From d2b1243be0c3f67e6bf78ace97b2c1b2382cb1c6 Mon Sep 17 00:00:00 2001
From: Jun Liu <jun.c.liu@rakuten.com>
Date: Wed, 15 Jul 2026 23:18:36 +0900
Subject: [PATCH] docs: document CUDA crash dump output (#31333)

Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
---
 docs_new/docs/advanced_features/observability.mdx    | 9 ++++++---
 docs_new/docs/advanced_features/server_arguments.mdx | 2 +-
 2 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/docs_new/docs/advanced_features/observability.mdx b/docs_new/docs/advanced_features/observability.mdx
index 4ddb193f2..4ff1b5bb9 100644
--- a/docs_new/docs/advanced_features/observability.mdx
+++ b/docs_new/docs/advanced_features/observability.mdx
@@ -38,7 +38,10 @@ To replay the request dump, use `scripts/playground/replay_request_dump.py`.
 
 ## Crash Dump and Replay
 Sometimes the server might crash, and you may want to debug the cause of the crash.
-SGLang supports crash dumping, which will dump all requests from the 5 minutes before the crash, allowing you to replay the requests and debug the reason later.
+SGLang can preserve recent request data for replay and collect CUDA device coredumps for low-level debugging.
 
-To enable crash dumping, use `--crash-dump-folder /tmp/crash_dump`.
-To replay the crash dump, use `scripts/playground/replay_request_dump.py`.
+Set the crash diagnostics folder with `--crash-dump-folder /tmp/crash_dump`.
+
+When SGLang handles a crash, it writes completed requests retained by the crash-dump buffer plus in-flight requests to `/tmp/crash_dump/<hostname>/crash_dump_<timestamp>.pkl`. The file also contains the server arguments and launch command. Replay it with `scripts/playground/replay_request_dump.py`.
+
+On NVIDIA CUDA, the option also sets default environment variables before CUDA initializes. These defaults enable device coredumps on CUDA exceptions and allow SGLang to trigger device coredumps for live scheduler processes when it handles a crash. CUDA device coredumps are written to `/tmp/crash_dump/<hostname>/core.cuda.<timestamp>.<pid>`. Explicitly configured CUDA coredump environment variables take precedence, including a custom `CUDA_COREDUMP_FILE` path. This option does not configure OS process core dumps.
diff --git a/docs_new/docs/advanced_features/server_arguments.mdx b/docs_new/docs/advanced_features/server_arguments.mdx
index 7694eac40..b85ee4cc3 100644
--- a/docs_new/docs/advanced_features/server_arguments.mdx
+++ b/docs_new/docs/advanced_features/server_arguments.mdx
@@ -840,7 +840,7 @@ Please consult the documentation below and [server_args.py](https://github.com/s
     </tr>
         <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--crash-dump-folder`</td>
-      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Folder path to dump requests from the last 5 min before a crash (if any). If not specified, crash dumping is disabled.</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Folder for crash diagnostics. Stores completed requests retained by the crash-dump buffer plus in-flight requests and, on NVIDIA CUDA, configures device coredumps under this folder. Existing CUDA coredump environment variables take precedence. If not specified, this option does not enable request crash dumping or set CUDA coredump defaults.</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)"}}>Type: str</td>
     </tr>