diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index fb3966ce6..5947724cf 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -1,5 +1,5 @@
 dsr1-fp4-mi355x-sglang:
-  image: lmsysorg/sglang:v0.5.12-rocm700-mi35x
+  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519
   model: amd/DeepSeek-R1-0528-MXFP4-Preview
   model-prefix: dsr1
   runner: mi355x
@@ -16,6 +16,7 @@ dsr1-fp4-mi355x-sglang:
     - isl: 8192
       osl: 1024
       search-space:
+      - { tp: 4, conc-start: 4, conc-end: 64 }
       - { tp: 8, conc-start: 4, conc-end: 64 }
     # Agentic-coding sweep commented out for this image-bump PR — the
     # 10-conc agentic matrix amplifies sweep cost and the bump validation
@@ -27,7 +28,7 @@ dsr1-fp4-mi355x-sglang:
     #   - { tp: 8, offloading: none, conc-list: [1, 2, 4, 8, 12, 16, 32, 64, 128, 256] }
 
 dsr1-fp4-mi355x-sglang-mtp:
-  image: lmsysorg/sglang:v0.5.12-rocm700-mi35x
+  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519
   model: amd/DeepSeek-R1-0528-MXFP4
   model-prefix: dsr1
   runner: mi355x
@@ -39,10 +40,12 @@ dsr1-fp4-mi355x-sglang-mtp:
     - isl: 1024
       osl: 1024
       search-space:
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
 
 dsr1-fp4-mi355x-atom:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d201e9f3b..03564a5e3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3395,3 +3395,12 @@
   description:
     - "Add DeepSeek-V4-Pro FP4 MI355X ATOM MTP3 benchmark; image rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.3"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1627
+
+- config-keys:
+    - dsr1-fp4-mi355x-sglang
+    - dsr1-fp4-mi355x-sglang-mtp
+  description:
+    - "Image: lmsysorg/sglang:v0.5.12-rocm700-mi35x -> lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 (ROCm 7.0 -> 7.2; same image already validated multi-node in PR #1566)"
+    - "Add tp:4 to search-space (8k1k for non-mtp; 1k1k+8k1k for mtp) to surface the TP4 throughput-per-GPU frontier"
+    - "Local MI355X bench (matched cfg: range-ratio 0.8 + ROCM_QUICK_REDUCE_QUANTIZATION=INT4): 8k1k decode TPOT 9-25% faster at every conc vs the rocm700 CI; c64 now leads both axes (TPOT 16.5 vs 18.1 ms, tput 2954 vs 2841 tok/s/gpu). TP4 c64 8k1k = 4254 tok/s/gpu (> TP8 2954)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1645