Spaces:

yitongl
/

video-opt-ablation

Running

App Files Files Community

yitongl commited on Jul 14

Commit

c19ff15

verified ·

1 Parent(s): 90c0a0d

per-prompt separate videos (no reel)

Browse files

Files changed (15) hide show

.gitattributes +13 -0
index.html +32 -31
videos/{lingbot.mp4 → lingbot/p00.mp4} +2 -2
videos/{wan14b.mp4 → lingbot/p01.mp4} +2 -2
videos/{wan5b.mp4 → lingbot/p02.mp4} +2 -2
videos/wan14b/p00.mp4 +3 -0
videos/wan14b/p01.mp4 +3 -0
videos/wan14b/p02.mp4 +3 -0
videos/wan14b/p03.mp4 +3 -0
videos/wan14b/p04.mp4 +3 -0
videos/wan5b/p00.mp4 +3 -0
videos/wan5b/p01.mp4 +3 -0
videos/wan5b/p02.mp4 +3 -0
videos/wan5b/p03.mp4 +3 -0
videos/wan5b/p04.mp4 +3 -0

.gitattributes CHANGED Viewed

@@ -36,3 +36,16 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 videos/lingbot.mp4 filter=lfs diff=lfs merge=lfs -text
 videos/wan14b.mp4 filter=lfs diff=lfs merge=lfs -text
 videos/wan5b.mp4 filter=lfs diff=lfs merge=lfs -text

 videos/lingbot.mp4 filter=lfs diff=lfs merge=lfs -text
 videos/wan14b.mp4 filter=lfs diff=lfs merge=lfs -text
 videos/wan5b.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/lingbot/p00.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/lingbot/p01.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/lingbot/p02.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan14b/p00.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan14b/p01.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan14b/p02.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan14b/p03.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan14b/p04.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan5b/p00.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan5b/p01.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan5b/p02.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan5b/p03.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/wan5b/p04.mp4 filter=lfs diff=lfs merge=lfs -text

index.html CHANGED Viewed

@@ -12,6 +12,9 @@
   h1{font-size:30px;margin:0 0 6px}
   .sub{color:var(--mut);margin:0 0 28px}
   h2{font-size:22px;border-bottom:1px solid var(--bd);padding-bottom:10px;margin:44px 0 20px}
   .cards{display:grid;grid-template-columns:repeat(auto-fit,minmax(320px,1fr));gap:18px}
   .card{background:var(--card);border:1px solid var(--bd);border-radius:12px;padding:20px}
   .card h3{margin:0 0 4px;font-size:18px}
@@ -23,16 +26,17 @@
   th{color:var(--mut);font-weight:600;font-size:12px;text-transform:uppercase;letter-spacing:.4px}
   td.x{color:var(--acc);font-weight:700;text-align:right;font-variant-numeric:tabular-nums}
   .drop{color:var(--red)!important}
-  .vid{background:var(--card);border:1px solid var(--bd);border-radius:12px;padding:16px;margin-bottom:22px}
-  .vid h3{margin:0 0 3px;font-size:18px}
-  .vid .meta{margin-bottom:12px}
-  video{width:100%;border-radius:8px;background:#000}
   .lr{display:flex;gap:10px;font-size:12px;color:var(--mut);margin-top:8px}
   .pill{padding:2px 8px;border-radius:20px}
   .pl{background:rgba(248,81,73,.18);color:#ff9d97}
   .pr{background:rgba(63,185,80,.18);color:#7ee787}
-  .note{color:var(--mut);font-size:12.5px;margin-top:10px;padding:10px 12px;background:rgba(88,166,255,.08);border-left:3px solid var(--blue);border-radius:4px}
   footer{color:var(--mut);font-size:12.5px;margin-top:50px;border-top:1px solid var(--bd);padding-top:16px}
 </style>
 </head>
 <body>
@@ -42,7 +46,6 @@
   <h2>① 加速方法与各方法贡献 (Ablation)</h2>
   <div class="cards">
     <div class="card">
       <h3>Wan2.2 TI2V-5B</h3>
       <p class="meta">单卡 · 704×1280 · 121 帧 · 50 步</p>
@@ -54,7 +57,6 @@
         <tr><td>Attention (PISA)</td><td class="x drop">弃(0.85×)</td></tr>
       </table>
     </div>
     <div class="card">
       <h3>Wan2.2-A14B (14B)</h3>
       <p class="meta">4-GPU CP4 · 720×1280 · 81 帧 · 40 步</p>
@@ -66,7 +68,6 @@
         <tr><td>Attention <span style="color:#8b949e">PISA density 0.10</span></td><td class="x">1.28×</td></tr>
       </table>
     </div>
     <div class="card">
       <h3>LingBot-Video (MoE 30B-A3B)</h3>
       <p class="meta">4-GPU CP4+FSDP · base 480p → refiner 1080p</p>
@@ -78,38 +79,38 @@
         <tr><td>Cache <span style="color:#8b949e">EasyCache base(1.18×) + refiner(1.10×)</span></td><td class="x">1.30×</td></tr>
       </table>
     </div>
   </div>
-  <div class="note">系数为增量口径（每个方法在前一个之上），相乘 ≈ 该模型总加速。各模型主导方法不同：5B=Cache 主导、14B=Cache+Attention、LingBot=Kernel(cuDNN) 主导。PISA sparse-attention 只在大注意力上有效（14B / LingBot-refiner ✓；5B ✗）。</div>
   <h2>② 端到端生成对比：Baseline vs Full-Opt</h2>
-  <div class="vid">
-    <h3>Wan2.2 TI2V-5B — 2.885×</h3>
-    <p class="meta">5 个官方 prompt 拼接 · 704×1280 · 50 步 · seed 42</p>
-    <video src="videos/wan5b.mp4" controls muted loop preload="metadata"></video>
-    <div class="lr"><span class="pill pl">左 = Baseline 70.25s</span><span class="pill pr">右 = Full-Opt 24.35s (kernel + EasyCache)</span></div>
-  </div>
-  <div class="vid">
-    <h3>Wan2.2-A14B (14B)</h3>
-    <p class="meta">5 个官方 prompt 拼接 · 720×1280 · 81 帧 · 40 步 · seed 1024</p>
-    <video src="videos/wan14b.mp4" controls muted loop preload="metadata"></video>
-    <div class="lr"><span class="pill pl">左 = Baseline (单卡 naive)</span><span class="pill pr">右 = Full-Opt 58.89s (CP4 + kernel + EasyCache + PISA)</span></div>
-    <div class="note">视频对比是相对<b>单卡 naive baseline</b>（含 4-GPU 并行化，视觉比值更大）；上方 Ablation 的 <b>2.19×</b> 是相对<b>同拓扑 CP4 baseline</b> 的纯优化贡献（kernel × cache × attention）。</div>
-  </div>
-  <div class="vid">
-    <h3>LingBot-Video (MoE 30B-A3B) — 2.6×</h3>
-    <p class="meta">3 个官方 t2v prompt 拼接 · refiner 1088×1920 · seed 42</p>
-    <video src="videos/lingbot.mp4" controls muted loop preload="metadata"></video>
-    <div class="lr"><span class="pill pl">左 = Baseline 375.5s</span><span class="pill pr">右 = Full-Opt 144.4s (cuDNN + PISA + EasyCache)</span></div>
-  </div>
   <footer>
     优化改变浮点 reduction 顺序与 attention 稀疏（PISA/EasyCache），输出非逐像素等价；同 seed 固定初始噪声。计时排除模型加载。
-    全部素材与元数据：<a style="color:#58a6ff" href="https://huggingface.co/datasets/yitongl/video-opt-baseline-vs-fullopt">HF dataset yitongl/video-opt-baseline-vs-fullopt</a>。
-    Quant (nvfp4) 一节暂略。
   </footer>
 </div>
 </body>

   h1{font-size:30px;margin:0 0 6px}
   .sub{color:var(--mut);margin:0 0 28px}
   h2{font-size:22px;border-bottom:1px solid var(--bd);padding-bottom:10px;margin:44px 0 20px}
+  h3.model{font-size:19px;margin:34px 0 4px}
+  .mmeta{color:var(--mut);font-size:13px;margin:0 0 16px}
+  .badge{display:inline-block;background:rgba(63,185,80,.16);color:#7ee787;font-weight:700;padding:2px 10px;border-radius:20px;font-size:14px;margin-left:8px}
   .cards{display:grid;grid-template-columns:repeat(auto-fit,minmax(320px,1fr));gap:18px}
   .card{background:var(--card);border:1px solid var(--bd);border-radius:12px;padding:20px}
   .card h3{margin:0 0 4px;font-size:18px}
   th{color:var(--mut);font-weight:600;font-size:12px;text-transform:uppercase;letter-spacing:.4px}
   td.x{color:var(--acc);font-weight:700;text-align:right;font-variant-numeric:tabular-nums}
   .drop{color:var(--red)!important}
+  .row{background:var(--card);border:1px solid var(--bd);border-radius:12px;padding:14px 14px 10px;margin-bottom:16px}
+  .plabel{font-size:14px;margin:0 0 10px;color:var(--fg)}
+  .pn{color:var(--blue);font-weight:700;margin-right:6px}
+  video{width:100%;border-radius:8px;background:#000;display:block}
   .lr{display:flex;gap:10px;font-size:12px;color:var(--mut);margin-top:8px}
   .pill{padding:2px 8px;border-radius:20px}
   .pl{background:rgba(248,81,73,.18);color:#ff9d97}
   .pr{background:rgba(63,185,80,.18);color:#7ee787}
+  .note{color:var(--mut);font-size:12.5px;margin-top:8px;padding:10px 12px;background:rgba(88,166,255,.08);border-left:3px solid var(--blue);border-radius:4px}
   footer{color:var(--mut);font-size:12.5px;margin-top:50px;border-top:1px solid var(--bd);padding-top:16px}
+  a{color:#58a6ff}
 </style>
 </head>
 <body>
   <h2>① 加速方法与各方法贡献 (Ablation)</h2>
   <div class="cards">
     <div class="card">
       <h3>Wan2.2 TI2V-5B</h3>
       <p class="meta">单卡 · 704×1280 · 121 帧 · 50 步</p>
         <tr><td>Attention (PISA)</td><td class="x drop">弃(0.85×)</td></tr>
       </table>
     </div>
     <div class="card">
       <h3>Wan2.2-A14B (14B)</h3>
       <p class="meta">4-GPU CP4 · 720×1280 · 81 帧 · 40 步</p>
         <tr><td>Attention <span style="color:#8b949e">PISA density 0.10</span></td><td class="x">1.28×</td></tr>
       </table>
     </div>
     <div class="card">
       <h3>LingBot-Video (MoE 30B-A3B)</h3>
       <p class="meta">4-GPU CP4+FSDP · base 480p → refiner 1080p</p>
         <tr><td>Cache <span style="color:#8b949e">EasyCache base(1.18×) + refiner(1.10×)</span></td><td class="x">1.30×</td></tr>
       </table>
     </div>
   </div>
+  <div class="note">系数为增量口径（每个方法在前一个之上），相乘 ≈ 该模型总加速。各模型主导方法不同：5B=Cache 主导、14B=Cache+Attention、LingBot=Kernel(cuDNN) 主导。PISA 只在大注意力上有效（14B / LingBot-refiner ✓；5B ✗）。</div>
   <h2>② 端到端生成对比：Baseline vs Full-Opt</h2>
+  <p class="sub" style="margin-top:-8px">每个 prompt 一个独立视频，左 = Baseline，右 = Full-Opt。</p>
+  <h3 class="model">Wan2.2 TI2V-5B <span class="badge">2.885×</span></h3>
+  <p class="mmeta">单卡 · 704×1280 · 121 帧 · 50 步 · seed 42 · 左 Baseline 70.25s / 右 Full-Opt 24.35s (kernel + EasyCache)</p>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 0</span>Will Smith casually eats noodles at a street food market</p><video src="videos/wan5b/p00.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 1</span>A lone hiker atop a towering cliff against the vast horizon</p><video src="videos/wan5b/p01.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 2</span>A hand tossing a bright yellow lemon from a wooden bowl</p><video src="videos/wan5b/p02.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 3</span>A curious raccoon peers through a field of yellow sunflowers</p><video src="videos/wan5b/p03.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 4</span>A superintelligent humanoid robot waking up in a lab</p><video src="videos/wan5b/p04.mp4" controls muted loop preload="metadata"></video></div>
+  <h3 class="model">Wan2.2-A14B (14B) <span class="badge">2.19× (纯优化)</span></h3>
+  <p class="mmeta">4-GPU CP4 · 720×1280 · 81 帧 · 40 步 · seed 1024 · 左 Baseline (单卡 naive) / 右 Full-Opt 58.89s</p>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 0</span>Will Smith casually eats noodles at a street food market</p><video src="videos/wan14b/p00.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 1</span>A lone hiker atop a towering cliff against the vast horizon</p><video src="videos/wan14b/p01.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 2</span>A hand tossing a bright yellow lemon from a wooden bowl</p><video src="videos/wan14b/p02.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 3</span>A curious raccoon peers through a field of yellow sunflowers</p><video src="videos/wan14b/p03.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 4</span>A superintelligent humanoid robot waking up in a lab</p><video src="videos/wan14b/p04.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="note">视频对比是相对<b>单卡 naive baseline</b>（含 4-GPU 并行化，视觉比值更大）；Ablation 的 <b>2.19×</b> 是相对<b>同拓扑 CP4 baseline</b> 的纯优化贡献。</div>
+  <h3 class="model">LingBot-Video (MoE 30B-A3B) <span class="badge">2.6×</span></h3>
+  <p class="mmeta">4-GPU CP4+FSDP · refiner 1088×1920 · 121 帧 · seed 42 · 左 Baseline 375.5s / 右 Full-Opt 144.4s (cuDNN + PISA + EasyCache)</p>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 0</span>A young woman modeling an outfit in a bright modern apartment</p><video src="videos/lingbot/p00.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 1</span>A child blowing shimmering bubbles outdoors on a sunny day</p><video src="videos/lingbot/p01.mp4" controls muted loop preload="metadata"></video></div>
+  <div class="row"><p class="plabel"><span class="pn">Prompt 2</span>First-person view of a desk workspace with a game controller</p><video src="videos/lingbot/p02.mp4" controls muted loop preload="metadata"></video></div>
   <footer>
     优化改变浮点 reduction 顺序与 attention 稀疏（PISA/EasyCache），输出非逐像素等价；同 seed 固定初始噪声。计时排除模型加载。
+    全部素材与元数据：<a href="https://huggingface.co/datasets/yitongl/video-opt-baseline-vs-fullopt">HF dataset yitongl/video-opt-baseline-vs-fullopt</a>。Quant (nvfp4) 一节暂略。
   </footer>
 </div>
 </body>

videos/{lingbot.mp4 → lingbot/p00.mp4} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8de7b5e8a9dcce92f138a1caced2cc0b0ce595814de3e5e1c85420c8a0685bda
-size 8537203

 version https://git-lfs.github.com/spec/v1
+oid sha256:e8a57ec83c4e8207bc36d471cda88764d5bb09c1a51624a6b131eddd8b86bf2d
+size 2579632

videos/{wan14b.mp4 → lingbot/p01.mp4} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c50f4c5595fa78e5c11a1955dc6709d1cb04993e4252aa03004dd4cdb90b9429
-size 15082136

 version https://git-lfs.github.com/spec/v1
+oid sha256:1cab11ac206bfb2981f073be1fe392ee5e0b47f634c161f5326ff86369e5f5d9
+size 3473789

videos/{wan5b.mp4 → lingbot/p02.mp4} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:69ec294da5f304687dbf707bb25881825193b2798bdb418da14644b35270a2f1
-size 20448809

 version https://git-lfs.github.com/spec/v1
+oid sha256:e87bc0be10fe40ecb48a128efa77c9cd8e0af86895bdbe5d6e6c0ae3c2d7ca50
+size 2485309

videos/wan14b/p00.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:08ec7b0aa5612eacc354836605f88e50a0e950a759156c4c851e9a3720e6303f
+size 3317480

videos/wan14b/p01.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:9c85bb403686352a8bf145d4fbef5b6118e8b767b52a6b0dfea42d846ea55745
+size 2195608

videos/wan14b/p02.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:fad8db6fd7e1de88efff82a1e757f4e61d1c98bb4d224c1a4036e9106dd59d1e
+size 2026268

videos/wan14b/p03.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:48f933cc6a71139619ad3dcd1a68e80e6f150707194ba4f121cd5c21a8a800cc
+size 3737828

videos/wan14b/p04.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:db5b50a0a715ac05e78ed9365bdea2085c3a248b447442c1a2f368c5e663faf2
+size 3808044

videos/wan5b/p00.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:1d9c20c970970616ea84307ba9b26d3ee5a4d40af1843265f4970cb519873920
+size 6166905

videos/wan5b/p01.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:89f820fc823b8bc2ee3a5ea84f8b39bc41ae1187dbe95d053d0ed52dd219c235
+size 4380720

videos/wan5b/p02.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:6b0c1e6ec5b1b4da2f4d0d62f9f4e73f66567d9c2253d8ff5a27d56894e45ad9
+size 2562324

videos/wan5b/p03.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:52c2eae25bbc7abfb0f85945e03a10cd4f1014d5a949f15893421eafb3fd1b9e
+size 4279020

videos/wan5b/p04.mp4 ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:80ec2796a29e9193369602ea154ca8bf013fc7c80408c2dda7a62f754f3623f5
+size 3062975