diff --git a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml
index bd0dcd6623..46a9b50485 100644
--- a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml
@@ -3,7 +3,7 @@ base:
   name: qwen3.5-fp8-mi325x-sglang-mtp-8k1k
   model:
     path: hf:Qwen/Qwen3.5-397B-A17B-FP8
-    container: lmsysorg/sglang:v0.5.12-rocm720-mi30x
+    container: lmsysorg/sglang:v0.5.20-rocm720-mi30x
     precision: fp8
   resources:
     gpu_type: mi325x
@@ -28,7 +28,7 @@ base:
         trust-remote-code: true
         tokenizer-worker-num: 6
         enable-aiter-allreduce-fusion: true
-        cuda-graph-max-bs: 4
+        cuda-graph-max-bs-decode: 4
         disable-radix-cache: true
         max-prefill-tokens: 32768
         scheduler-recv-interval: 30
@@ -54,7 +54,7 @@ zip_override_concurrency:
   roles:
     agg:
       args:
-        cuda-graph-max-bs: [4, 8, 16, 32, 64]
+        cuda-graph-max-bs-decode: [4, 8, 16, 32, 64]
   benchmark:
     env:
       CONC: ['4', '8', '16', '32', '64']
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 543c404580..d7f8727f4f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -871,7 +871,7 @@ dsr1-fp8-mi325x-sglang-mtp:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsr1/sglang/mi325x-fp8-mtp/8k1k.yaml
 
 qwen3.5-fp8-mi325x-sglang-mtp:
-  image: lmsysorg/sglang:v0.5.12-rocm720-mi30x
+  image: lmsysorg/sglang:v0.5.20-rocm720-mi30x
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: mi325x