From 320bdd1ee2d6aface704c53d4674fd91a671fa74 Mon Sep 17 00:00:00 2001
From: Martin Hickey <martin.hickey@ie.ibm.com>
Date: Fri, 4 Sep 2026 22:26:28 +0100
Subject: [PATCH] [Docs] Document --retraction-policy,
 --return-hidden-states-mode, --language-model-only (#37989)

Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
Co-authored-by: mottopanikeiku <fcetin@hawk.iit.edu>
Co-authored-by: alp <falpercetin@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
---
 .../advanced_features/server_arguments.mdx     | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/docs/docs/advanced_features/server_arguments.mdx b/docs/docs/advanced_features/server_arguments.mdx
index 897e1dc83..fccc06f5a 100644
--- a/docs/docs/advanced_features/server_arguments.mdx
+++ b/docs/docs/advanced_features/server_arguments.mdx
@@ -519,6 +519,12 @@ Please consult the documentation below and [server_args.py](https://github.com/s
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>How conservative the schedule policy is. A larger value means more conservative scheduling. Use a larger value if you see requests being retracted frequently.</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`1.0`</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Type: float</td>
+    </tr>
+        <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--retraction-policy`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>The decode retraction policy to use when the KV cache is full. 'length' preserves the existing behavior and retracts short-output, long-input requests first. 'priority' retracts lower-priority requests first, using the same priority direction as priority scheduling.</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`length`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>length</code>, <code>priority</code></td>
     </tr>
         <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--page-size`</td>
@@ -2763,6 +2769,12 @@ Please consult the documentation below and [server_args.py](https://github.com/s
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`False`</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>bool flag (set to enable)</td>
     </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--return-hidden-states-mode`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Set the maximum hidden-state return mode supported by the server. `last` allows requests with return_hidden_states=False or `last`; `full` also allows return_hidden_states=True.</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`None`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>last</code>, <code>full</code></td>
+    </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--enable-return-routed-experts`</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Enable returning routed experts of each layer with responses.</td>
@@ -3106,6 +3118,12 @@ Please consult the documentation below and [server_args.py](https://github.com/s
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>For VLM, load weights for the language model only.</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`False`</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>bool flag (set to enable)</td>
+    </tr>
+        <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--language-model-only`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Skip the multimodal encoder entirely: its weights are never loaded and the tower is never built, freeing that GPU memory for KV cache. Multimodal requests are rejected. Unlike --language-only this is a standalone mode, not part of encoder/decoder disaggregation.</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>`False`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>bool flag (set to enable)</td>
     </tr>
         <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--encoder-transfer-backend`</td>