feat: execute DeepSeek absorbed MLA cache

src/components/DeepSeekArtifactLab.astro

@@ -1,9 +1,18 @@
 ---
 import rawTrace from "@/data/deepseek-v2-lite-trace.json";
 import rawRepro from "@/data/deepseek-v2-lite-trace-repro.json";
+import rawAbsorb from "@/data/deepseek-v2-lite-absorb.json";
+import rawAbsorbRepro from "@/data/deepseek-v2-lite-absorb-repro.json";
 
 const trace = rawTrace as any;
 const repro = rawRepro as any;
+const absorb = rawAbsorb as any;
+const absorbRepro = rawAbsorbRepro as any;
+const absorbExact = JSON.stringify(absorb) === JSON.stringify(absorbRepro);
+const bytes = (value: number) => value >= 1024
+  ? `${(value / 1024).toFixed(2)} KiB`
+  : `${value.toLocaleString()} B`;
+const scientific = (value: number) => value.toExponential(2);
 const compact = {
   prompts: trace.prompts.map((prompt: any) => ({
     id: prompt.id,
@@ -46,7 +55,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     </div>
     <p>
       固定官方 revision、tokenizer、模型代码和 BF16 第一分片；RTX 5090 连续执行 layer 0–6，
-      捕获真实 MLA 状态与 3,240 次 routed-expert 选择。所有结论都带证据身份与停止线。
+      捕获 3,240 次真实路由，并让 layer-1 权重继续走入官方吸收式 cache。所有结论都带证据身份与停止线。
     </p>
   </header>
 
@@ -67,8 +76,11 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     <button type="button" role="tab" data-artifact-tab="cache" aria-selected="false" tabindex="-1">
       <span>03</span><b>缓存实现账</b><small>latent vs HF eager</small>
     </button>
+    <button type="button" role="tab" data-artifact-tab="absorb" aria-selected="false" tabindex="-1">
+      <span>04</span><b>吸收式执行</b><small>real cache · SM120</small>
+    </button>
     <button type="button" role="tab" data-artifact-tab="evidence" aria-selected="false" tabindex="-1">
-      <span>04</span><b>证据断面</b><small>revision · shards · rerun</small>
+      <span>05</span><b>证据断面</b><small>revision · shards · rerun</small>
     </button>
   </div>
 
@@ -243,7 +255,114 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
 
     <div class="artifact-boundary">
       <b>IMPLEMENTATION CONTRACT</b>
-      <p>算法上“可以缓存 576 元素”与本次框架“实际缓存 5,120 元素”同时为真；生产吞吐还需要真实 latent-cache kernel 与 serving runtime。</p>
+      <p>算法上“可以缓存 576 元素”与 HF eager“实际缓存 5,120 元素”同时为真；下一页继续执行官方 absorb 参考路径，生产吞吐仍需要受支持的优化 kernel 与 serving runtime。</p>
+    </div>
+  </section>
+
+  <section class="artifact-panel" data-artifact-panel="absorb" hidden>
+    <div class="panel-lead">
+      <div><span>X + O / WEIGHT ABSORPTION</span><h4>这一次 576 元素真的进入了缓存</h4></div>
+      <p>
+        同一组真实 V2-Lite layer-1 权重、同一条 26-token hidden-state 轨迹：
+        先用 V3 官方 naive 路径展开 K/V，再用官方 absorb 路径做 25-token prefill + 1-token decode。
+      </p>
+    </div>
+
+    <div class="absorb-algebra" aria-label="MLA 权重吸收的两条代数变换">
+      <article>
+        <span>SCORE / 把 Wᴷ 搬到 query</span>
+        <code>qᵀ(Wᴷc) = (Wᴷᵀq)ᵀc</code>
+        <p>不再为每个历史 token、每个 head 存展开后的 no-RoPE key；query 先变成 512 维，再与 latent <i>c</i> 相乘。</p>
+      </article>
+      <article>
+        <span>VALUE / 把 Wⱽ 搬到 attention 后</span>
+        <code>Σ pₜ(Wⱽcₜ) = Wⱽ(Σ pₜcₜ)</code>
+        <p>先对 512 维 latent 做加权和，再展开成各 head 的 value。含位置的 64 维 RoPE key 不能这样吸收，必须单独缓存。</p>
+      </article>
+    </div>
+
+    <div class="absorb-cache-flow">
+      <article class="naive">
+        <header><span>X / V3 NAIVE · BF16</span><b>{bytes(absorb.cache_accounting.naive_active_bytes)}</b></header>
+        <div>
+          <code>K [1, 26, 16, 192]</code>
+          <code>V [1, 26, 16, 128]</code>
+        </div>
+        <p>26 × 5,120 元素；与 HF eager 是同一种展开状态。</p>
+      </article>
+      <i aria-hidden="true">→</i>
+      <article class="absorbed">
+        <header><span>X / V3 ABSORB · BF16</span><b>{bytes(absorb.cache_accounting.absorb_active_bytes)}</b></header>
+        <div>
+          <code>latent [1, 26, 512]</code>
+          <code>RoPE [1, 26, 64]</code>
+        </div>
+        <p>26 × 576 元素；两个官方 buffer 的真实 active slice。</p>
+      </article>
+    </div>
+
+    <div class="absorb-metrics">
+      <article><span>REAL CACHE RATIO</span><b>{absorb.cache_accounting.naive_over_absorb_ratio.toFixed(4)}×</b><p>展开 K/V ÷ latent + RoPE</p></article>
+      <article><span>BF16 MAX |Δ|</span><b>{absorb.correctness.v3_naive_vs_absorb_bfloat16.max_abs}</b><p>官方 V3 naive ↔ absorb decode</p></article>
+      <article><span>FP32 MAX |Δ|</span><b>{scientific(absorb.correctness.v3_naive_vs_absorb_float32.max_abs)}</b><p>同一 BF16 权重转 FP32 做代数审计</p></article>
+      <article class="exact"><span>INDEPENDENT RERUN</span><b>{absorbExact ? "BYTE-EXACT" : "MISMATCH"}</b><p>完整 JSON SHA-256 6b4c714a…e63d</p></article>
+    </div>
+
+    <div class="precision-lens">
+      <div>
+        <span>HF EAGER ↔ V3 NAIVE / BF16</span>
+        <i><em style={`--error:${absorb.correctness.hf_eager_vs_v3_naive_bfloat16.max_abs / .004}`}></em></i>
+        <b>max {absorb.correctness.hf_eager_vs_v3_naive_bfloat16.max_abs}</b>
+      </div>
+      <div>
+        <span>V3 NAIVE ↔ ABSORB / BF16</span>
+        <i><em style={`--error:${absorb.correctness.v3_naive_vs_absorb_bfloat16.max_abs / .004}`}></em></i>
+        <b>max {absorb.correctness.v3_naive_vs_absorb_bfloat16.max_abs}</b>
+      </div>
+      <div>
+        <span>V3 NAIVE ↔ ABSORB / FP32</span>
+        <i><em style={`--error:${absorb.correctness.v3_naive_vs_absorb_float32.max_abs / .004}`}></em></i>
+        <b>max {scientific(absorb.correctness.v3_naive_vs_absorb_float32.max_abs)}</b>
+      </div>
+      <p>
+        三组输出均 finite。FP32 不是新的 checkpoint 精度，而是把同一组 BF16 权重提升后，
+        隔离矩阵乘法顺序造成的舍入误差。
+      </p>
+    </div>
+
+    <div class="kernel-contract">
+      <div>
+        <span>OFFICIAL FLASHMLA / PINNED {absorb.flashmla_boundary.official_revision.slice(0, 8)}</span>
+        <h5>算法路径已执行，不等于优化 kernel 已执行</h5>
+        <p>官方支持矩阵与编译目标只覆盖 SM90 / SM100；本机 RTX 5090 是 SM120。</p>
+      </div>
+      <div class="kernel-matrix" role="table" aria-label="FlashMLA 官方架构支持矩阵">
+        <div class="head" role="row"><b>KERNEL</b><b>SM90</b><b>SM100</b><b>LOCAL SM120</b></div>
+        <div role="row"><span>dense decode</span><i class="yes">支持</i><i>—</i><i class="no">未支持</i></div>
+        <div role="row"><span>sparse decode</span><i class="yes">支持</i><i class="yes">支持</i><i class="no">未支持</i></div>
+        <div role="row"><span>dense prefill</span><i>—</i><i class="yes">支持</i><i class="no">未支持</i></div>
+        <div role="row"><span>sparse prefill</span><i class="yes">支持</i><i class="yes">支持</i><i class="no">未支持</i></div>
+      </div>
+    </div>
+
+    <div class="execution-split">
+      <article><span>EXECUTED</span><b>官方 V3 pure-PyTorch absorb</b><p>真实权重、真实 latent / RoPE buffers、真实 incremental decode。</p></article>
+      <article><span>BUILD ATTEMPTS</span><b>2 个环境，均未产出 wheel</b><p>host 编译器边界；隔离 CUDA 13 环境缺少 <code>cuda/std/utility</code>。</p></article>
+      <article class="blocked"><span>NOT EXECUTED</span><b>FlashMLA optimized kernel</b><p>源码只生成 <code>sm_90a</code> / <code>sm_100f</code>，dense decode 运行时还拒绝非 SM90a。</p></article>
+    </div>
+
+    <div class="evidence-links">
+      <a href="https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py" rel="noreferrer">官方 V3 absorb 参考实现 ↗</a>
+      <a href="https://github.com/deepseek-ai/FlashMLA" rel="noreferrer">官方 FlashMLA 支持矩阵 ↗</a>
+      <a href="https://arxiv.org/abs/2405.04434" rel="noreferrer">DeepSeek-V2 MLA 报告 ↗</a>
+    </div>
+
+    <div class="artifact-boundary">
+      <b>CORRECT CLAIM</b>
+      <p>
+        本实验复现的是“真实权重 + 官方参考实现中的压缩缓存与 decode 等价性”，不是 FlashMLA 性能、
+        生产 serving 吞吐或 V2 论文完整模型相对 MHA 的 93.3% 缓存降幅。
+      </p>
     </div>
   </section>
 
@@ -328,15 +447,16 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
 
     <div class="artifact-boundary">
       <b>U / STILL OPEN</b>
-      <p>完整 27 层生成、真实 latent-cache kernel、生产服务、训练负载、FP8/pipeline 与 R1-like 训练 trace 仍未覆盖。</p>
+      <p>完整 27 层生成、受支持硬件上的 FlashMLA 优化 kernel、生产服务、训练负载、FP8/pipeline 与 R1-like 训练 trace 仍未覆盖。</p>
     </div>
   </section>
 
   <figcaption>
     <span>可复现入口</span>
     <code>experiments/deepseek/v2_lite_trace.py</code> ·
+    <code>experiments/deepseek/v2_lite_absorb_probe.py</code> ·
     <code>src/data/deepseek-v2-lite-trace.json</code> ·
-    <code>research/DEEPSEEK_V2_LITE_TRACE.md</code>
+    <code>research/DEEPSEEK_MLA_ABSORB_AUDIT.md</code>
   </figcaption>
 
   <script is:inline type="application/json" data-dsv2-trace set:html={compactJson}></script>
@@ -575,6 +695,11 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .jaccard-block span,
   .observed-cache span,
   .artifact-identity span,
+  .absorb-algebra span,
+  .absorb-cache-flow span,
+  .absorb-metrics span,
+  .kernel-contract span,
+  .execution-split span,
   .repro-gate span,
   .dependency-split span,
   .checksum-grid span {
@@ -631,7 +756,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .artifact-status b { color: var(--ink); font-size: .72rem; }
   .artifact-tabs {
     display: grid;
-    grid-template-columns: repeat(4, 1fr);
+    grid-template-columns: repeat(5, 1fr);
     background: var(--ink);
   }
   .artifact-tabs button {
@@ -792,6 +917,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .route-metrics,
   .cache-ratio,
   .load-lessons,
+  .absorb-metrics,
   .artifact-identity,
   .checksum-grid {
     display: grid;
@@ -802,6 +928,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .route-metrics article,
   .cache-ratio article,
   .load-lessons article,
+  .absorb-metrics article,
   .artifact-identity article,
   .checksum-grid article {
     padding: .8rem;
@@ -810,20 +937,24 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .route-metrics article:last-child,
   .cache-ratio article:last-child,
   .load-lessons article:last-child,
+  .absorb-metrics article:last-child,
   .artifact-identity article:last-child,
   .checksum-grid article:last-child { border-right: 0; }
   .route-metrics span,
   .cache-ratio span,
-  .load-lessons span {
+  .load-lessons span,
+  .absorb-metrics span {
     display: block;
     color: var(--blue);
     font: 700 .61rem/1 var(--font-mono);
   }
   .route-metrics b,
-  .cache-ratio b { display: block; margin-top: .35rem; font: 750 1.25rem/1 var(--font-display); }
+  .cache-ratio b,
+  .absorb-metrics b { display: block; margin-top: .35rem; font: 750 1.25rem/1 var(--font-display); }
   .route-metrics p,
   .cache-ratio p,
-  .load-lessons p { margin: .3rem 0 0; color: rgba(32,32,39,.56); font-size: .65rem; line-height: 1.45; }
+  .load-lessons p,
+  .absorb-metrics p { margin: .3rem 0 0; color: rgba(32,32,39,.56); font-size: .65rem; line-height: 1.45; }
   .artifact-boundary {
     display: grid;
     grid-template-columns: 11rem 1fr;
@@ -950,6 +1081,127 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .cache-ledger p { margin: .35rem 0 0; color: rgba(32,32,39,.58); font-size: .68rem; line-height: 1.45; }
   .cache-ratio { grid-template-columns: repeat(3, 1fr); }
   .cache-ratio .warning { background: rgba(186,118,44,.09); }
+  .absorb-algebra {
+    display: grid;
+    grid-template-columns: 1fr 1fr;
+    gap: .8rem;
+  }
+  .absorb-algebra article {
+    padding: 1rem;
+    border: 1px solid rgba(32,32,39,.15);
+    background: #fffdf8;
+  }
+  .absorb-algebra code {
+    display: block;
+    margin: .75rem 0;
+    color: var(--ink);
+    font: 750 clamp(.88rem, 1.7vw, 1.2rem)/1.2 var(--font-mono);
+  }
+  .absorb-algebra p,
+  .absorb-cache-flow p,
+  .kernel-contract p,
+  .execution-split p {
+    margin: 0;
+    color: rgba(32,32,39,.62);
+    font-size: .69rem;
+    line-height: 1.55;
+  }
+  .absorb-cache-flow {
+    display: grid;
+    grid-template-columns: 1fr auto 1fr;
+    gap: .8rem;
+    align-items: center;
+    margin-top: .8rem;
+  }
+  .absorb-cache-flow > i { color: var(--teal); font: 800 1.5rem/1 var(--font-display); }
+  .absorb-cache-flow article {
+    padding: 1rem;
+    border: 1px solid rgba(32,32,39,.15);
+  }
+  .absorb-cache-flow article.naive { background: rgba(161,77,77,.07); }
+  .absorb-cache-flow article.absorbed { background: rgba(57,120,110,.09); }
+  .absorb-cache-flow header { display: flex; justify-content: space-between; gap: 1rem; align-items: center; }
+  .absorb-cache-flow header b { font: 760 1.25rem/1 var(--font-display); }
+  .absorb-cache-flow div { display: flex; flex-wrap: wrap; gap: .4rem; margin: .75rem 0; }
+  .absorb-cache-flow code {
+    padding: .35rem .45rem;
+    background: rgba(32,32,39,.08);
+    font-size: .66rem;
+  }
+  .absorb-metrics { grid-template-columns: repeat(4, 1fr); }
+  .absorb-metrics article.exact { background: rgba(57,120,110,.1); }
+  .precision-lens {
+    display: grid;
+    gap: .55rem;
+    margin-top: .8rem;
+    padding: 1rem;
+    background: var(--ink);
+    color: white;
+  }
+  .precision-lens > div {
+    display: grid;
+    grid-template-columns: minmax(13rem, 1.3fr) 1fr minmax(8rem, auto);
+    gap: .8rem;
+    align-items: center;
+  }
+  .precision-lens span,
+  .precision-lens b { font: 650 .62rem/1.3 var(--font-mono); }
+  .precision-lens b { text-align: right; }
+  .precision-lens i { height: 7px; overflow: hidden; background: rgba(255,255,255,.13); }
+  .precision-lens em {
+    display: block;
+    width: max(calc(var(--error) * 100%), 2px);
+    max-width: 100%;
+    height: 100%;
+    background: #b7bce5;
+  }
+  .precision-lens > div:last-of-type em { background: #91cfc1; }
+  .precision-lens > p {
+    margin: .25rem 0 0;
+    padding-top: .7rem;
+    border-top: 1px solid rgba(255,255,255,.15);
+    color: rgba(255,255,255,.58);
+    font-size: .66rem;
+    line-height: 1.5;
+  }
+  .kernel-contract {
+    display: grid;
+    grid-template-columns: .7fr 1.3fr;
+    gap: 1rem;
+    margin-top: .8rem;
+    padding: 1rem;
+    border: 1px solid rgba(32,32,39,.15);
+    background: #e6dfd3;
+  }
+  .kernel-contract h5 { margin: .45rem 0; font: 720 1rem/1.15 var(--font-display); }
+  .kernel-matrix { overflow-x: auto; border: 1px solid rgba(32,32,39,.14); background: #fffdf8; }
+  .kernel-matrix > div {
+    display: grid;
+    grid-template-columns: 1.4fr repeat(3, minmax(5rem, .7fr));
+    min-width: 520px;
+  }
+  .kernel-matrix b,
+  .kernel-matrix span,
+  .kernel-matrix i {
+    padding: .55rem;
+    border-right: 1px solid rgba(32,32,39,.1);
+    border-bottom: 1px solid rgba(32,32,39,.1);
+    font: 650 .61rem/1.2 var(--font-mono);
+  }
+  .kernel-matrix .head { background: var(--ink); color: white; }
+  .kernel-matrix i { color: rgba(32,32,39,.45); font-style: normal; text-align: center; }
+  .kernel-matrix i.yes { color: var(--teal); background: rgba(57,120,110,.08); }
+  .kernel-matrix i.no { color: var(--red); background: rgba(161,77,77,.08); }
+  .execution-split {
+    display: grid;
+    grid-template-columns: repeat(3, 1fr);
+    margin-top: .8rem;
+    border: 1px solid rgba(32,32,39,.14);
+  }
+  .execution-split article { padding: .9rem; border-right: 1px solid rgba(32,32,39,.12); }
+  .execution-split article:last-child { border-right: 0; }
+  .execution-split article.blocked { background: rgba(161,77,77,.08); }
+  .execution-split b { display: block; margin: .4rem 0; font-size: .77rem; }
   .artifact-identity { grid-template-columns: repeat(3, 1fr); }
   .artifact-identity article { display: grid; gap: .35rem; }
   .artifact-identity b { font-size: .78rem; }
@@ -1030,7 +1282,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .panel-lead,
     .heat-head,
     .aggregate-card,
-    .jaccard-block { grid-template-columns: 1fr; }
+    .jaccard-block,
+    .kernel-contract { grid-template-columns: 1fr; }
     .artifact-status { grid-template-columns: 1fr 1fr; }
     .artifact-tabs { grid-template-columns: 1fr 1fr; }
     .route-controls { grid-template-columns: 1fr 1fr; }
@@ -1039,7 +1292,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .route-stage > i { transform: rotate(90deg); justify-self: center; }
     .cache-controls { grid-template-columns: 1fr 1fr; }
     .route-metrics,
-    .checksum-grid { grid-template-columns: 1fr 1fr; }
+    .checksum-grid,
+    .absorb-metrics { grid-template-columns: 1fr 1fr; }
     .layer-evidence { grid-template-columns: repeat(9, 1fr); }
     .repro-gate { grid-template-columns: 1fr 1fr; }
     .repro-gate > p { grid-column: 1 / -1; padding: .8rem 0 0; border-left: 0; border-top: 1px solid rgba(255,255,255,.18); }
@@ -1062,20 +1316,28 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .route-metrics,
     .cache-ratio,
     .load-lessons,
+    .absorb-metrics,
     .artifact-identity,
     .checksum-grid,
     .cache-ledger { grid-template-columns: 1fr; }
     .route-metrics article,
     .cache-ratio article,
     .load-lessons article,
+    .absorb-metrics article,
     .artifact-identity article,
     .checksum-grid article { border-right: 0; border-bottom: 1px solid rgba(32,32,39,.12); }
     .artifact-boundary { grid-template-columns: 1fr; }
     .load-dials { grid-template-columns: 1fr; }
     .observed-cache,
-    .dependency-split { grid-template-columns: 1fr; }
+    .dependency-split,
+    .absorb-algebra,
+    .absorb-cache-flow,
+    .execution-split { grid-template-columns: 1fr; }
     .observed-cache > i,
     .dependency-split > i { transform: rotate(90deg); justify-self: center; }
+    .absorb-cache-flow > i { transform: rotate(90deg); justify-self: center; }
+    .precision-lens > div { grid-template-columns: 1fr; }
+    .precision-lens b { text-align: left; }
     .layer-evidence { grid-template-columns: repeat(7, 1fr); }
     .repro-gate { grid-template-columns: 1fr; }
     .repro-gate > p { grid-column: auto; }