From 5b52e4a44a62a32edbc12030ad0f3fee93778db0 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 24 Sep 2026 22:31:29 -0400
Subject: [PATCH] [Klaud Cold] Update qwen3.5-fp8-mi355x-sglang (+mtp) SGLang
 ROCm image to v0.5.20-rocm720-mi35x-20260924

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---
 .../qwen3.5/sglang/mi355x-fp8-mtp/8k1k.yaml                | 6 +++---
 .../srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml  | 6 +++---
 configs/amd-master.yaml                                    | 4 ++--
 perf-changelog.yaml                                        | 7 +++++++
 4 files changed, 15 insertions(+), 8 deletions(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8-mtp/8k1k.yaml
index 488994e3bf..b97ad8cea5 100644
--- a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8-mtp/8k1k.yaml
@@ -3,7 +3,7 @@ base:
   name: qwen3.5-fp8-mi355x-sglang-mtp-8k1k
   model:
     path: hf:Qwen/Qwen3.5-397B-A17B-FP8
-    container: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260828
+    container: lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260924
     precision: fp8
   resources:
     gpu_type: mi355x
@@ -29,7 +29,7 @@ base:
         tokenizer-worker-num: 6
         enable-aiter-allreduce-fusion: true
         max-running-requests: 4
-        cuda-graph-max-bs: 4
+        cuda-graph-max-bs-decode: 4
         disable-radix-cache: true
         chunked-prefill-size: 32768
         scheduler-recv-interval: 30
@@ -60,7 +60,7 @@ zip_override_concurrency:
   roles:
     agg:
       args:
-        cuda-graph-max-bs: [4, 8, 16, 32, 64, 128, 256]
+        cuda-graph-max-bs-decode: [4, 8, 16, 32, 64, 128, 256]
         max-running-requests: [4, 8, 16, 32, 64, 128, 256]
   benchmark:
     env:
diff --git a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml
index c753c2b465..ddbe1db8c4 100644
--- a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml
@@ -3,7 +3,7 @@ base:
   name: qwen3.5-fp8-mi355x-sglang-8k1k
   model:
     path: hf:Qwen/Qwen3.5-397B-A17B-FP8
-    container: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260828
+    container: lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260924
     precision: fp8
   resources:
     gpu_type: mi355x
@@ -29,7 +29,7 @@ base:
         tokenizer-worker-num: 6
         enable-aiter-allreduce-fusion: true
         max-running-requests: 4
-        cuda-graph-max-bs: 4
+        cuda-graph-max-bs-decode: 4
         disable-radix-cache: true
         chunked-prefill-size: 32768
         scheduler-recv-interval: 30
@@ -56,7 +56,7 @@ zip_override_concurrency:
   roles:
     agg:
       args:
-        cuda-graph-max-bs: [4, 8, 16, 32, 64, 128, 256]
+        cuda-graph-max-bs-decode: [4, 8, 16, 32, 64, 128, 256]
         max-running-requests: [4, 8, 16, 32, 64, 128, 256]
   benchmark:
     env:
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index dd6c375967..51695b4066 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -182,7 +182,7 @@ qwen3.5-fp8-mi325x-sglang:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8/8k1k.yaml
 
 qwen3.5-fp8-mi355x-sglang:
-  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260828
+  image: lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260924
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: mi355x
@@ -201,7 +201,7 @@ qwen3.5-fp8-mi355x-sglang:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/mi355x-fp8/8k1k.yaml
 
 qwen3.5-fp8-mi355x-sglang-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260828
+  image: lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260924
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: mi355x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 89f152cacc..da223b7cd7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8887,3 +8887,10 @@
     - "Add DeepSeek-V4-Pro-0813 golden AL for draft lengths 4, 5, 7 and 8 (3.36 / 3.61 / 3.73 / 3.47)."
     - "Agentic PD router: pin --decode-policy round_robin so decode no longer inherits the prefill --policy (consistent_hashing)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3256
+
+- config-keys:
+    - qwen3.5-fp8-mi355x-sglang
+    - qwen3.5-fp8-mi355x-sglang-mtp
+  description:
+    - "Update SGLang ROCm image from v0.5.18-rocm720-mi35x-20260828 to v0.5.20-rocm720-mi35x-20260924 (latest nightly)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3423