bupalinyu commited on
Commit
cf4c3d9
路
verified 路
1 Parent(s): a3778c6

docs: say where strata-vision comes from (not the Edge0 source tree)

Browse files
Files changed (1) hide show
  1. README.md +15 -5
README.md CHANGED
@@ -447,9 +447,19 @@ and then serves OpenAI- and Anthropic-compatible endpoints:</p>
447
  </details>
448
  <details open>
449
  <summary><b>Images</b> 路 the vision encoder in this folder</summary>
450
- <p><code>mmproj-BF16.gguf</code> is the vision encoder. It turns a picture into the embedding rows
451
- the language model reads; the language model itself is unchanged. Two entries in <code>model.json</code> wire it in,
452
- next to <code>args</code>:</p>
 
 
 
 
 
 
 
 
 
 
453
  <pre><code class="language-json">"exe": "/path/to/strata",
454
  "vision": {
455
  "exe": "/path/to/strata-vision",
@@ -460,14 +470,14 @@ next to <code>args</code>:</p>
460
  }
461
  </code></pre>
462
  <ul>
463
- <li><code>strata-vision</code> ships with the engine build, in the same folder as the engine.</li>
464
  <li><code>"model"</code> is the text backbone this folder already uses.</li>
465
  <li><code>"gpu": false</code> runs the encoder on the CPU and takes no VRAM. <code>true</code> runs it on the card and
466
  keeps about 1.4 GB free for it, so the expert cache gets that much less.</li>
467
  <li><code>"max_tokens"</code> caps how many tokens one picture becomes. More keeps more detail and takes longer.</li>
468
  </ul>
469
  <p>Add <code>--vision</code> to <code>"args"</code> as well. The server adds it by itself when the config has a
470
- <code>vision</code> section and prints a line saying so; writing it keeps the two in step.</p>
 
471
  <p>The log says <code>images on</code> once the encoder is up. Then send a picture: <code>/image /path/to.png</code>
472
  followed by the question in the terminal chat, or an <code>image_url</code> part over the OpenAI API. That part takes a
473
  <code>data:</code> URL, an <code>http(s)</code> URL or a local file path.</p>
 
447
  </details>
448
  <details open>
449
  <summary><b>Images</b> 路 the vision encoder in this folder</summary>
450
+ <p><code>mmproj-BF16.gguf</code> is the vision encoder: it turns a picture into the embedding rows the language model
451
+ reads, and the language model itself is unchanged. Reading pictures also needs a small helper program,
452
+ <code>strata-vision</code>. That program is <strong>not</strong> in this repository and the Edge0 source tree does not
453
+ build it, so it comes from one of these two places:</p>
454
+ <ul>
455
+ <li><strong>The prebuilt engine archive.</strong> <code>setup.py</code> and <code>strata launch</code> download one and
456
+ unpack <code>strata</code> and <code>strata-vision</code> side by side, so an install that ran with <code>--vision</code>
457
+ already has the helper next to the engine.</li>
458
+ <li><strong>Building it.</strong> Take a Strata source checkout (the <code>tools/vision</code> project lives there, not
459
+ in Edge0) and run <code>cmake --build build-vision --target strata-vision</code>, then put the result next to
460
+ <code>strata</code>.</li>
461
+ </ul>
462
+ <p>Then two entries in <code>model.json</code>, next to <code>args</code>:</p>
463
  <pre><code class="language-json">"exe": "/path/to/strata",
464
  "vision": {
465
  "exe": "/path/to/strata-vision",
 
470
  }
471
  </code></pre>
472
  <ul>
 
473
  <li><code>"model"</code> is the text backbone this folder already uses.</li>
474
  <li><code>"gpu": false</code> runs the encoder on the CPU and takes no VRAM. <code>true</code> runs it on the card and
475
  keeps about 1.4 GB free for it, so the expert cache gets that much less.</li>
476
  <li><code>"max_tokens"</code> caps how many tokens one picture becomes. More keeps more detail and takes longer.</li>
477
  </ul>
478
  <p>Add <code>--vision</code> to <code>"args"</code> as well. The server adds it by itself when the config has a
479
+ <code>vision</code> section and prints a line saying so; writing it keeps the two in step. Without the helper beside the
480
+ engine, a request that carries a picture fails with <code>this server was started without the vision encoder</code>.</p>
481
  <p>The log says <code>images on</code> once the encoder is up. Then send a picture: <code>/image /path/to.png</code>
482
  followed by the question in the terminal chat, or an <code>image_url</code> part over the OpenAI API. That part takes a
483
  <code>data:</code> URL, an <code>http(s)</code> URL or a local file path.</p>