 }
 
 
-export function estimateWeightsBytes({ paramCount = null, paramsB = null, quant = 'q4_k_m', weightBytesOverride = null } = {}) {
+export const KV_CACHE_DTYPE_BYTES = Object.freeze({
+  fp32: 4.0,
+  f32: 4.0,
+  fp16: 2.0,
+  f16: 2.0,
+  bf16: 2.0,
+  q8_0: 1.0625, 
+  q8: 1.0625,
+  q4_0: 0.5625,
+  q4: 0.5625,
+  q4_1: 0.625,
+  q5_0: 0.6875,
+  q5_1: 0.75,
+});
+
+export function resolveKvBytesPerValue(cacheBits = 16, cacheType = null) {
+  if (cacheType && typeof cacheType === 'string') {
+    const key = cacheType.toLowerCase().replace(/[^a-z0-9_]/g, '');
+    if (KV_CACHE_DTYPE_BYTES[key] !== undefined) return KV_CACHE_DTYPE_BYTES[key];
+  }
+  if (typeof cacheBits === 'string') {
+    const key = cacheBits.toLowerCase().replace(/[^a-z0-9_]/g, '');
+    if (KV_CACHE_DTYPE_BYTES[key] !== undefined) return KV_CACHE_DTYPE_BYTES[key];
+  }
+  const n = Number(cacheBits);
+  if (Number.isFinite(n) && n > 0) {
+    return n / 8;
+  }
+  return 2.0;
+}
+
+export function estimateWeightsBytes({
+  paramCount = null,
+  paramsB = null,
+  quant = 'q4_k_m',
+  weightBytesOverride = null,
+  gpuLayers = null,
+  totalLayers = null,
+} = {}) {
   const bpw = parseBpw(quant);
+  let rawBytes = null;
+  let count = paramCount;
   if (weightBytesOverride !== null && weightBytesOverride !== undefined && Number.isFinite(Number(weightBytesOverride)) && Number(weightBytesOverride) > 0) {
-    let count = paramCount;
+    rawBytes = Math.round(Number(weightBytesOverride));
+    if (count === null || count === undefined) {
+      count = paramsB !== null && paramsB !== undefined && Number.isFinite(Number(paramsB))
+        ? Math.round(Number(paramsB) * 1e9)
+        : null;
+    }
+  } else {
     if (count === null || count === undefined) {
       count = paramsB !== null && paramsB !== undefined && Number.isFinite(Number(paramsB))
         ? Math.round(Number(paramsB) * 1e9)
         : null;
     }
-    return { bytes: Math.round(Number(weightBytesOverride)), bpw, paramCount: count, source: 'override' };
+    if (count && count > 0 && bpw !== null) {
+      rawBytes = Math.round(count * bpw / 8);
+    }
   }
-  let count = paramCount;
-  if (count === null || count === undefined) {
-    count = paramsB !== null && paramsB !== undefined && Number.isFinite(Number(paramsB))
-      ? Math.round(Number(paramsB) * 1e9)
-      : null;
+
+  if (rawBytes === null) {
+    return { bytes: null, bpw, paramCount: count };
+  }
+
+  if (gpuLayers !== null && gpuLayers !== undefined && totalLayers !== null && totalLayers !== undefined) {
+    const gL = Number(gpuLayers);
+    const tL = Number(totalLayers);
+    if (Number.isFinite(gL) && Number.isFinite(tL) && tL > 0 && gL >= 0 && gL < tL) {
+      const fraction = Math.min(1, Math.max(0, gL / tL));
+      return {
+        bytes: Math.round(rawBytes * fraction),
+        bpw,
+        paramCount: count,
+        totalWeightBytes: rawBytes,
+        offloadFraction: fraction,
+        gpuLayers: gL,
+        totalLayers: tL,
+        source: weightBytesOverride ? 'override-fractional' : 'fractional',
+      };
+    }
   }
-  if (!count || count <= 0 || bpw === null) return { bytes: null, bpw, paramCount: count };
-  return { bytes: Math.round(count * bpw / 8), bpw, paramCount: count };
+
+  return { bytes: rawBytes, bpw, paramCount: count, source: weightBytesOverride ? 'override' : undefined };
 }
 
 
-export function computeKvBytesPerToken({ layers, kvHeads, headDim, cacheBits = 16 } = {}) {
+export function computeKvBytesPerToken({ layers, kvHeads, headDim, cacheBits = 16, cacheType = null } = {}) {
   if (!layers || !kvHeads || !headDim) return null;
-  return 2 * layers * kvHeads * headDim * (Number(cacheBits || 16) / 8);
+  const bytesPerVal = resolveKvBytesPerValue(cacheBits, cacheType);
+  return 2 * layers * kvHeads * headDim * bytesPerVal;
 }
 
 export function estimateKvCacheBytes({
@@ -149,10 +214,12 @@ export function estimateKvCacheBytes({
   contextLength = null,
   batchSize = 1,
   cacheBits = 16,
+  cacheType = null,
   swa = null,
 } = {}) {
+  const bytesPerVal = resolveKvBytesPerValue(cacheBits, cacheType);
   const perLayerPerToken = (layers && kvHeads && headDim)
-    ? 2 * kvHeads * headDim * (Number(cacheBits || 16) / 8)
+    ? 2 * kvHeads * headDim * bytesPerVal
     : null;
   const bytesPerToken = perLayerPerToken !== null ? perLayerPerToken * layers : null;
   const ctx = Number(contextLength);
@@ -197,16 +264,25 @@ export function estimateActivationBytes({
   queryHeads = null,
   kvHeads = null,
   activationBits = 16,
+  contextLength = null,
 } = {}) {
+  const batch = Math.max(1, Number(batchSize) || 1);
+  const ctx = Number(contextLength) || 0;
+  let dynamicScratchGb = DEFAULT_ACTIVATION_GB;
+  if (ctx >= 32768) {
+    dynamicScratchGb = Math.max(dynamicScratchGb, 0.70);
+  } else if (ctx >= 16384) {
+    dynamicScratchGb = Math.max(dynamicScratchGb, 0.50);
+  }
+
   if (!hiddenSize) {
-    return { analyticBytes: null, reservedBytes: Math.round(DEFAULT_ACTIVATION_GB * GIB) };
+    return { analyticBytes: null, reservedBytes: Math.round(dynamicScratchGb * GIB) };
   }
   const gqaFactor = queryHeads && kvHeads ? 5 * (queryHeads / kvHeads) : 5;
-  const batch = Math.max(1, Number(batchSize) || 1);
   const analyticBytes = Math.round(
     batch * hiddenSize * (34 + gqaFactor) * (Number(activationBits || 16) / 8)
   );
-  return { analyticBytes, reservedBytes: Math.round(DEFAULT_ACTIVATION_GB * GIB) };
+  return { analyticBytes, reservedBytes: Math.round(dynamicScratchGb * GIB) };
 }
 
 /**
@@ -219,10 +295,14 @@ export function estimateActivationBytes({
  * @param {string}  [input.quant='q4_k_m']       Weight quantization.
  * @param {number}  [input.contextLength]        Runtime sequence limit (text tokens).
  * @param {number}  [input.batchSize=1]
- * @param {number}  [input.cacheBits=16]         KV cache precision (FP16=16, Q8_0=8).
+ * @param {number}  [input.cacheBits=16]         KV cache precision (FP16=16, Q8_0=8, or 'q8_0').
+ * @param {string}  [input.cacheType]            KV cache type ('fp16', 'q8_0', 'q4_0').
+ * @param {number}  [input.gpuLayers]            Number of layers offloaded to GPU.
+ * @param {number}  [input.totalLayers]          Total model layers.
  * @param {number}  [input.layers|kvHeads|queryHeads|headDim|hiddenSize]
  * @param {object}  [input.vision]               { images, tokensPerImage, encoderParamsB }
  * @param {number}  [input.activationGb]         Override static activation reservation.
+ * @param {number}  [input.scratchGb]            Alias for activationGb.
  * @param {number}  [input.overheadGb]           Override driver/runtime overhead.
  * @returns {object} Breakdown in bytes + binary GB, warnings included.
  */
@@ -235,6 +315,8 @@ export function estimateVramFootprint(input = {}) {
     paramsB: input.paramsB,
     quant,
     weightBytesOverride: input.weightBytesOverride,
+    gpuLayers: input.gpuLayers,
+    totalLayers: input.totalLayers || input.layers,
   });
 
   const vision = input.vision || null;
@@ -262,6 +344,7 @@ export function estimateVramFootprint(input = {}) {
     contextLength: effectiveContextLength || null,
     batchSize: input.batchSize,
     cacheBits: input.cacheBits,
+    cacheType: input.cacheType,
     swa: input.swa,
   });
 
@@ -271,9 +354,11 @@ export function estimateVramFootprint(input = {}) {
     queryHeads: input.queryHeads,
     kvHeads: input.kvHeads,
     activationBits: 16,
+    contextLength: effectiveContextLength || textCtx,
   });
-  const activationBytes = input.activationGb !== undefined && input.activationGb !== null
-    ? Math.round(Number(input.activationGb) * GIB)
+  const actOverride = input.activationGb ?? input.scratchGb;
+  const activationBytes = actOverride !== undefined && actOverride !== null
+    ? Math.round(Number(actOverride) * GIB)
     : activation.reservedBytes;
   const overheadBytes = input.overheadGb !== undefined && input.overheadGb !== null
     ? Math.round(Number(input.overheadGb) * GIB)
@@ -462,3 +547,94 @@ export function recommendAdjustments(estimate, availableBytes, { targetZone = 'G
 
   return suggestions;
 }
+
+/**
+ * Simulate VRAM distribution across multiple GPUs in layer-split mode.
+ *
+ * @param {object} options
+ * @param {number} [options.totalBytes]
+ * @param {number} [options.weightsBytes]
+ * @param {number} [options.kvCacheBytes]
+ * @param {number} [options.activationBytes]
+ * @param {number} [options.overheadBytes]
+ * @param {number} [options.gpuCount=2]
+ * @param {string|number[]} [options.tensorSplit] e.g. '20,12' or [20, 12]
+ * @param {number} [options.mainGpu=0]
+ * @param {number} [options.mainGpuScratchHeadroomGb=2.0] Headroom for scratchpad/context on main GPU
+ * @returns {object} Breakdown per GPU
+ */
+export function simulateMultiGpuFootprint({
+  totalBytes = null,
+  weightsBytes = null,
+  kvCacheBytes = null,
+  activationBytes = null,
+  overheadBytes = null,
+  gpuCount = 2,
+  tensorSplit = null,
+  mainGpu = 0,
+  mainGpuScratchHeadroomGb = 2.0,
+} = {}) {
+  const count = Math.max(1, Number(gpuCount) || 2);
+  const mg = Math.max(0, Math.min(count - 1, Number(mainGpu) || 0));
+
+  let ratios = [];
+  if (typeof tensorSplit === 'string' && tensorSplit.trim()) {
+    ratios = tensorSplit.split(',').map(s => Number(s.trim())).filter(n => Number.isFinite(n) && n >= 0);
+  } else if (Array.isArray(tensorSplit)) {
+    ratios = tensorSplit.map(n => Number(n)).filter(n => Number.isFinite(n) && n >= 0);
+  }
+
+  if (ratios.length < count) {
+    while (ratios.length < count) {
+      ratios.push(1);
+    }
+  } else if (ratios.length > count) {
+    ratios = ratios.slice(0, count);
+  }
+
+  const sumRatios = ratios.reduce((a, b) => a + b, 0) || count;
+  const normalizedRatios = ratios.map(r => r / sumRatios);
+
+  const wBytes = weightsBytes ?? 0;
+  const kvBytes = kvCacheBytes ?? 0;
+  const actBytes = activationBytes ?? Math.round(DEFAULT_ACTIVATION_GB * GIB);
+  const ovBytes = overheadBytes ?? Math.round(DEFAULT_OVERHEAD_GB * GIB);
+  const mgHeadroomBytes = Math.round(Number(mainGpuScratchHeadroomGb || 2.0) * GIB);
+
+  const gpus = normalizedRatios.map((ratio, idx) => {
+    const isMain = idx === mg;
+    const gpuWeights = Math.round(wBytes * ratio);
+    const gpuKv = Math.round(kvBytes * ratio);
+    const gpuAct = isMain ? actBytes + mgHeadroomBytes : Math.round(actBytes * 0.5);
+    const gpuOv = ovBytes;
+    const gpuTotal = gpuWeights + gpuKv + gpuAct + gpuOv;
+
+    return {
+      gpuIndex: idx,
+      isMain,
+      ratio: Number((ratio * 100).toFixed(1)),
+      ratioRaw: ratios[idx],
+      weightsBytes: gpuWeights,
+      kvCacheBytes: gpuKv,
+      activationBytes: gpuAct,
+      overheadBytes: gpuOv,
+      totalBytes: gpuTotal,
+      gb: {
+        weights: Number((gpuWeights / GIB).toFixed(2)),
+        kvCache: Number((gpuKv / GIB).toFixed(2)),
+        activation: Number((gpuAct / GIB).toFixed(2)),
+        overhead: Number((gpuOv / GIB).toFixed(2)),
+        total: Number((gpuTotal / GIB).toFixed(2)),
+      },
+    };
+  });
+
+  return {
+    gpuCount: count,
+    mainGpu: mg,
+    tensorSplit: ratios.join(','),
+    gpus,
+    totalAllocatedBytes: gpus.reduce((acc, g) => acc + g.totalBytes, 0),
+  };
+}
+