minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

docs(cookbook): enable FlashInfer GDN for Qwen3.5 B200 (#37360)

Browse Source
This commit is contained in:
Ankur Singh
2026-09-01 11:48:41 -07:00
committed by GitHub
parent a2b8681d1d
commit 3315356cc0
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/src/snippets/autoregressive/qwen35-deployment.jsx
+9
View File
@@ -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 // Append backend configurations
if (hardware === 'b200' || (hardware === 'b300' && quantization === 'fp4')) { if (hardware === 'b200' || (hardware === 'b300' && quantization === 'fp4')) {
cmd += ` \\\n --attention-backend trtllm_mha`; cmd += ` \\\n --attention-backend trtllm_mha`;