ui: document ds4 performance setup in global + per-model settings

Add a "tuning ds4 for performance" note to the global ds4 Settings section and a
condensed inline note to the per-model ds4 streaming section on the Models page:
NVMe/SSD placement (≈10× prefill), capping the expert cache by count, sizing the
VRAM reserve, DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512, avoiding DS4_CUDA_WEIGHT_CACHE,
and that decode of a model larger than VRAM is streaming-bound (smaller quant is
the real fix). Distilled from measured tuning on the 154GB DeepSeek-V4 MoE.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

codai/admin/templates/models.html

@@ -602,6 +602,11 @@ window.__DEFAULT_WHISPER_SERVER_PATH__ = {{ default_whisper_server_path|tojson }
           <input type="text" id="cfg-ds4-extra-env" class="form-input" placeholder="DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512">
           <span class="form-hint" style="font-size:11px"><code>KEY=VALUE</code> pairs. e.g. <code>DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512</code> fixes weight-arena OOM under heavy expert caching. Blank = inherit global.</span>
         </div>
+        <span class="form-hint" style="font-size:11px;line-height:1.5;margin-top:.4rem;display:block">
+          <b>⚡ Performance:</b> put the GGUF on <b>NVMe/SSD</b> (≈10× prefill for a streamed MoE) ·
+          cap the expert cache with <code>--ssd-streaming-cache-experts N</code> (a count, e.g. <code>400</code>) — higher helps decode but too high OOMs the weight arena ·
+          set <b>reserve</b> ≈ weights + ~2 GiB · add <code>DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512</code> · <b>avoid</b> <code>DS4_CUDA_WEIGHT_CACHE</code> (slower). Decode of a model bigger than VRAM is streaming-bound (~0.5 tok/s) — the real fix is a smaller quant. Context = <b>n_ctx</b> above.
+        </span>
       </div>
       <div style="display:grid;grid-template-columns:1fr 1fr;gap:.75rem;margin-top:.75rem">
         <div class="form-row" style="margin:0">

codai/admin/templates/settings.html

@@ -479,6 +479,18 @@
       <input type="text" id="s-ds4-extra-env" class="form-input" placeholder="DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512">
       <span class="form-hint">Whitespace-separated <code>KEY=VALUE</code> pairs added to ds4-server's environment — for the many CUDA tunables it exposes only via env. Notably <code>DS4_CUDA_WEIGHT_ARENA_CHUNK_MB</code> (default 1792): lower it (e.g. <code>512</code>) so the weight arena allocates in smaller chunks that fit a heap fragmented by the streaming expert cache, fixing <i>model arena alloc failed … out of memory</i>.</span>
     </div>
+    <div class="form-row" style="background:var(--bg-2,#1a1a1a);border:1px solid var(--border,#333);border-radius:8px;padding:.7rem .8rem">
+      <label class="form-label" style="margin-bottom:.3rem">⚡ Tuning ds4 for performance</label>
+      <span class="form-hint" style="font-size:11.5px;line-height:1.55">
+        These are global <b>defaults</b>; override them per-model on the <b>Models</b> page (a big-quant streamer and a small one that fits VRAM want different settings). Order of impact:
+        <br>• <b>Put the GGUF on a fast NVMe/SSD</b>, not a spinning disk — for a streamed MoE this is the single biggest win (≈10× prefill). A symlink from the old path works.
+        <br>• <b>Cap the expert cache</b> with <code>--ssd-streaming-cache-experts N</code> (a COUNT, e.g. <code>400</code> — not NGB). Bigger = better decode hit-rate, but too big fragments VRAM and the weight arena OOMs; find the ceiling for your card.
+        <br>• <b>Reserve</b> ≈ resident weights + ~2 GiB (field above). ds4 defaults to half the card, which starves the cache.
+        <br>• Add <code>DS4_CUDA_WEIGHT_ARENA_CHUNK_MB=512</code> (env) to stop the weight-arena OOM under heavy caching.
+        <br>• <b>Avoid</b> <code>DS4_CUDA_WEIGHT_CACHE</code> — it disables ds4's zero-copy path and was slower in testing.
+        <br>• Decode of a model far larger than VRAM is streaming-bound (~0.5 tok/s here); the only real fix is a <b>smaller quant</b> (IQ2/IQ3) that mostly fits VRAM+RAM. Context window is set per-model via <b>n_ctx</b>.
+      </span>
+    </div>
   </div>
 </div>
 {% endblock %}