feat: add AttnRes gradient scale lab
This commit is contained in:
wuyang
@@ -0,0 +1,813 @@
|
||||
---
|
||||
import rawLab from "@/data/k3-attnres-gradient-compact.json";
|
||||
|
||||
const lab = rawLab as any;
|
||||
const json = JSON.stringify(lab).replaceAll("<", "\\u003c");
|
||||
const depth16 = lab.depth_summaries["16"];
|
||||
const depth32 = lab.depth_summaries["32"];
|
||||
const formatSigned = (value: number, digits = 4) =>
|
||||
`${value < 0 ? "−" : value > 0 ? "+" : ""}${Math.abs(value).toFixed(digits)}`;
|
||||
const pct = (value: number, digits = 1) => `${(value * 100).toFixed(digits)}%`;
|
||||
const shortHash = (value: string) => `${value.slice(0, 10)}…${value.slice(-8)}`;
|
||||
---
|
||||
|
||||
<figure class="gradient-lab" data-gradient-lab>
|
||||
<header class="gradient-head">
|
||||
<div>
|
||||
<p>ROUND 05 / GRADIENT DEFINITION × DEPTH SCALE</p>
|
||||
<h3>“早层不再过大”与“整条谱更均匀”不是同一件事</h3>
|
||||
</div>
|
||||
<p>
|
||||
16 / 32 Transformer blocks · Baseline / Block · 3 seeds · 8,000 steps。
|
||||
被测对象是 post-MLP output activation gradient，不冒充论文未公开的 Figure 5 实现。
|
||||
</p>
|
||||
</header>
|
||||
|
||||
<div class="gradient-ledger">
|
||||
<article><span>FORMAL GRID</span><b>2 × 2 × 3</b><p>12 个独立训练格</p></article>
|
||||
<article><span>TARGET BYTES</span><b>786.432M</b><p>每格 65,536,000</p></article>
|
||||
<article><span>DEPTH</span><b>16 → 32</b><p>宽度固定 192</p></article>
|
||||
<article class="split"><span>CV</span><b>6 / 6 更差</b><p>中后段局部尖峰</p></article>
|
||||
<article class="pass"><span>FIRST ↔ LAST</span><b>6 / 6 更近</b><p>失衡改善 56%–81%</p></article>
|
||||
<article class="pass"><span>FULL REPLAY</span><b>exact</b><p>model + optimizer state</p></article>
|
||||
</div>
|
||||
|
||||
<div class="gradient-tabs" role="tablist" aria-label="选择 AttnRes 梯度定义与深度实验视图">
|
||||
<button type="button" role="tab" data-gradient-tab="definition" aria-selected="true">
|
||||
<span>01</span><b>论文到底定义了什么</b><small>known · underdefined · operationalization</small>
|
||||
</button>
|
||||
<button type="button" role="tab" data-gradient-tab="spectrum" aria-selected="false">
|
||||
<span>02</span><b>绝对谱与归一化谱</b><small>depth · seed · checkpoint</small>
|
||||
</button>
|
||||
<button type="button" role="tab" data-gradient-tab="timeline" aria-selected="false">
|
||||
<span>03</span><b>六个时点怎样演化</b><small>CV · imbalance · mean scale</small>
|
||||
</button>
|
||||
<button type="button" role="tab" data-gradient-tab="output" aria-selected="false">
|
||||
<span>04</span><b>Output RMS 与块节律</b><small>growth · reset · group boundary</small>
|
||||
</button>
|
||||
<button type="button" role="tab" data-gradient-tab="verdict" aria-selected="false">
|
||||
<span>05</span><b>联合判定、成本与重放</b><small>activation ≠ parameter</small>
|
||||
</button>
|
||||
</div>
|
||||
|
||||
<section class="gradient-panel" data-gradient-panel="definition">
|
||||
<div class="panel-lead">
|
||||
<div><span>I / DEFINITION AUDIT</span><h4>论文画出一条“梯度曲线”，却没有给出足以唯一重算的测量合同</h4></div>
|
||||
<p>
|
||||
Figure 5(c) 只写 “Each transformer block’s gradient magnitude”。
|
||||
官方仓库没有训练代码、checkpoint、统计脚本或原始数组。
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<div class="known-grid">
|
||||
<article class="known">
|
||||
<span>OFFICIAL / KNOWN</span>
|
||||
<h5>图与文字能确认</h5>
|
||||
<ul>
|
||||
<li>横轴是 Transformer block index。</li>
|
||||
<li>Figure 5(b) 同时画 block output magnitude。</li>
|
||||
<li>正文说 Baseline 早层梯度过大，Block 更均匀。</li>
|
||||
<li>最终模型约 27 blocks / 54 residual layers。</li>
|
||||
</ul>
|
||||
</article>
|
||||
<article class="unknown">
|
||||
<span>OFFICIAL / UNDERDEFINED</span>
|
||||
<h5>无法从公开工件唯一恢复</h5>
|
||||
<ul>
|
||||
<li>activation、branch 还是 parameter gradient？</li>
|
||||
<li>L2、RMS、mean absolute 还是别的 norm？</li>
|
||||
<li>batch / token / channel 怎样 reduction？</li>
|
||||
<li>哪个 checkpoint、AMP / clip 前还是后？</li>
|
||||
</ul>
|
||||
</article>
|
||||
</div>
|
||||
|
||||
<div class="object-chain" aria-label="Round 05 activation gradient 测量对象">
|
||||
<div><span>FIXED INPUT</span><b>16 × 256 bytes</b><p>同一 diagnostic tensor</p></div>
|
||||
<i>→</i>
|
||||
<div><span>POST-MLP OUTPUT</span><b>h₁ … h<sub>L</sub></b><p>每个 Transformer block 一个</p></div>
|
||||
<i>→</i>
|
||||
<div><span>TOKEN-MEAN CE</span><b>ℒ</b><p>FP32 cross entropy</p></div>
|
||||
<i>→</i>
|
||||
<div class="accent"><span>MEASURE</span><b>RMS(∂ℒ/∂h<sub>l</sub>)</b><p>B × T × C 联合 RMS</p></div>
|
||||
</div>
|
||||
|
||||
<div class="object-compare">
|
||||
<article><span>ROUND 04</span><b>∇<sub>θl</sub>ℒ</b><p>核心参数梯度：权重收到多大更新信号。</p></article>
|
||||
<i>≠</i>
|
||||
<article><span>ROUND 05</span><b>∂ℒ/∂h<sub>l</sub></b><p>activation gradient：损失对这一深度表示有多敏感。</p></article>
|
||||
<p>两种对象都公开；新指标不会覆盖上一轮反结果。</p>
|
||||
</div>
|
||||
|
||||
<div class="definition-boundary">
|
||||
<b>能说</b><p>“这是与 Figure 5 叙述对齐的一种公开 operationalization。”</p>
|
||||
<b>不能说</b><p>“论文作者就是这样算的，或本站复画了 Figure 5(c)。”</p>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="gradient-panel" data-gradient-panel="spectrum" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>II / ALIGNED DEPTH SPECTRUM</span><h4>同一条梯度谱，绝对值与归一化形状要一起看</h4></div>
|
||||
<p>竖线是 8 个 AttnRes aggregation groups 的边界；Baseline 也画同位置，方便逐层配对。</p>
|
||||
</div>
|
||||
|
||||
<div class="lab-controls">
|
||||
<div role="group" aria-label="选择深度">
|
||||
<button type="button" data-spectrum-depth="16" aria-pressed="false">DEPTH 16</button>
|
||||
<button type="button" data-spectrum-depth="32" aria-pressed="true">DEPTH 32</button>
|
||||
</div>
|
||||
<label>SEED
|
||||
<select data-spectrum-seed>
|
||||
<option value="mean">3-SEED MEAN</option>
|
||||
{lab.study.seeds.map((seed: number) => <option value={String(seed)}>{seed}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<label>CHECKPOINT
|
||||
<select data-spectrum-step>
|
||||
{lab.study.diagnostic_steps.map((step: number) => <option value={String(step)} selected={step === 8000}>STEP {step.toLocaleString("en-US")}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<div role="group" aria-label="选择绝对或归一化梯度谱">
|
||||
<button type="button" data-spectrum-scale="absolute" aria-pressed="true">ABSOLUTE RMS</button>
|
||||
<button type="button" data-spectrum-scale="normalized" aria-pressed="false">÷ LAYER MEAN</button>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="spectrum-layout">
|
||||
<div class="chart-shell">
|
||||
<header><span data-spectrum-title>ACTIVATION GRADIENT RMS</span><b>POST-MLP BLOCK OUTPUT</b></header>
|
||||
<svg data-spectrum-chart viewBox="0 0 920 360" role="img" aria-label="Baseline 与 Block 的逐层 activation gradient 谱">
|
||||
<g data-chart-groups></g>
|
||||
<g data-chart-grid></g>
|
||||
<polyline data-chart-line="baseline" class="series baseline"></polyline>
|
||||
<polyline data-chart-line="block" class="series block"></polyline>
|
||||
<g data-chart-points="baseline"></g>
|
||||
<g data-chart-points="block"></g>
|
||||
<text x="460" y="350" class="axis-title">TRANSFORMER BLOCK INDEX</text>
|
||||
</svg>
|
||||
<div class="chart-legend">
|
||||
<span><i class="baseline"></i>Baseline</span>
|
||||
<span><i class="block"></i>Block AttnRes</span>
|
||||
<span><i class="boundary"></i>aggregation boundary</span>
|
||||
</div>
|
||||
</div>
|
||||
<div class="spectrum-readout">
|
||||
<span data-spectrum-state>DEPTH 32 · 3-SEED MEAN · STEP 8,000</span>
|
||||
<article><b>POPULATION CV</b><div><span>BASE</span><strong data-spectrum-base-cv>0.3786</strong></div><div><span>BLOCK</span><strong data-spectrum-block-cv>0.5996</strong></div></article>
|
||||
<article><b>FIRST / LAST QUARTILE</b><div><span>BASE</span><strong data-spectrum-base-ratio>3.21×</strong></div><div><span>BLOCK</span><strong data-spectrum-block-ratio>0.95×</strong></div></article>
|
||||
<article><b>MEAN ABSOLUTE SCALE</b><div><span>BASE</span><strong data-spectrum-base-mean>1.44e−4</strong></div><div><span>BLOCK</span><strong data-spectrum-block-mean>0.78e−4</strong></div></article>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="split-result">
|
||||
<article><span>FIRST ↔ LAST</span><b>更接近</b><p>Baseline 的早层整体隆起被削弱。</p></article>
|
||||
<i>但</i>
|
||||
<article><span>ALL-LAYER CV</span><b>反而更高</b><p>中后段少数位置形成更尖的峰。</p></article>
|
||||
<i>所以</i>
|
||||
<article class="accent"><span>PRE-REGISTERED</span><b>mixed</b><p>“更均匀”必须拆成至少两个指标。</p></article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="gradient-panel" data-gradient-panel="timeline" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>III / SIX FROZEN CHECKPOINTS</span><h4>Block 的中期优势会反转；不能挑一个 checkpoint 讲故事</h4></div>
|
||||
<p>横轴按六个预注册诊断时点等距排列；标签保留真实 step，不暗示实际时间等距。</p>
|
||||
</div>
|
||||
|
||||
<div class="lab-controls">
|
||||
<div role="group" aria-label="选择时间轨迹深度">
|
||||
<button type="button" data-time-depth="16" aria-pressed="false">DEPTH 16</button>
|
||||
<button type="button" data-time-depth="32" aria-pressed="true">DEPTH 32</button>
|
||||
</div>
|
||||
<label>SEED
|
||||
<select data-time-seed>
|
||||
<option value="mean">3-SEED MEAN</option>
|
||||
{lab.study.seeds.map((seed: number) => <option value={String(seed)}>{seed}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<div role="group" aria-label="选择时间轨迹指标">
|
||||
<button type="button" data-time-metric="cv" aria-pressed="true">CV</button>
|
||||
<button type="button" data-time-metric="imbalance" aria-pressed="false">FIRST/LAST IMBALANCE</button>
|
||||
<button type="button" data-time-metric="mean" aria-pressed="false">ABSOLUTE MEAN</button>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="chart-shell timeline-chart">
|
||||
<header><span data-time-title>POPULATION CV</span><b data-time-copy>DEPTH 32 · 3-SEED MEAN</b></header>
|
||||
<svg data-time-chart viewBox="0 0 920 360" role="img" aria-label="六个固定训练时点的 activation gradient 指标轨迹">
|
||||
<g data-time-grid></g>
|
||||
<polyline data-time-line="baseline" class="series baseline"></polyline>
|
||||
<polyline data-time-line="block" class="series block"></polyline>
|
||||
<g data-time-points="baseline"></g>
|
||||
<g data-time-points="block"></g>
|
||||
<text x="460" y="350" class="axis-title">PREREGISTERED DIAGNOSTIC STEP</text>
|
||||
</svg>
|
||||
<div class="chart-legend"><span><i class="baseline"></i>Baseline</span><span><i class="block"></i>Block AttnRes</span></div>
|
||||
</div>
|
||||
|
||||
<div class="timeline-notes">
|
||||
<article><span>SEED 01 / DEPTH 32</span><b>0.344 → 0.538</b><p>step 2,000：Block 已显著更尖。</p></article>
|
||||
<article><span>SEED 02 / DEPTH 32</span><b>0.352 → 0.618</b><p>同一时点复现恶化方向。</p></article>
|
||||
<article class="counter"><span>SEED 03 / DEPTH 32</span><b>0.400 → 0.349</b><p>中期反例：Block 此时反而更平。</p></article>
|
||||
<article><span>SEED 03 / FINAL</span><b>0.403 → 0.427</b><p>到 8,000 step 才轻微反转。</p></article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="gradient-panel" data-gradient-panel="output" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>IV / OUTPUT MAGNITUDE</span><h4>梯度结论 mixed，不代表论文所有训练动力学叙述都没有出现</h4></div>
|
||||
<p>同一个 post-MLP 位置计算 output RMS；这里不做 backward，也不改变主判定。</p>
|
||||
</div>
|
||||
|
||||
<div class="lab-controls">
|
||||
<div role="group" aria-label="选择 output RMS 深度">
|
||||
<button type="button" data-output-depth="16" aria-pressed="false">DEPTH 16</button>
|
||||
<button type="button" data-output-depth="32" aria-pressed="true">DEPTH 32</button>
|
||||
</div>
|
||||
<label>SEED
|
||||
<select data-output-seed>
|
||||
<option value="mean">3-SEED MEAN</option>
|
||||
{lab.study.seeds.map((seed: number) => <option value={String(seed)}>{seed}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<label>CHECKPOINT
|
||||
<select data-output-step>
|
||||
{lab.study.diagnostic_steps.map((step: number) => <option value={String(step)} selected={step === 8000}>STEP {step.toLocaleString("en-US")}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
</div>
|
||||
|
||||
<div class="chart-shell">
|
||||
<header><span>POST-MLP OUTPUT RMS</span><b data-output-copy>DEPTH 32 · 3-SEED MEAN · STEP 8,000</b></header>
|
||||
<svg data-output-chart viewBox="0 0 920 360" role="img" aria-label="Baseline 与 Block 的逐层 output RMS">
|
||||
<g data-output-groups></g>
|
||||
<g data-output-grid></g>
|
||||
<polyline data-output-line="baseline" class="series baseline"></polyline>
|
||||
<polyline data-output-line="block" class="series block"></polyline>
|
||||
<g data-output-points="baseline"></g>
|
||||
<g data-output-points="block"></g>
|
||||
<text x="460" y="350" class="axis-title">TRANSFORMER BLOCK INDEX</text>
|
||||
</svg>
|
||||
<div class="chart-legend"><span><i class="baseline"></i>Baseline</span><span><i class="block"></i>Block AttnRes</span><span><i class="boundary"></i>aggregation boundary</span></div>
|
||||
</div>
|
||||
|
||||
<div class="output-ratios">
|
||||
<article><span>DEPTH 16 / LAST ÷ FIRST</span><div><b>Baseline</b><strong>4.59×</strong></div><div><b>Block</b><strong>1.17×</strong></div></article>
|
||||
<article><span>DEPTH 32 / LAST ÷ FIRST</span><div><b>Baseline</b><strong>6.08×</strong></div><div><b>Block</b><strong>1.89×</strong></div></article>
|
||||
<article class="accent"><span>WHAT THE CURVE SAYS</span><b>全局累积 → 组内锯齿</b><p>Block 限制 output magnitude 持续跨深度增长；这一方向与 Figure 5(b) 叙述一致。</p></article>
|
||||
</div>
|
||||
|
||||
<div class="group-rhythm">
|
||||
<header><span data-rhythm-title>BLOCK ATTNRES · DEPTH 32</span><b>8 AGGREGATION GROUPS</b></header>
|
||||
<div data-output-bars aria-label="Block AttnRes output RMS 组内节律"></div>
|
||||
<p>粗分隔线是 group boundary；柱高来自当前选择的 checkpoint / seed，不是示意动画。</p>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="gradient-panel" data-gradient-panel="verdict" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>V / JOINT VERDICT</span><h4>一个正结果、两个反结果和一张成本账，要同时摆在桌面上</h4></div>
|
||||
<p>Round 05 的主判定只读 activation CV + imbalance；BPC、参数梯度与成本是必须公开的次要结果。</p>
|
||||
</div>
|
||||
|
||||
<div class="verdict-table-wrap">
|
||||
<table class="verdict-table">
|
||||
<thead><tr><th>Depth</th><th>Δ BPC</th><th>Activation CV</th><th>First/last imbalance</th><th>Mean grad scale</th><th>Parameter CV</th><th>Verdict</th></tr></thead>
|
||||
<tbody>
|
||||
<tr>
|
||||
<th>16</th>
|
||||
<td>{formatSigned(depth16.means.block_minus_baseline_bpc, 5)}</td>
|
||||
<td class="bad">{pct(depth16.means.relative_cv_reduction)} reduction</td>
|
||||
<td class="good">+{pct(depth16.means.relative_imbalance_reduction)}</td>
|
||||
<td>{pct(depth16.means.block_to_baseline_activation_grad_mean)} of Base</td>
|
||||
<td>{depth16.means.baseline_parameter_grad_cv.toFixed(3)} → {depth16.means.block_parameter_grad_cv.toFixed(3)}</td>
|
||||
<td>mixed</td>
|
||||
</tr>
|
||||
<tr>
|
||||
<th>32</th>
|
||||
<td>{formatSigned(depth32.means.block_minus_baseline_bpc, 5)}</td>
|
||||
<td class="bad">{pct(depth32.means.relative_cv_reduction)} reduction</td>
|
||||
<td class="good">+{pct(depth32.means.relative_imbalance_reduction)}</td>
|
||||
<td>{pct(depth32.means.block_to_baseline_activation_grad_mean)} of Base</td>
|
||||
<td>{depth32.means.baseline_parameter_grad_cv.toFixed(3)} → {depth32.means.block_parameter_grad_cv.toFixed(3)}</td>
|
||||
<td>mixed</td>
|
||||
</tr>
|
||||
</tbody>
|
||||
</table>
|
||||
</div>
|
||||
|
||||
<div class="metric-pairs">
|
||||
<article><span>ACTIVATION / FIRST-LAST</span><b>Block 改善</b><p>6 / 6 配对更接近 1。</p></article>
|
||||
<article class="warn"><span>ACTIVATION / ALL-LAYER CV</span><b>Block 恶化</b><p>6 / 6 配对 CV 更高。</p></article>
|
||||
<article class="warn"><span>PARAMETER / ALL-LAYER CV</span><b>Block 恶化</b><p>0.416→0.683；0.397→0.772。</p></article>
|
||||
<article class="pass"><span>VALIDATION BPC</span><b>Block 更低</b><p>6 / 6 配对为负；非同算力。</p></article>
|
||||
</div>
|
||||
|
||||
<div class="cost-compare">
|
||||
<article><span>DEPTH 16</span><div><b>STEP TIME</b><strong>21.47 → 54.71 ms</strong></div><div><b>PEAK ALLOC</b><strong>3.04 → 6.54 GB</strong></div><p>约 2.55× time · 2.15× memory</p></article>
|
||||
<article><span>DEPTH 32</span><div><b>STEP TIME</b><strong>42.11 → 109.38 ms</strong></div><div><b>PEAK ALLOC</b><strong>5.94 → 12.82 GB</strong></div><p>约 2.60× time · 2.16× memory</p></article>
|
||||
<article class="boundary"><span>CLAIM BOUNDARY</span><b>同 token / step</b><p>不能写成同 FLOPs、同 wall time，或 K3 生产成本。</p></article>
|
||||
</div>
|
||||
|
||||
<div class="replay-ledger">
|
||||
<article><span>FORMAL</span><b>depth-32 · Block · seed-01</b><p>8,000 steps from initialization</p></article>
|
||||
<i>≡</i>
|
||||
<article><span>FRESH REPLAY</span><b>all frozen fields exact</b><p>额外 65,536,000 target bytes</p></article>
|
||||
<i>→</i>
|
||||
<article class="pass"><span>STATE HASH</span><b>model + optimizer exact</b><p>{shortHash(lab.cells.find((cell: any) => cell.depth === 32 && cell.architecture === "block" && cell.seed === 2026073001).hashes.final_model_state)}</p></article>
|
||||
</div>
|
||||
|
||||
<div class="hash-ledger">
|
||||
<article><span>MANIFEST</span><code>{lab.manifest_summary.file_sha256}</code></article>
|
||||
<article><span>SCHEDULE</span><code>{lab.manifest_summary.formal_schedule_sha256}</code></article>
|
||||
<article><span>COMPACT CANONICAL</span><code>{lab.canonical_sha256_without_self}</code></article>
|
||||
<article><span>OVERALL</span><code>{lab.overall_verdict}</code></article>
|
||||
</div>
|
||||
|
||||
<div class="claim-grid">
|
||||
<article class="yes"><span>THIS STUDY SUPPORTS</span><ul><li>在公开定义下，Block 改写了梯度失衡的形态。</li><li>早/晚深度更接近，但中后段局部峰更尖。</li><li>Output RMS 的跨深度增长显著受限。</li></ul></article>
|
||||
<article class="no"><span>THIS STUDY DOES NOT SUPPORT</span><ul><li>复现论文 Figure 5(c) 的数值或隐藏实现。</li><li>测到 K3 checkpoint 的真实梯度。</li><li>证明 AttnRes 一般更稳定或更省算力。</li></ul></article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<script is:inline type="application/json" data-gradient-payload set:html={json}></script>
|
||||
</figure>
|
||||
|
||||
<script>
|
||||
const initializeGradientLab = (root: HTMLElement) => {
|
||||
if (root.dataset.gradientReady === "true") return;
|
||||
root.dataset.gradientReady = "true";
|
||||
const payload = root.querySelector<HTMLScriptElement>("[data-gradient-payload]");
|
||||
if (!payload?.textContent) return;
|
||||
const data = JSON.parse(payload.textContent);
|
||||
const architectures = ["baseline", "block"];
|
||||
const steps = data.study.diagnostic_steps;
|
||||
const seeds = data.study.seeds;
|
||||
const ns = "http://www.w3.org/2000/svg";
|
||||
const colors: Record<string, string> = { baseline: "#77746b", block: "#ba603b" };
|
||||
|
||||
const tabs = [...root.querySelectorAll<HTMLButtonElement>("[data-gradient-tab]")];
|
||||
const panels = [...root.querySelectorAll<HTMLElement>("[data-gradient-panel]")];
|
||||
const selectTab = (id: string) => {
|
||||
tabs.forEach((tab) => tab.setAttribute("aria-selected", String(tab.dataset.gradientTab === id)));
|
||||
panels.forEach((panel) => { panel.hidden = panel.dataset.gradientPanel !== id; });
|
||||
};
|
||||
tabs.forEach((tab, index) => {
|
||||
tab.addEventListener("click", () => selectTab(tab.dataset.gradientTab || "definition"));
|
||||
tab.addEventListener("keydown", (event) => {
|
||||
if (!["ArrowLeft", "ArrowRight", "Home", "End"].includes(event.key)) return;
|
||||
event.preventDefault();
|
||||
const nextIndex = event.key === "Home" ? 0
|
||||
: event.key === "End" ? tabs.length - 1
|
||||
: (index + (event.key === "ArrowRight" ? 1 : -1) + tabs.length) % tabs.length;
|
||||
tabs[nextIndex].focus();
|
||||
selectTab(tabs[nextIndex].dataset.gradientTab || "definition");
|
||||
});
|
||||
});
|
||||
|
||||
const findCell = (depth: number, architecture: string, seed: number) =>
|
||||
data.cells.find((cell: any) => cell.depth === depth && cell.architecture === architecture && cell.seed === seed);
|
||||
const findDiagnostic = (cell: any, step: number) =>
|
||||
cell.diagnostics.find((row: any) => row.step === step);
|
||||
const seedList = (seedKey: string) => seedKey === "mean" ? seeds : [Number(seedKey)];
|
||||
const average = (values: number[]) => values.reduce((sum, value) => sum + value, 0) / values.length;
|
||||
const arrayFor = (depth: number, architecture: string, seedKey: string, step: number, field: string) => {
|
||||
const rows = seedList(seedKey).map((seed: number) => findDiagnostic(findCell(depth, architecture, seed), step)[field]);
|
||||
return rows[0].map((_: number, index: number) => average(rows.map((row: number[]) => row[index])));
|
||||
};
|
||||
const statFor = (depth: number, architecture: string, seedKey: string, step: number, group: string, field: string) =>
|
||||
average(seedList(seedKey).map((seed: number) => findDiagnostic(findCell(depth, architecture, seed), step)[group][field]));
|
||||
const formatStep = (value: number) => value.toLocaleString("en-US");
|
||||
const formatScientific = (value: number) => value.toExponential(2).replace("e-", "e−");
|
||||
const svgNode = (name: string, attrs: Record<string, string | number>) => {
|
||||
const node = document.createElementNS(ns, name);
|
||||
Object.entries(attrs).forEach(([key, value]) => node.setAttribute(key, String(value)));
|
||||
return node;
|
||||
};
|
||||
|
||||
const drawGroups = (group: SVGGElement, depth: number) => {
|
||||
group.replaceChildren();
|
||||
for (let index = 1; index < 8; index += 1) {
|
||||
const block = depth / 8 * index;
|
||||
const x = 58 + block / (depth - 1) * 822;
|
||||
group.append(svgNode("line", { x1: x, x2: x, y1: 24, y2: 304, class: "group-boundary" }));
|
||||
}
|
||||
};
|
||||
const drawChart = (
|
||||
svg: SVGSVGElement,
|
||||
groups: SVGGElement | null,
|
||||
grid: SVGGElement,
|
||||
lines: Record<string, SVGPolylineElement | null>,
|
||||
pointGroups: Record<string, SVGGElement | null>,
|
||||
series: Record<string, number[]>,
|
||||
labels: string[],
|
||||
options: { depth?: number; formatter?: (value: number) => string; zero?: boolean } = {},
|
||||
) => {
|
||||
grid.replaceChildren();
|
||||
const values = Object.values(series).flat();
|
||||
const rawMin = options.zero === false ? Math.min(...values) : 0;
|
||||
const rawMax = Math.max(...values);
|
||||
const padding = rawMax === rawMin ? 1 : (rawMax - rawMin) * 0.08;
|
||||
const minimum = Math.max(0, rawMin - padding);
|
||||
const maximum = rawMax + padding;
|
||||
const xAt = (index: number) => 58 + index / Math.max(1, labels.length - 1) * 822;
|
||||
const yAt = (value: number) => 24 + (maximum - value) / Math.max(1e-30, maximum - minimum) * 280;
|
||||
const formatter = options.formatter || ((value: number) => value.toFixed(2));
|
||||
for (let index = 0; index < 5; index += 1) {
|
||||
const value = minimum + (maximum - minimum) * (4 - index) / 4;
|
||||
const y = 24 + index / 4 * 280;
|
||||
grid.append(svgNode("line", { x1: 58, x2: 880, y1: y, y2: y, class: "grid-line" }));
|
||||
const text = svgNode("text", { x: 48, y: y + 4, class: "axis-label y" });
|
||||
text.textContent = formatter(value);
|
||||
grid.append(text);
|
||||
}
|
||||
labels.forEach((label, index) => {
|
||||
const text = svgNode("text", { x: xAt(index), y: 326, class: "axis-label x" });
|
||||
text.textContent = label;
|
||||
grid.append(text);
|
||||
});
|
||||
if (groups && options.depth) drawGroups(groups, options.depth);
|
||||
architectures.forEach((architecture) => {
|
||||
const points = series[architecture].map((value, index) => `${xAt(index)},${yAt(value)}`).join(" ");
|
||||
lines[architecture]?.setAttribute("points", points);
|
||||
pointGroups[architecture]?.replaceChildren();
|
||||
series[architecture].forEach((value, index) => {
|
||||
const point = svgNode("circle", {
|
||||
cx: xAt(index), cy: yAt(value), r: 3.6, fill: colors[architecture],
|
||||
});
|
||||
const title = svgNode("title", {});
|
||||
title.textContent = `${architecture === "baseline" ? "Baseline" : "Block"} · ${labels[index]} · ${formatter(value)}`;
|
||||
point.append(title);
|
||||
pointGroups[architecture]?.append(point);
|
||||
});
|
||||
});
|
||||
svg.dataset.maximum = String(maximum);
|
||||
};
|
||||
|
||||
let spectrumDepth = 32;
|
||||
let spectrumScale = "absolute";
|
||||
const spectrumSeed = root.querySelector<HTMLSelectElement>("[data-spectrum-seed]")!;
|
||||
const spectrumStep = root.querySelector<HTMLSelectElement>("[data-spectrum-step]")!;
|
||||
const spectrumChart = root.querySelector<SVGSVGElement>("[data-spectrum-chart]")!;
|
||||
const updateSpectrum = () => {
|
||||
const seedKey = spectrumSeed.value;
|
||||
const step = Number(spectrumStep.value);
|
||||
const absolute = Object.fromEntries(architectures.map((architecture) => [
|
||||
architecture,
|
||||
arrayFor(spectrumDepth, architecture, seedKey, step, "activation_grad_rms_by_block"),
|
||||
]));
|
||||
const series = spectrumScale === "normalized"
|
||||
? Object.fromEntries(architectures.map((architecture) => {
|
||||
const mean = average(absolute[architecture]);
|
||||
return [architecture, absolute[architecture].map((value: number) => value / mean)];
|
||||
}))
|
||||
: absolute;
|
||||
drawChart(
|
||||
spectrumChart,
|
||||
spectrumChart.querySelector("[data-chart-groups]"),
|
||||
spectrumChart.querySelector("[data-chart-grid]")!,
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, spectrumChart.querySelector(`[data-chart-line="${architecture}"]`)])),
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, spectrumChart.querySelector(`[data-chart-points="${architecture}"]`)])),
|
||||
series,
|
||||
Array.from({ length: spectrumDepth }, (_, index) => String(index + 1)),
|
||||
{
|
||||
depth: spectrumDepth,
|
||||
formatter: spectrumScale === "absolute" ? formatScientific : (value) => `${value.toFixed(1)}×`,
|
||||
},
|
||||
);
|
||||
const title = root.querySelector<HTMLElement>("[data-spectrum-title]");
|
||||
if (title) title.textContent = spectrumScale === "absolute" ? "ACTIVATION GRADIENT RMS" : "NORMALIZED BY LAYER MEAN";
|
||||
const state = root.querySelector<HTMLElement>("[data-spectrum-state]");
|
||||
if (state) state.textContent = `DEPTH ${spectrumDepth} · ${seedKey === "mean" ? "3-SEED MEAN" : seedKey} · STEP ${formatStep(step)}`;
|
||||
architectures.forEach((architecture) => {
|
||||
const cv = statFor(spectrumDepth, architecture, seedKey, step, "activation_grad_statistics", "population_cv");
|
||||
const ratio = statFor(spectrumDepth, architecture, seedKey, step, "activation_grad_statistics", "first_to_last_ratio");
|
||||
const mean = statFor(spectrumDepth, architecture, seedKey, step, "activation_grad_statistics", "mean");
|
||||
const prefix = architecture === "baseline" ? "base" : "block";
|
||||
root.querySelector<HTMLElement>(`[data-spectrum-${prefix}-cv]`)!.textContent = cv.toFixed(4);
|
||||
root.querySelector<HTMLElement>(`[data-spectrum-${prefix}-ratio]`)!.textContent = `${ratio.toFixed(2)}×`;
|
||||
root.querySelector<HTMLElement>(`[data-spectrum-${prefix}-mean]`)!.textContent = formatScientific(mean);
|
||||
});
|
||||
};
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-spectrum-depth]").forEach((button) => button.addEventListener("click", () => {
|
||||
spectrumDepth = Number(button.dataset.spectrumDepth);
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-spectrum-depth]").forEach((peer) => peer.setAttribute("aria-pressed", String(peer === button)));
|
||||
updateSpectrum();
|
||||
}));
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-spectrum-scale]").forEach((button) => button.addEventListener("click", () => {
|
||||
spectrumScale = button.dataset.spectrumScale || "absolute";
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-spectrum-scale]").forEach((peer) => peer.setAttribute("aria-pressed", String(peer === button)));
|
||||
updateSpectrum();
|
||||
}));
|
||||
spectrumSeed.addEventListener("change", updateSpectrum);
|
||||
spectrumStep.addEventListener("change", updateSpectrum);
|
||||
updateSpectrum();
|
||||
|
||||
let timeDepth = 32;
|
||||
let timeMetric = "cv";
|
||||
const timeSeed = root.querySelector<HTMLSelectElement>("[data-time-seed]")!;
|
||||
const timeChart = root.querySelector<SVGSVGElement>("[data-time-chart]")!;
|
||||
const timeMetricMap: Record<string, [string, string, (value: number) => string]> = {
|
||||
cv: ["activation_grad_statistics", "population_cv", (value) => value.toFixed(2)],
|
||||
imbalance: ["activation_grad_statistics", "imbalance_abs_log_ratio", (value) => value.toFixed(2)],
|
||||
mean: ["activation_grad_statistics", "mean", formatScientific],
|
||||
};
|
||||
const updateTimeline = () => {
|
||||
const seedKey = timeSeed.value;
|
||||
const [group, field, formatter] = timeMetricMap[timeMetric];
|
||||
const series = Object.fromEntries(architectures.map((architecture) => [
|
||||
architecture,
|
||||
steps.map((step: number) => statFor(timeDepth, architecture, seedKey, step, group, field)),
|
||||
]));
|
||||
drawChart(
|
||||
timeChart,
|
||||
null,
|
||||
timeChart.querySelector("[data-time-grid]")!,
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, timeChart.querySelector(`[data-time-line="${architecture}"]`)])),
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, timeChart.querySelector(`[data-time-points="${architecture}"]`)])),
|
||||
series,
|
||||
steps.map((step: number) => step >= 1000 ? `${step / 1000}K` : String(step)),
|
||||
{ formatter },
|
||||
);
|
||||
const titles: Record<string, string> = {
|
||||
cv: "POPULATION CV",
|
||||
imbalance: "ABS(LOG(FIRST / LAST QUARTILE))",
|
||||
mean: "MEAN ACTIVATION GRADIENT RMS",
|
||||
};
|
||||
root.querySelector<HTMLElement>("[data-time-title]")!.textContent = titles[timeMetric];
|
||||
root.querySelector<HTMLElement>("[data-time-copy]")!.textContent = `DEPTH ${timeDepth} · ${seedKey === "mean" ? "3-SEED MEAN" : seedKey}`;
|
||||
};
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-time-depth]").forEach((button) => button.addEventListener("click", () => {
|
||||
timeDepth = Number(button.dataset.timeDepth);
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-time-depth]").forEach((peer) => peer.setAttribute("aria-pressed", String(peer === button)));
|
||||
updateTimeline();
|
||||
}));
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-time-metric]").forEach((button) => button.addEventListener("click", () => {
|
||||
timeMetric = button.dataset.timeMetric || "cv";
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-time-metric]").forEach((peer) => peer.setAttribute("aria-pressed", String(peer === button)));
|
||||
updateTimeline();
|
||||
}));
|
||||
timeSeed.addEventListener("change", updateTimeline);
|
||||
updateTimeline();
|
||||
|
||||
let outputDepth = 32;
|
||||
const outputSeed = root.querySelector<HTMLSelectElement>("[data-output-seed]")!;
|
||||
const outputStep = root.querySelector<HTMLSelectElement>("[data-output-step]")!;
|
||||
const outputChart = root.querySelector<SVGSVGElement>("[data-output-chart]")!;
|
||||
const updateOutput = () => {
|
||||
const seedKey = outputSeed.value;
|
||||
const step = Number(outputStep.value);
|
||||
const series = Object.fromEntries(architectures.map((architecture) => [
|
||||
architecture,
|
||||
arrayFor(outputDepth, architecture, seedKey, step, "activation_output_rms_by_block"),
|
||||
]));
|
||||
drawChart(
|
||||
outputChart,
|
||||
outputChart.querySelector("[data-output-groups]"),
|
||||
outputChart.querySelector("[data-output-grid]")!,
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, outputChart.querySelector(`[data-output-line="${architecture}"]`)])),
|
||||
Object.fromEntries(architectures.map((architecture) => [architecture, outputChart.querySelector(`[data-output-points="${architecture}"]`)])),
|
||||
series,
|
||||
Array.from({ length: outputDepth }, (_, index) => String(index + 1)),
|
||||
{ depth: outputDepth, formatter: (value) => value.toFixed(2) },
|
||||
);
|
||||
root.querySelector<HTMLElement>("[data-output-copy]")!.textContent = `DEPTH ${outputDepth} · ${seedKey === "mean" ? "3-SEED MEAN" : seedKey} · STEP ${formatStep(step)}`;
|
||||
root.querySelector<HTMLElement>("[data-rhythm-title]")!.textContent = `BLOCK ATTNRES · DEPTH ${outputDepth}`;
|
||||
const bars = root.querySelector<HTMLElement>("[data-output-bars]")!;
|
||||
const values: number[] = series.block;
|
||||
const maximum = Math.max(...values);
|
||||
const blocksPerGroup = outputDepth / 8;
|
||||
bars.replaceChildren();
|
||||
values.forEach((value, index) => {
|
||||
const bar = document.createElement("i");
|
||||
bar.style.setProperty("--bar", `${value / maximum * 100}%`);
|
||||
if ((index + 1) % blocksPerGroup === 0) bar.classList.add("boundary");
|
||||
const label = document.createElement("span");
|
||||
label.textContent = String(index + 1);
|
||||
bar.append(label);
|
||||
bars.append(bar);
|
||||
});
|
||||
};
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-output-depth]").forEach((button) => button.addEventListener("click", () => {
|
||||
outputDepth = Number(button.dataset.outputDepth);
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-output-depth]").forEach((peer) => peer.setAttribute("aria-pressed", String(peer === button)));
|
||||
updateOutput();
|
||||
}));
|
||||
outputSeed.addEventListener("change", updateOutput);
|
||||
outputStep.addEventListener("change", updateOutput);
|
||||
updateOutput();
|
||||
};
|
||||
|
||||
document.querySelectorAll<HTMLElement>("[data-gradient-lab]").forEach(initializeGradientLab);
|
||||
document.addEventListener("astro:page-load", () => {
|
||||
document.querySelectorAll<HTMLElement>("[data-gradient-lab]").forEach(initializeGradientLab);
|
||||
});
|
||||
</script>
|
||||
|
||||
<style>
|
||||
.gradient-lab {
|
||||
--g-ink: #1c201e;
|
||||
--g-muted: #77746b;
|
||||
--g-line: rgba(28, 32, 30, .16);
|
||||
--g-paper: #f4f0e7;
|
||||
--g-raised: #faf7ef;
|
||||
--g-copper: #ba603b;
|
||||
--g-green: #163f3b;
|
||||
width: min(1120px, 100%);
|
||||
margin: 42px 0;
|
||||
color: var(--g-ink);
|
||||
border: 1px solid var(--g-line);
|
||||
background: var(--g-paper);
|
||||
box-shadow: 0 30px 80px rgba(28, 32, 30, .09);
|
||||
}
|
||||
.gradient-head {
|
||||
display: grid;
|
||||
grid-template-columns: minmax(0, 1.45fr) minmax(260px, .7fr);
|
||||
gap: 44px;
|
||||
padding: 30px;
|
||||
color: #f5efe4;
|
||||
background: var(--g-green);
|
||||
}
|
||||
.gradient-head p { margin: 0; color: rgba(245,239,228,.7); font: .65rem/1.7 var(--mono); }
|
||||
.gradient-head div > p { color: #d58a68; letter-spacing: .08em; }
|
||||
.gradient-head h3 { max-width: 720px; margin: 14px 0 0; color: inherit; font-size: clamp(1.15rem, 2.2vw, 1.75rem); line-height: 1.35; }
|
||||
.gradient-ledger { display: grid; grid-template-columns: repeat(6, 1fr); border-bottom: 1px solid var(--g-line); }
|
||||
.gradient-ledger article { min-height: 126px; padding: 18px 15px; border-right: 1px solid var(--g-line); }
|
||||
.gradient-ledger article:last-child { border-right: 0; }
|
||||
.gradient-ledger span, .panel-lead span { color: var(--g-muted); font: .56rem/1.2 var(--mono); letter-spacing: .08em; }
|
||||
.gradient-ledger b { display: block; margin-top: 23px; font: 700 .95rem/1 var(--mono); }
|
||||
.gradient-ledger p { margin: 8px 0 0; color: var(--g-muted); font-size: .6rem; line-height: 1.45; }
|
||||
.gradient-ledger .pass { color: #f7f0e6; background: var(--g-green); }
|
||||
.gradient-ledger .split { color: #f7f0e6; background: var(--g-copper); }
|
||||
.gradient-ledger .pass span, .gradient-ledger .pass p, .gradient-ledger .split span, .gradient-ledger .split p { color: rgba(247,240,230,.72); }
|
||||
.gradient-tabs { display: grid; grid-template-columns: repeat(5, 1fr); border-bottom: 1px solid var(--g-line); background: #e9e4da; }
|
||||
.gradient-tabs button { min-height: 116px; padding: 16px; text-align: left; color: inherit; border: 0; border-right: 1px solid var(--g-line); background: transparent; cursor: pointer; }
|
||||
.gradient-tabs button:last-child { border-right: 0; }
|
||||
.gradient-tabs button[aria-selected="true"] { color: #f7f0e6; background: var(--g-copper); }
|
||||
.gradient-tabs span, .gradient-tabs small { display: block; color: var(--g-muted); font: .54rem/1.25 var(--mono); }
|
||||
.gradient-tabs b { display: block; margin: 15px 0 8px; font-size: .69rem; line-height: 1.35; }
|
||||
.gradient-tabs button[aria-selected="true"] span, .gradient-tabs button[aria-selected="true"] small { color: rgba(247,240,230,.72); }
|
||||
.gradient-panel { padding: 30px; }
|
||||
.panel-lead { display: grid; grid-template-columns: 1.05fr .95fr; gap: 48px; align-items: end; margin-bottom: 28px; }
|
||||
.panel-lead h4 { max-width: 680px; margin: 10px 0 0; font-size: 1.2rem; line-height: 1.4; }
|
||||
.panel-lead p { margin: 0; color: var(--g-muted); font-size: .7rem; line-height: 1.7; }
|
||||
.known-grid { display: grid; grid-template-columns: repeat(2, 1fr); border: 1px solid var(--g-line); }
|
||||
.known-grid article { min-height: 270px; padding: 24px; }
|
||||
.known-grid article + article { border-left: 1px solid var(--g-line); background: #ece2d6; }
|
||||
.known-grid span, .object-chain span, .object-compare span, .split-result span, .timeline-notes span, .output-ratios span, .metric-pairs span, .cost-compare > article > span, .replay-ledger span, .hash-ledger span, .claim-grid span {
|
||||
color: var(--g-copper); font: .56rem/1 var(--mono); letter-spacing: .06em;
|
||||
}
|
||||
.known-grid h5 { margin: 26px 0 16px; font-size: .95rem; }
|
||||
.known-grid ul { margin: 0; padding-left: 18px; }
|
||||
.known-grid li { margin-top: 11px; color: var(--g-muted); font-size: .68rem; line-height: 1.55; }
|
||||
.object-chain { display: grid; grid-template-columns: 1fr 30px 1fr 30px .8fr 30px 1.2fr; gap: 6px; align-items: center; margin-top: 20px; }
|
||||
.object-chain div { min-height: 145px; padding: 18px; border: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.object-chain .accent { color: #f7f0e6; background: var(--g-green); }
|
||||
.object-chain .accent span, .object-chain .accent p { color: rgba(247,240,230,.68); }
|
||||
.object-chain b { display: block; margin-top: 25px; font: 700 .77rem/1.35 var(--mono); }
|
||||
.object-chain p { color: var(--g-muted); font-size: .61rem; line-height: 1.45; }
|
||||
.object-chain > i, .object-compare > i, .split-result > i, .replay-ledger > i { color: var(--g-copper); font-style: normal; text-align: center; }
|
||||
.object-compare { display: grid; grid-template-columns: 1fr 50px 1fr; gap: 12px; align-items: center; margin-top: 20px; padding: 20px; background: #e9e4da; }
|
||||
.object-compare article { padding: 12px; }
|
||||
.object-compare b { display: block; margin-top: 18px; font: 700 1.1rem/1 var(--mono); }
|
||||
.object-compare p { color: var(--g-muted); font-size: .65rem; line-height: 1.55; }
|
||||
.object-compare > p { grid-column: 1/-1; margin: 0; padding-top: 16px; border-top: 1px solid var(--g-line); }
|
||||
.definition-boundary { display: grid; grid-template-columns: 80px 1fr; margin-top: 20px; border-top: 1px solid var(--g-line); }
|
||||
.definition-boundary > * { margin: 0; padding: 15px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); }
|
||||
.definition-boundary b { color: var(--g-copper); font: .6rem/1.4 var(--mono); }
|
||||
.definition-boundary p { color: var(--g-muted); font-size: .66rem; line-height: 1.55; }
|
||||
.lab-controls { display: flex; flex-wrap: wrap; gap: 10px 18px; align-items: end; margin-bottom: 20px; }
|
||||
.lab-controls > div { display: flex; }
|
||||
.lab-controls button, .lab-controls select { min-height: 38px; padding: 10px 12px; color: var(--g-muted); font: 700 .56rem/1 var(--mono); border: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.lab-controls button { cursor: pointer; }
|
||||
.lab-controls button + button { border-left: 0; }
|
||||
.lab-controls button[aria-pressed="true"] { color: #fff9ef; background: var(--g-green); }
|
||||
.lab-controls label { display: grid; gap: 6px; color: var(--g-muted); font: .52rem/1 var(--mono); }
|
||||
.spectrum-layout { display: grid; grid-template-columns: minmax(0, 1fr) 235px; border: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.chart-shell { min-width: 0; padding: 18px; border: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.spectrum-layout .chart-shell { border: 0; border-right: 1px solid var(--g-line); }
|
||||
.chart-shell header { display: flex; justify-content: space-between; gap: 12px; color: var(--g-muted); font: .55rem/1 var(--mono); }
|
||||
.chart-shell svg { display: block; width: 100%; height: auto; margin-top: 10px; overflow: visible; }
|
||||
.series { fill: none; stroke-width: 3; stroke-linejoin: round; stroke-linecap: round; }
|
||||
.series.baseline { stroke: #77746b; }
|
||||
.series.block { stroke: #ba603b; }
|
||||
.grid-line { stroke: rgba(28,32,30,.1); stroke-width: 1; }
|
||||
.group-boundary { stroke: rgba(22,63,59,.22); stroke-width: 1.5; stroke-dasharray: 4 4; }
|
||||
.axis-label { fill: #8b867c; font: 11px var(--mono); }
|
||||
.axis-label.y { text-anchor: end; }
|
||||
.axis-label.x { text-anchor: middle; }
|
||||
.axis-title { fill: #8b867c; font: 11px var(--mono); text-anchor: middle; }
|
||||
.chart-legend { display: flex; flex-wrap: wrap; gap: 18px; margin-top: 4px; color: var(--g-muted); font: .56rem/1 var(--mono); }
|
||||
.chart-legend span { display: inline-flex; gap: 7px; align-items: center; }
|
||||
.chart-legend i { width: 22px; height: 3px; }
|
||||
.chart-legend i.baseline { background: #77746b; }
|
||||
.chart-legend i.block { background: #ba603b; }
|
||||
.chart-legend i.boundary { height: 0; border-top: 2px dashed var(--g-green); background: transparent; }
|
||||
.spectrum-readout { padding: 20px 18px; }
|
||||
.spectrum-readout > span { color: var(--g-copper); font: .54rem/1.4 var(--mono); }
|
||||
.spectrum-readout article { padding: 18px 0; border-bottom: 1px solid var(--g-line); }
|
||||
.spectrum-readout article b { color: var(--g-muted); font: .52rem/1 var(--mono); }
|
||||
.spectrum-readout article div { display: flex; justify-content: space-between; margin-top: 13px; }
|
||||
.spectrum-readout article span { color: var(--g-muted); font: .5rem/1 var(--mono); }
|
||||
.spectrum-readout article strong { font: 700 .7rem/1 var(--mono); }
|
||||
.split-result { display: grid; grid-template-columns: 1fr 45px 1fr 45px 1fr; gap: 8px; align-items: center; margin-top: 20px; }
|
||||
.split-result article { min-height: 145px; padding: 20px; border: 1px solid var(--g-line); }
|
||||
.split-result .accent { color: #f7f0e6; background: var(--g-green); }
|
||||
.split-result .accent span, .split-result .accent p { color: rgba(247,240,230,.68); }
|
||||
.split-result b { display: block; margin-top: 25px; font-size: .83rem; }
|
||||
.split-result p { color: var(--g-muted); font-size: .62rem; line-height: 1.5; }
|
||||
.timeline-chart { margin-top: 0; }
|
||||
.timeline-notes { display: grid; grid-template-columns: repeat(4, 1fr); margin-top: 20px; border-top: 1px solid var(--g-line); border-left: 1px solid var(--g-line); }
|
||||
.timeline-notes article { min-height: 145px; padding: 18px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.timeline-notes .counter { background: #e9e4da; }
|
||||
.timeline-notes b { display: block; margin-top: 25px; font: 700 .78rem/1 var(--mono); }
|
||||
.timeline-notes p { color: var(--g-muted); font-size: .61rem; line-height: 1.5; }
|
||||
.output-ratios { display: grid; grid-template-columns: 1fr 1fr 1.25fr; margin-top: 20px; border-top: 1px solid var(--g-line); border-left: 1px solid var(--g-line); }
|
||||
.output-ratios article { min-height: 170px; padding: 20px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); }
|
||||
.output-ratios article > div { display: flex; justify-content: space-between; margin-top: 24px; }
|
||||
.output-ratios article > div b { color: var(--g-muted); font: .55rem/1 var(--mono); }
|
||||
.output-ratios article > div strong { font: 700 .75rem/1 var(--mono); }
|
||||
.output-ratios .accent { color: #f7f0e6; background: var(--g-green); }
|
||||
.output-ratios .accent span, .output-ratios .accent p { color: rgba(247,240,230,.68); }
|
||||
.output-ratios .accent b { display: block; margin-top: 25px; font-size: .8rem; }
|
||||
.output-ratios p { color: var(--g-muted); font-size: .62rem; line-height: 1.5; }
|
||||
.group-rhythm { margin-top: 20px; padding: 20px; color: #f7f0e6; background: var(--g-green); overflow: hidden; }
|
||||
.group-rhythm header { display: flex; justify-content: space-between; color: rgba(247,240,230,.68); font: .55rem/1 var(--mono); }
|
||||
.group-rhythm > div { display: grid; grid-template-columns: repeat(auto-fit, minmax(8px, 1fr)); align-items: end; height: 170px; margin-top: 18px; border-bottom: 1px solid rgba(255,255,255,.3); }
|
||||
.group-rhythm i { position: relative; display: block; height: max(5px, var(--bar)); margin-right: 2px; background: #d4835d; }
|
||||
.group-rhythm i.boundary { margin-right: 8px; border-right: 2px solid rgba(255,255,255,.75); }
|
||||
.group-rhythm i span { position: absolute; bottom: -18px; left: 50%; color: rgba(255,255,255,.55); font: .43rem/1 var(--mono); transform: translateX(-50%); }
|
||||
.group-rhythm p { margin: 35px 0 0; color: rgba(247,240,230,.72); font-size: .64rem; }
|
||||
.verdict-table-wrap { overflow-x: auto; }
|
||||
.verdict-table { width: 100%; min-width: 850px; border-collapse: collapse; font-size: .62rem; }
|
||||
.verdict-table th, .verdict-table td { padding: 15px 12px; border-bottom: 1px solid var(--g-line); text-align: left; }
|
||||
.verdict-table thead th { color: var(--g-muted); font: .53rem/1.3 var(--mono); }
|
||||
.verdict-table tbody th, .verdict-table tbody td:last-child { font: 700 .66rem/1 var(--mono); }
|
||||
.verdict-table .good { color: var(--g-green); }
|
||||
.verdict-table .bad { color: var(--g-copper); }
|
||||
.metric-pairs { display: grid; grid-template-columns: repeat(4, 1fr); margin-top: 20px; border-top: 1px solid var(--g-line); border-left: 1px solid var(--g-line); }
|
||||
.metric-pairs article { min-height: 150px; padding: 18px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.metric-pairs article.warn { border-top: 4px solid var(--g-copper); }
|
||||
.metric-pairs article.pass { border-top: 4px solid var(--g-green); }
|
||||
.metric-pairs b { display: block; margin-top: 25px; font-size: .77rem; }
|
||||
.metric-pairs p { color: var(--g-muted); font-size: .61rem; line-height: 1.5; }
|
||||
.cost-compare { display: grid; grid-template-columns: repeat(3, 1fr); margin-top: 20px; border-top: 1px solid var(--g-line); border-left: 1px solid var(--g-line); }
|
||||
.cost-compare article { min-height: 205px; padding: 20px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); }
|
||||
.cost-compare article > div { display: flex; justify-content: space-between; gap: 12px; margin-top: 25px; }
|
||||
.cost-compare article > div b { color: var(--g-muted); font: .52rem/1 var(--mono); }
|
||||
.cost-compare article > div strong { font: 700 .62rem/1 var(--mono); text-align: right; }
|
||||
.cost-compare p { color: var(--g-copper); font: .56rem/1.5 var(--mono); }
|
||||
.cost-compare .boundary { color: #f7f0e6; background: var(--g-green); }
|
||||
.cost-compare .boundary span, .cost-compare .boundary p { color: rgba(247,240,230,.68); }
|
||||
.cost-compare .boundary b { display: block; margin-top: 30px; font-size: .85rem; }
|
||||
.replay-ledger { display: grid; grid-template-columns: 1fr 40px 1fr 40px 1fr; gap: 8px; align-items: center; margin-top: 20px; }
|
||||
.replay-ledger article { min-height: 145px; padding: 18px; border: 1px solid var(--g-line); background: var(--g-raised); }
|
||||
.replay-ledger article.pass { color: #f7f0e6; background: var(--g-green); }
|
||||
.replay-ledger article.pass span, .replay-ledger article.pass p { color: rgba(247,240,230,.68); }
|
||||
.replay-ledger b { display: block; margin-top: 25px; font-size: .7rem; }
|
||||
.replay-ledger p { color: var(--g-muted); font-size: .58rem; line-height: 1.5; overflow-wrap: anywhere; }
|
||||
.hash-ledger { display: grid; grid-template-columns: repeat(2, 1fr); margin-top: 20px; border-top: 1px solid var(--g-line); border-left: 1px solid var(--g-line); }
|
||||
.hash-ledger article { min-width: 0; padding: 16px; border-right: 1px solid var(--g-line); border-bottom: 1px solid var(--g-line); background: #e9e4da; }
|
||||
.hash-ledger code { display: block; margin-top: 12px; overflow: hidden; color: var(--g-green); font: .54rem/1.3 var(--mono); text-overflow: ellipsis; }
|
||||
.claim-grid { display: grid; grid-template-columns: repeat(2, 1fr); margin-top: 20px; }
|
||||
.claim-grid article { padding: 22px; }
|
||||
.claim-grid .yes { color: #f7f0e6; background: var(--g-green); }
|
||||
.claim-grid .no { background: #e7d8ca; }
|
||||
.claim-grid ul { margin: 18px 0 0; padding-left: 17px; }
|
||||
.claim-grid li { margin-top: 10px; font-size: .65rem; line-height: 1.55; }
|
||||
.claim-grid .yes span, .claim-grid .yes li { color: rgba(247,240,230,.78); }
|
||||
@media (max-width: 920px) {
|
||||
.gradient-ledger { grid-template-columns: repeat(3, 1fr); }
|
||||
.gradient-ledger article:nth-child(3) { border-right: 0; }
|
||||
.gradient-tabs { grid-template-columns: repeat(3, 1fr); }
|
||||
.spectrum-layout { grid-template-columns: 1fr; }
|
||||
.spectrum-layout .chart-shell { border-right: 0; border-bottom: 1px solid var(--g-line); }
|
||||
.timeline-notes, .metric-pairs { grid-template-columns: repeat(2, 1fr); }
|
||||
.object-chain { grid-template-columns: 1fr 24px 1fr; }
|
||||
.object-chain > i:nth-of-type(n+3) { display: none; }
|
||||
}
|
||||
@media (max-width: 680px) {
|
||||
.gradient-head, .panel-lead, .known-grid { grid-template-columns: 1fr; gap: 20px; }
|
||||
.gradient-head, .gradient-panel { padding: 20px; }
|
||||
.known-grid article + article { border-left: 0; border-top: 1px solid var(--g-line); }
|
||||
.gradient-ledger { grid-template-columns: repeat(2, 1fr); }
|
||||
.gradient-ledger article:nth-child(3) { border-right: 1px solid var(--g-line); }
|
||||
.gradient-ledger article:nth-child(even) { border-right: 0; }
|
||||
.gradient-tabs { display: flex; overflow-x: auto; }
|
||||
.gradient-tabs button { flex: 0 0 190px; }
|
||||
.lab-controls { align-items: stretch; }
|
||||
.lab-controls > div, .lab-controls label { flex: 1 0 100%; }
|
||||
.lab-controls button { flex: 1; }
|
||||
.lab-controls select { width: 100%; }
|
||||
.object-chain, .object-compare, .split-result, .replay-ledger { grid-template-columns: 1fr; }
|
||||
.object-chain > i, .object-compare > i, .split-result > i, .replay-ledger > i { display: block !important; transform: rotate(90deg); }
|
||||
.object-compare > p { grid-column: 1; }
|
||||
.definition-boundary { grid-template-columns: 70px 1fr; }
|
||||
.timeline-notes, .output-ratios, .metric-pairs, .cost-compare, .hash-ledger, .claim-grid { grid-template-columns: 1fr; }
|
||||
.chart-shell { padding: 12px 8px; }
|
||||
.chart-shell header { padding: 0 8px; }
|
||||
.axis-label { font-size: 9px; }
|
||||
.group-rhythm > div { min-width: 620px; }
|
||||
.group-rhythm { overflow-x: auto; }
|
||||
}
|
||||
</style>
|
||||
@@ -2,6 +2,7 @@
|
||||
import BaseLayout from "@/layouts/BaseLayout.astro";
|
||||
import ArchitectureExplorer from "@/components/ArchitectureExplorer.astro";
|
||||
import K3ArtifactLab from "@/components/K3ArtifactLab.astro";
|
||||
import K3AttnResGradientLab from "@/components/K3AttnResGradientLab.astro";
|
||||
import K3AttnResTraceLab from "@/components/K3AttnResTraceLab.astro";
|
||||
import K3ReportLab from "@/components/K3ReportLab.astro";
|
||||
import { k3FigureAtlas, k3Ledgers, k3PaperChain, k3ReportMap } from "@/data/k3";
|
||||
@@ -38,7 +39,8 @@ const toc = [
|
||||
["28", "lab", "八联交互实验"],
|
||||
["29", "artifacts", "开放权重工件审计"],
|
||||
["30", "attnres-reduced", "AttnRes 缩小机制实验"],
|
||||
["31", "audit", "21 张图表审计"],
|
||||
["31", "attnres-gradient", "梯度定义与深度扩展"],
|
||||
["32", "audit", "21 张图表审计"],
|
||||
["↳", "papers", "100 节点阅读链"],
|
||||
];
|
||||
|
||||
@@ -107,13 +109,13 @@ const paperGroups = [
|
||||
|
||||
<BaseLayout
|
||||
title="Kimi K3 技术报告完整深读：架构、训练、RL、系统与评测"
|
||||
description="用三十二张问题账、二十一张图表审计、八个机制实验、四个开放工件视图、五个 AttnRes 独立实验视图与一百个一手阅读节点，逐节读懂 Kimi K3。"
|
||||
description="用三十二张问题账、二十一张图表审计、八个机制实验、四个开放工件视图、两轮十个 AttnRes 独立实验视图与一百个一手阅读节点，逐节读懂 Kimi K3。"
|
||||
section="k3"
|
||||
>
|
||||
<header class="page-hero k3-hero">
|
||||
<div class="page-hero-inner">
|
||||
<div>
|
||||
<p class="eyebrow"><span>ANCHOR REPORT / ROUND 04</span> KIMI K3 · REPORT → ARTIFACTS → INDEPENDENT PROBE</p>
|
||||
<p class="eyebrow"><span>ANCHOR REPORT / ROUND 05</span> KIMI K3 · REPORT → ARTIFACTS → INDEPENDENT PROBE</p>
|
||||
<h1>不把报告压成摘要<br />把每个因果环节<br />重新展开</h1>
|
||||
<p class="lead">
|
||||
K3 同时扩展序列、深度、宽度、视觉与 Agent 轨迹。真正值得读的不是 2.8T 这个最大数字，
|
||||
@@ -123,11 +125,11 @@ const paperGroups = [
|
||||
<dl class="page-facts">
|
||||
<div><dt>QUESTIONS</dt><dd>32 张问题账</dd></div>
|
||||
<div><dt>REPORT</dt><dd>16 Figures · 5 Tables</dd></div>
|
||||
<div><dt>LABS</dt><dd>8 + 4 + 5 个交互视图</dd></div>
|
||||
<div><dt>LABS</dt><dd>8 + 4 + 5 + 5 个交互视图</dd></div>
|
||||
<div><dt>READING</dt><dd>100 个一手 / 官方节点</dd></div>
|
||||
<div><dt>MODEL</dt><dd>2.78T total / 104.2B active</dd></div>
|
||||
<div><dt>ARTIFACTS</dt><dd>96 shards · 497,220 tensors</dd></div>
|
||||
<div><dt>STATUS</dt><dd>K3 四轮 · AttnRes 实验</dd></div>
|
||||
<div><dt>STATUS</dt><dd>K3 五轮 · 梯度定义闭环</dd></div>
|
||||
</dl>
|
||||
</div>
|
||||
</header>
|
||||
@@ -892,12 +894,36 @@ const paperGroups = [
|
||||
<a class="button primary" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_REDUCED_AUDIT.md">阅读完整研究审计</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/experiments/k3/attnres">复跑公开实验代码</a>
|
||||
<a class="button" href="https://arxiv.org/abs/2603.15031">Attention Residuals 原论文</a>
|
||||
<a class="button" href="https://github.com/MoonshotAI/Attention-Residuals">官方实现</a>
|
||||
<a class="button" href="https://github.com/MoonshotAI/Attention-Residuals">官方论文工件</a>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="article-section" id="attnres-gradient">
|
||||
<p class="eyebrow"><span>31</span> GRADIENT DEFINITION × DEPTH SCALE</p>
|
||||
<h2>“论文说梯度更均匀”，和上一轮参数梯度反结果，测的是同一件事吗？</h2>
|
||||
<p class="lede">
|
||||
第五轮先审计 Attention Residuals 官方论文与仓库：Figure 5(c) 没有公开 gradient tensor、
|
||||
norm、reduction、diagnostic batch、AMP / clipping 时点或统计代码。本站因此冻结一个可复现的
|
||||
post-MLP output activation-gradient 定义，把深度扩到 16 / 32 blocks、预算扩到 8,000 steps，
|
||||
再用三 seed 检查“首尾平衡”和“全层离散度”是否真的同方向。
|
||||
</p>
|
||||
<div class="artifact-callout">
|
||||
<article><span>F / FROZEN</span><b>12 × 8,000 steps</b><p>786,432,000 formal target bytes；两深度、两结构、三 seed。</p></article>
|
||||
<article><span>X / OBSERVED</span><b>first/last 6 / 6 改善</b><p>depth-16 平均 61.0%；depth-32 平均 72.0%。</p></article>
|
||||
<article class="warning"><span>X / COUNTEREVIDENCE</span><b>CV 6 / 6 恶化</b><p>局部尖峰让 depth-16 / 32 平均相对恶化 10.3% / 60.0%。</p></article>
|
||||
<article><span>R / REPLAY</span><b>model + optimizer exact</b><p>指定 32-layer Block 格从零重训 8,000 steps，冻结字段逐项一致。</p></article>
|
||||
</div>
|
||||
<K3AttnResGradientLab />
|
||||
<div class="hero-actions">
|
||||
<a class="button primary" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_GRADIENT_SCALE_AUDIT.md">阅读完整结果审计</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_GRADIENT_DEFINITION_AUDIT.md">核对论文定义边界</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/experiments/k3/attnres_gradient">复跑 12 格实验</a>
|
||||
<a class="button" href="https://github.com/MoonshotAI/Attention-Residuals">官方一手工件</a>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="article-section" id="audit">
|
||||
<p class="eyebrow"><span>31</span> FIGURE & TABLE AUDIT</p>
|
||||
<p class="eyebrow"><span>32</span> FIGURE & TABLE AUDIT</p>
|
||||
<h2>Figure 1–16、Table 1–5：每张图究竟支持什么，不能支持什么</h2>
|
||||
<div class="figure-atlas">
|
||||
{k3FigureAtlas.map(([id, report, title, contract]) => (
|
||||
|
||||
@@ -9,7 +9,7 @@ const researching = chapters.filter((chapter) => ["researching", "drafting"].inc
|
||||
const workstreams = [
|
||||
{ label: "研究框架与规范", value: 83, next: "给 Scaling 与推理专题补逐篇图表/实验精读层级" },
|
||||
{ label: "网站设计系统", value: 89, next: "打印样式与更多通用可视化组件" },
|
||||
{ label: "Kimi K3 深读", value: 96, next: "对齐 AttnRes 梯度定义并扩展深度/预算；等待 A_log 官方转换合同" },
|
||||
{ label: "Kimi K3 深读", value: 98, next: "对齐 layer 21–25 梯度尖峰与 mixer weights；等待 A_log 官方转换合同" },
|
||||
{ label: "语言模型前史", value: 78, next: "逐图精读 Kneser–Ney、LSTM 与 Bahdanau，并加入真实小语料复现" },
|
||||
{ label: "Transformer 基础", value: 79, next: "逐图精读多头电路、Pre/Post-LN 与真实 kernel / KV 配置" },
|
||||
{ label: "表示、位置与残差高速公路", value: 81, next: "加入真实 hidden-state / norm traces、长上下文位置外推复现与更多深层稳定性消融" },
|
||||
@@ -50,7 +50,7 @@ const workstreams = [
|
||||
<div><dt>OVERALL</dt><dd>专题平均 {average}%</dd></div>
|
||||
<div><dt>READABLE</dt><dd>{published} 个首版可读专题</dd></div>
|
||||
<div><dt>ACTIVE</dt><dd>{researching} 个研究/写作中</dd></div>
|
||||
<div><dt>UPDATED</dt><dd>2026-07-30 07:30 CST</dd></div>
|
||||
<div><dt>UPDATED</dt><dd>2026-07-30 10:05 CST</dd></div>
|
||||
<div><dt>MODE</dt><dd>持续迭代，不锁死版本</dd></div>
|
||||
</dl>
|
||||
</div>
|
||||
@@ -97,7 +97,7 @@ const workstreams = [
|
||||
<article><span>✓</span><h3>K3 报告已结构化拆解</h3><p>47 页报告目录、151 条参考来源和架构/后训练/系统主线已经提取。</p></article>
|
||||
<article><span>✓</span><h3>17 专题知识图</h3><p>从语言模型基础到评测安全，包含先修依赖和三条贯穿案例。</p></article>
|
||||
<article><span>✓</span><h3>编辑式网站系统</h3><p>响应式导航、章节模板、侧栏、进度、论文链和证据提示组件。</p></article>
|
||||
<article><span>✓</span><h3>九十四个原创交互视图</h3><p>K3 三轴图、八联报告实验、四联开放工件实验与五联 AttnRes 独立实验，DeepSeek 四联公式实验、十三联 Base 工件实验、Chat 行为、completion/full-depth、multi-seed、cross-source 与 task-bootstrap CRN 五轮实验，以及语言模型前史、Transformer、表示深度、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>九十九个原创交互视图</h3><p>K3 三轴图、八联报告实验、四联开放工件实验与两轮十联 AttnRes 独立实验，DeepSeek 四联公式实验、十三联 Base 工件实验、Chat 行为、completion/full-depth、multi-seed、cross-source 与 task-bootstrap CRN 五轮实验，以及语言模型前史、Transformer、表示深度、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>十七篇首版长文</h3><p>K3、语言模型前史、Transformer、表示/位置/残差、DeepSeek、Scaling、数据工程、长上下文、MoE、后训练、推理、Agent、原生多模态、训练系统、推理服务、数值优化与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>语言模型前史深度专题</h3><p>八张独立问题账、33 个正式节点、20 段长文与概率—向量—记忆—对齐四联实验。</p></article>
|
||||
<article><span>✓</span><h3>Transformer 深度专题</h3><p>十张独立问题账、40 个正式节点、21 段正文与 QKV—Mask—多头位置—Block 成本四联实验。</p></article>
|
||||
@@ -106,6 +106,7 @@ const workstreams = [
|
||||
<article><span>✓</span><h3>Kimi K3 技术报告二轮深读</h3><p>三十二张问题账、Figure 1–16 / Table 1–5 审计、100 节点阅读链，以及 Delta—Decay—AttnRes—LatentMoE—SiTU—QB—MOPD—Cache 八联实验。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 三轮开放工件里程碑</h3><p>固定官方 revisions，审计 96 个 shards、497,220 个 tensor entries 与真实 KDA / MLA / MoE / MoonViT shapes；四联实验分开显示层型、tensor anatomy、参数范围和复现边界。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 四轮 AttnRes 独立实验</h3><p>冻结三结构 × 三 seed 的 9 个 2,000-step 格；Full / Block 相对 Baseline 的平均 paired delta 为 −0.01457 / −0.04247 BPC，但核心参数梯度 CV 没有复现论文叙述。指定正式格全新进程八字段 exact，五视图同时展示结果、反证、成本与 claim boundary。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 五轮梯度定义与深度扩展</h3><p>先确认 Figure 5 没有公开唯一 gradient telemetry 合同，再冻结 16/32 blocks × Baseline/Block × 3 seeds 的 12 个 8,000-step 格。Block 的首尾失衡 6/6 改善但全层 CV 6/6 恶化，两个深度都判为 mixed；指定 32 层格完整重训的模型、优化器与全部冻结字段 exact。</p></article>
|
||||
<article><span>✓</span><h3>FlashKDA RTX 5090 执行闸门</h3><p>隔离 CUDA 13.0 / glibc 2.39 编译 sm_120a wheel；6/6 官方参考逐元素相等，并完成 fixed / varlen、三种 state mode 的 1,800 个 CUDA Event samples。</p></article>
|
||||
<article><span>✓</span><h3>Scaling Laws 深度专题</h3><p>九张账、29 个一手节点、DeepSeek/Kimi 双谱系与曲面—部署—复用—涌现四联实验。</p></article>
|
||||
<article><span>✓</span><h3>数据工程深度专题</h3><p>十二张账、31 个一手节点、DeepSeek/Kimi 双谱系与流水线—去重—混合—改写四联实验。</p></article>
|
||||
@@ -134,7 +135,7 @@ const workstreams = [
|
||||
</div>
|
||||
<div class="queue-table">
|
||||
<div class="head"><b>优先级</b><b>专题</b><b>本轮交付</b><b>完成闸门</b></div>
|
||||
<div><span>P0</span><strong>K3 四轮后续</strong><p>对齐论文梯度定义 → 增加 depth / budget → 等待 A_log 官方合同后进入真实 checkpoint forward</p><em>尺度复查 + 工件边界</em></div>
|
||||
<div><span>P0</span><strong>K3 五轮后续</strong><p>对齐 layer 21–25 尖峰、pre-attention / pre-MLP 与 mixer source weights → 等待 A_log 官方合同后进入真实 checkpoint forward</p><em>局部机制 + 工件边界</em></div>
|
||||
<div><span>P0</span><strong>DeepSeek 八轮后续</strong><p>干预式 mediation → SM90 FlashMLA / FP8 / pipeline traces → R1-like RL 小模型复现</p><em>运行证据 + 独立复现</em></div>
|
||||
<div><span>P0</span><strong>Transformer 二轮</strong><p>多头电路逐图 → Pre/Post-LN 真实 traces → Flash/KV 配置与 kernel 对照</p><em>逐图笔记 + 实测边界</em></div>
|
||||
<div><span>P0</span><strong>表示、位置与残差二轮</strong><p>真实 hidden-state / norm traces → 长上下文位置外推 → mHC / AttnRes 深层稳定性消融</p><em>可复现实验 + 逐图笔记</em></div>
|
||||
@@ -220,6 +221,10 @@ const workstreams = [
|
||||
<div><time>2026-07-30</time><b>AttnRes 缩小实验先冻结、后运行</b><p>三结构共享公共主干、初始化、窗口与优化器；只按三个 paired seed 和预注册 −0.010 BPC 阈值给出本协议内方向判断。</p></div>
|
||||
<div><time>2026-07-30</time><b>支持结果与梯度反结果同时进入主视区</b><p>Full / Block 的最终 BPC 同向改善；核心参数 gradient RMS CV 却高于 Baseline，不换指标掩盖。</p></div>
|
||||
<div><time>2026-07-30</time><b>正式重放不把 wall time 纳入 exact</b><p>Block / seed-1 的模型、优化器、曲线、历史、诊断和环境八字段 exact；计时受调度影响，单独报告。</p></div>
|
||||
<div><time>2026-07-30</time><b>Figure 5 的“梯度”不再靠猜测补合同</b><p>官方未公开 gradient tensor、norm、reduction 与统计代码；本站 activation-gradient 定义只叫 operationalization，不叫论文复画。</p></div>
|
||||
<div><time>2026-07-30</time><b>首尾平衡与全层 CV 永久分账</b><p>Block 在 6/6 配对中改善 first/last，却因中后段局部尖峰让 CV 在 6/6 配对中恶化；联合判定保持 mixed。</p></div>
|
||||
<div><time>2026-07-30</time><b>绝对梯度尺度必须与归一化谱同屏</b><p>Block mean gradient 约为 Baseline 的 54%–57%；更接近 1 的首尾比不能偷换成各层信号更强。</p></div>
|
||||
<div><time>2026-07-30</time><b>32 层完整重放扩到状态哈希</b><p>8,000-step fresh replay 的全部冻结字段以及 model / optimizer state hashes exact；额外 replay bytes 单列，不混入 formal 预算。</p></div>
|
||||
<div><time>2026-07-29</time><b>32-token 对照改为同源 16→24</b><p>TNEWS 只有 105/10,000 条达到 32 tokens，强行统一会落入约 1% 极端长尾；24-token eligibility 仍保留 1,609 条中文候选。</p></div>
|
||||
<div><time>2026-07-29</time><b>长度敏感性必须成对重采样</b><p>16-token 输入严格是 24-token 输入前缀，2,000 次 bootstrap 共用 prompt indices；结果只描述固定 cohort 的长度敏感性。</p></div>
|
||||
<div><time>2026-07-29</time><b>三类 cohort 永久分身份</b><p>自然长度回答本批样本如何路由；matched-16 / 24 回答同一 prompt 多看 8 tokens 后如何变化，不把二者混成内容因果。</p></div>
|
||||
|
||||
Reference in New Issue
Block a user