From 3315356cc043d621c83df5ff5f989adc9ce7b623 Mon Sep 17 00:00:00 2001
From: Ankur Singh <as.ankursingh3.1@gmail.com>
Date: Tue, 1 Sep 2026 11:48:41 -0700
Subject: [PATCH] docs(cookbook): enable FlashInfer GDN for Qwen3.5 B200
 (#37360)

---
 docs/src/snippets/autoregressive/qwen35-deployment.jsx | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/docs/src/snippets/autoregressive/qwen35-deployment.jsx b/docs/src/snippets/autoregressive/qwen35-deployment.jsx
index 0085283a9..28a6c6184 100644
--- a/docs/src/snippets/autoregressive/qwen35-deployment.jsx
+++ b/docs/src/snippets/autoregressive/qwen35-deployment.jsx
@@ -438,6 +438,15 @@ export const Qwen35Deployment = () => {
       }
     }
 
+    // B200 NVFP4 with MTP runs TP2/EP2 (set above). Keep Triton as the base
+    // linear-attention backend while routing GDN decode and prefill through
+    // FlashInfer.
+    if (model === '397b' && hardware === 'b200' && quantization === 'fp4' && speculative === 'enabled') {
+      cmd += ` \\\n  --linear-attn-backend triton`;
+      cmd += ` \\\n  --linear-attn-decode-backend flashinfer`;
+      cmd += ` \\\n  --linear-attn-prefill-backend flashinfer`;
+    }
+
     // Append backend configurations
     if (hardware === 'b200' || (hardware === 'b300' && quantization === 'fp4')) {
       cmd += ` \\\n  --attention-backend trtllm_mha`;