diff --git a/.github/workflows/nightly-test-amd-rocm720.yml b/.github/workflows/nightly-test-amd-rocm720.yml
index 6808f2f20..8af8e9838 100644
--- a/.github/workflows/nightly-test-amd-rocm720.yml
+++ b/.github/workflows/nightly-test-amd-rocm720.yml
@@ -144,7 +144,7 @@ jobs:
 
   nightly-test-1-gpu-unit-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-test-1-gpu-unit-rocm720,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -174,7 +174,7 @@ jobs:
 
   nightly-test-1-gpu-kernel-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-test-1-gpu-kernel-rocm720,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -239,7 +239,7 @@ jobs:
 
   nightly-accuracy-2-gpu-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu-rocm720,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -338,7 +338,7 @@ jobs:
 
   nightly-accuracy-2-gpu-vlm-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu-vlm-rocm720,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -369,7 +369,7 @@ jobs:
 
   nightly-perf-2-gpu-text-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-2-gpu-text-rocm720,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -401,7 +401,7 @@ jobs:
 
   nightly-perf-2-gpu-vlm-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-2-gpu-vlm-rocm720,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -433,7 +433,7 @@ jobs:
 
   nightly-4-gpu-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-4-gpu-rocm720,'))
-    runs-on: linux-mi325-4gpu-sglang
+    runs-on: linux-mi300-4gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -469,7 +469,7 @@ jobs:
 
   nightly-accuracy-8-gpu-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-8-gpu-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -548,7 +548,7 @@ jobs:
 
   nightly-8-gpu-grok1-int4-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-grok1-int4-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -645,7 +645,7 @@ jobs:
 
   nightly-8-gpu-grok2-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-grok2-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -742,7 +742,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v31-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v31-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -788,7 +788,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v32-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v32-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -832,7 +832,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v32-mtp-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v32-mtp-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -876,7 +876,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v3-kv-fp8-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v3-kv-fp8-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1437,7 +1437,7 @@ jobs:
 
   nightly-8-gpu-kimi-k26-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-kimi-k26-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1510,7 +1510,7 @@ jobs:
 
   nightly-8-gpu-qwen3-235b-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen3-235b-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1582,7 +1582,7 @@ jobs:
 
   nightly-8-gpu-qwen35-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen35-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1714,7 +1714,7 @@ jobs:
 
   nightly-8-gpu-glm51-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm51-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1936,7 +1936,7 @@ jobs:
 
   nightly-8-gpu-minimax-m27-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-minimax-m27-rocm720,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1986,7 +1986,7 @@ jobs:
 
   nightly-1-gpu-zimage-turbo-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-1-gpu-zimage-turbo-rocm720,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
diff --git a/.github/workflows/nightly-test-amd.yml b/.github/workflows/nightly-test-amd.yml
index 73b4dcd8b..da995a2c3 100644
--- a/.github/workflows/nightly-test-amd.yml
+++ b/.github/workflows/nightly-test-amd.yml
@@ -140,7 +140,7 @@ jobs:
 
   nightly-test-1-gpu-unit:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-test-1-gpu-unit,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -171,7 +171,7 @@ jobs:
 
   nightly-test-1-gpu-kernel:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-test-1-gpu-kernel,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -240,7 +240,7 @@ jobs:
 
   nightly-accuracy-2-gpu:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -340,7 +340,7 @@ jobs:
 
   nightly-accuracy-2-gpu-vlm:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu-vlm,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -372,7 +372,7 @@ jobs:
 
   nightly-perf-2-gpu-text:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-2-gpu-text,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -405,7 +405,7 @@ jobs:
 
   nightly-perf-2-gpu-vlm:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-2-gpu-vlm,'))
-    runs-on: linux-mi325-2gpu-sglang
+    runs-on: linux-mi300-2gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -438,7 +438,7 @@ jobs:
 
   nightly-4-gpu:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-4-gpu,'))
-    runs-on: linux-mi325-4gpu-sglang
+    runs-on: linux-mi300-4gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -474,7 +474,7 @@ jobs:
 
   nightly-accuracy-8-gpu:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-8-gpu,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -553,7 +553,7 @@ jobs:
 
   nightly-8-gpu-grok1-int4:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-grok1-int4,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -650,7 +650,7 @@ jobs:
 
   nightly-8-gpu-grok2:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-grok2,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -747,7 +747,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v31:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v31,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -793,7 +793,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v32:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v32,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -837,7 +837,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v32-mtp:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v32-mtp,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -881,7 +881,7 @@ jobs:
 
   nightly-8-gpu-deepseek-v3-kv-fp8:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-deepseek-v3-kv-fp8,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1291,7 +1291,7 @@ jobs:
 
   nightly-8-gpu-kimi-k26:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-kimi-k26,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1364,7 +1364,7 @@ jobs:
 
   nightly-8-gpu-qwen3-235b:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen3-235b,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1436,7 +1436,7 @@ jobs:
 
   nightly-8-gpu-qwen35:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen35,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1568,7 +1568,7 @@ jobs:
 
   nightly-8-gpu-glm51:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm51,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1790,7 +1790,7 @@ jobs:
 
   nightly-8-gpu-minimax-m27:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-minimax-m27,'))
-    runs-on: linux-mi325-8gpu-sglang
+    runs-on: linux-mi300-8gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1840,7 +1840,7 @@ jobs:
 
   nightly-1-gpu-zimage-turbo:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-1-gpu-zimage-turbo,'))
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
diff --git a/.github/workflows/pr-test-amd-extra.yml b/.github/workflows/pr-test-amd-extra.yml
index 35f5c99fe..1f0d440eb 100644
--- a/.github/workflows/pr-test-amd-extra.yml
+++ b/.github/workflows/pr-test-amd-extra.yml
@@ -32,7 +32,7 @@ on:
         description: 'AMD runner pool to dispatch GPU jobs to'
         required: false
         type: choice
-        default: mi325
+        default: mi300
         options:
           - mi300
           - mi325
@@ -61,6 +61,11 @@ on:
         required: false
         type: string
         default: ''
+      runner_arch:
+        description: 'AMD runner pool to dispatch GPU jobs to'
+        required: false
+        type: string
+        default: mi300
       rocm_version:
         description: 'ROCm container variant (empty = Dockerfile default; rocm720 = ROCm 7.2.0)'
         required: false
@@ -125,10 +130,10 @@ jobs:
   # setup across scarce AMD GPUs to shave only a couple minutes of test time,
   # so one GPU running the whole suite sequentially is the better trade.
   extra-a-test-1-gpu-small-amd:
-    name: ${{ format('extra-a-test-1-gpu-small-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('extra-a-test-1-gpu-small-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi300') }}
     needs: [call-gate]
     if: ${{ !cancelled() && needs.call-gate.result == 'success' }}
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -157,10 +162,10 @@ jobs:
   # pool as small (AMD GPUs carry enough VRAM that "large" here is a CUDA
   # memory-tier label, not a separate AMD runner pool).
   extra-a-test-1-gpu-large-amd:
-    name: ${{ format('extra-a-test-1-gpu-large-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('extra-a-test-1-gpu-large-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi300') }}
     needs: [call-gate]
     if: ${{ !cancelled() && needs.call-gate.result == 'success' }}
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -187,10 +192,10 @@ jobs:
   # Multi-GPU TP / PP / PD mock-model + kv_canary e2e tests. Mirrors CUDA's
   # extra-a 2-gpu-large stage; runs on the 2-GPU AMD pool.
   extra-a-test-2-gpu-large-amd:
-    name: ${{ format('extra-a-test-2-gpu-large-amd{0} (linux-{1}-2gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('extra-a-test-2-gpu-large-amd{0} (linux-{1}-2gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi300') }}
     needs: [call-gate]
     if: ${{ !cancelled() && needs.call-gate.result == 'success' }}
-    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
diff --git a/.github/workflows/pr-test-amd-rocm720.yml b/.github/workflows/pr-test-amd-rocm720.yml
index d437e13e8..eb92573cf 100644
--- a/.github/workflows/pr-test-amd-rocm720.yml
+++ b/.github/workflows/pr-test-amd-rocm720.yml
@@ -229,6 +229,7 @@ jobs:
     uses: ./.github/workflows/pr-test-amd-extra.yml
     with:
       ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
+      runner_arch: mi300
       rocm_version: rocm720
       aiter_ref: ${{ inputs.aiter_ref }}
       continue_on_error: true
@@ -249,7 +250,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -295,7 +296,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-2gpu-sglang]
+        runner: [linux-mi300-2gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -336,7 +337,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -374,7 +375,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -412,7 +413,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -450,8 +451,9 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 4
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
         part: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
     runs-on: ${{matrix.runner}}
     steps:
@@ -471,9 +473,9 @@ jobs:
       - name: Install dependencies
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
       - name: Run test
-        timeout-minutes: 60
+        timeout-minutes: 75
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 14 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
+          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 14 --timeout-per-file 2400 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-b-test-1-gpu-small-amd-nondeterministic-rocm720:
     needs: [check-changes]
@@ -490,7 +492,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -565,8 +567,9 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
         part: [0, 1, 2]
     runs-on: ${{matrix.runner}}
     steps:
@@ -586,9 +589,9 @@ jobs:
       - name: Install dependencies
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
       - name: Run test
-        timeout-minutes: 45
+        timeout-minutes: 60
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
+          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 2700 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-b-test-2-gpu-large-amd-rocm720:
     needs: [check-changes]
@@ -604,8 +607,9 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
-        runner: [linux-mi325-2gpu-sglang]
+        runner: [linux-mi300-2gpu-sglang]
         part: [0, 1]
     runs-on: ${{matrix.runner}}
     steps:
@@ -625,9 +629,9 @@ jobs:
       - name: Install dependencies
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
       - name: Run test
-        timeout-minutes: 45
+        timeout-minutes: 120
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
+          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 5400 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   multimodal-gen-test-1-gpu-amd-rocm720:
     needs: [check-changes]
@@ -643,8 +647,9 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
-        runner: [linux-mi325-1gpu-sglang]
+        runner: [linux-mi300-1gpu-sglang]
         part: [0, 1, 2, 3]
     runs-on: ${{matrix.runner}}
     steps:
@@ -676,16 +681,7 @@ jobs:
 
       - name: Setup kernel caches
         run: |
-          # Use the persistent /sgl-data directory (mounted from /home/runner/sgl-data)
-          # This directory persists across container restarts on the self-hosted runner
-          docker exec ci_sglang mkdir -p /sgl-data/aiter-kernels /sgl-data/miopen-cache /sgl-data/hf-cache/hub
-
-          # Clear pre-built AITER kernels from Docker image to avoid segfaults
-          # The image may have stale/incompatible kernels at /sgl-workspace/aiter/aiter/jit/
-          echo "Clearing pre-built AITER kernels from Docker image..."
-          docker exec ci_sglang rm -rf /sgl-workspace/aiter/aiter/jit/*.so 2>/dev/null || true
-          docker exec ci_sglang rm -rf /sgl-data/aiter-kernels/*.so 2>/dev/null || true
-          echo "AITER kernels cleared - will be rebuilt on first use"
+          docker exec ci_sglang mkdir -p /sgl-data/miopen-cache /sgl-data/hf-cache/hub
 
           # Create persistent cache marker if /sgl-data is a real mount (not ephemeral)
           # This tells the test cleanup code to NOT delete downloaded models
@@ -706,7 +702,7 @@ jobs:
           free -h
           echo ""
           echo "=== Disk Space ==="
-          df -h /home/runner/sgl-data 2>/dev/null || df -h
+          df -h /home/runner/sglang-data 2>/dev/null || df -h
           echo ""
           echo "=== HF Cache Directory Structure ==="
           docker exec ci_sglang ls -la /sgl-data/hf-cache/ 2>/dev/null || echo "HF cache dir not found"
@@ -728,7 +724,7 @@ jobs:
           docker exec ci_sglang rocm-smi --showmeminfo vram 2>/dev/null || echo "rocm-smi not available"
 
       - name: Run diffusion server tests (1-GPU)
-        timeout-minutes: 60
+        timeout-minutes: 90
         run: |
           # AMD CI: All 1-GPU tests except FLUX.2 (FLUX.1 covers same code path)
           # Tests: T2V, T2I, I2V, LoRA
@@ -745,7 +741,6 @@ jobs:
             -e SGLANG_SKIP_CONSISTENCY=1 \
             -e SGLANG_TEST_NUM_INFERENCE_STEPS=5 \
             -e SGLANG_DIFFUSION_ARTIFACT_DIR=/sglang-checkout/diffusion-failures \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -e HF_HUB_ENABLE_HF_TRANSFER=1 \
             -e HF_HUB_DISABLE_SYMLINKS_WARNING=1 \
@@ -783,8 +778,9 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
-        runner: [linux-mi325-2gpu-sglang]
+        runner: [linux-mi300-2gpu-sglang]
         part: [0, 1, 2]  # 3 partitions: 2 parametrized + 1 standalone (single_test_file/test_disagg_server.py)
     runs-on: ${{matrix.runner}}
     steps:
@@ -816,15 +812,7 @@ jobs:
 
       - name: Setup kernel caches
         run: |
-          # Use the persistent /sgl-data directory (mounted from /home/runner/sgl-data)
-          docker exec ci_sglang mkdir -p /sgl-data/aiter-kernels /sgl-data/miopen-cache /sgl-data/hf-cache/hub
-
-          # Clear pre-built AITER kernels from Docker image to avoid segfaults
-          # The image may have stale/incompatible kernels at /sgl-workspace/aiter/aiter/jit/
-          echo "Clearing pre-built AITER kernels from Docker image..."
-          docker exec ci_sglang rm -rf /sgl-workspace/aiter/aiter/jit/*.so 2>/dev/null || true
-          docker exec ci_sglang rm -rf /sgl-data/aiter-kernels/*.so 2>/dev/null || true
-          echo "AITER kernels cleared - will be rebuilt on first use"
+          docker exec ci_sglang mkdir -p /sgl-data/miopen-cache /sgl-data/hf-cache/hub
 
           # Create persistent cache marker if /sgl-data is a real mount (not ephemeral)
           # This tells the test cleanup code to NOT delete downloaded models
@@ -845,7 +833,7 @@ jobs:
           free -h
           echo ""
           echo "=== Disk Space ==="
-          df -h /home/runner/sgl-data 2>/dev/null || df -h
+          df -h /home/runner/sglang-data 2>/dev/null || df -h
           echo ""
           echo "=== HF Cache Directory Structure ==="
           docker exec ci_sglang ls -la /sgl-data/hf-cache/ 2>/dev/null || echo "HF cache dir not found"
@@ -867,7 +855,7 @@ jobs:
           docker exec ci_sglang rocm-smi --showmeminfo vram 2>/dev/null || echo "rocm-smi not available"
 
       - name: Run diffusion server tests (2-GPU)
-        timeout-minutes: 150
+        timeout-minutes: 180
         run: |
           # AMD CI: All 2-GPU tests including LoRA
           # Tests: T2V, T2I, I2V, LoRA
@@ -884,7 +872,6 @@ jobs:
             -e SGLANG_SKIP_CONSISTENCY=1 \
             -e SGLANG_TEST_NUM_INFERENCE_STEPS=5 \
             -e SGLANG_DIFFUSION_ARTIFACT_DIR=/sglang-checkout/diffusion-failures \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -e HF_HUB_ENABLE_HF_TRANSFER=1 \
             -e HF_HUB_DISABLE_SYMLINKS_WARNING=1 \
@@ -923,7 +910,7 @@ jobs:
           ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
         )
       )
-    runs-on: linux-mi325-1gpu-sglang
+    runs-on: linux-mi300-1gpu-sglang
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -957,7 +944,6 @@ jobs:
           # Skip ltx2_vae_channels_last: it asserts CUDA `channels_last_3d`
           # memory-format behavior that the ROCm conv path doesn't reproduce.
           docker exec \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -w /sglang-checkout/python \
             ci_sglang python3 sglang/multimodal_gen/test/run_suite.py \
@@ -979,7 +965,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux-mi325-4gpu-sglang]
+        runner: [linux-mi300-4gpu-sglang]
         part: [0]
     runs-on: ${{matrix.runner}}
     steps:
@@ -1000,7 +986,7 @@ jobs:
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 60
+        timeout-minutes: 90
         run: |
           bash scripts/ci/amd/amd_ci_exec.sh \
             -e NCCL_CUMEM_ENABLE=0 \
@@ -1013,7 +999,7 @@ jobs:
               --suite stage-c-test-4-gpu-amd \
               --auto-partition-id ${{ matrix.part }} \
               --auto-partition-size 1 \
-              --timeout-per-file 1800 \
+              --timeout-per-file 5400 \
               --enable-retry \
               --max-attempts 2 \
               --retry-wait-seconds 120 \
@@ -1033,11 +1019,12 @@ jobs:
         )
       )
     env:
-      RUNNER_LABELS: linux-mi325-8gpu-sglang
+      RUNNER_LABELS: linux-mi300-8gpu-sglang
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
-        runner: [linux-mi325-8gpu-sglang]
+        runner: [linux-mi300-8gpu-sglang]
         part: [0, 1, 2, 3]
     runs-on: ${{matrix.runner}}
     steps:
diff --git a/.github/workflows/pr-test-amd.yml b/.github/workflows/pr-test-amd.yml
index 4e5320805..7dba7448d 100644
--- a/.github/workflows/pr-test-amd.yml
+++ b/.github/workflows/pr-test-amd.yml
@@ -64,7 +64,7 @@ on:
         description: 'AMD runner pool to dispatch GPU jobs to'
         required: false
         type: choice
-        default: mi325
+        default: mi300
         options:
           - mi300
           - mi325
@@ -80,6 +80,11 @@ on:
         required: false
         type: string
         default: ''
+      runner_arch:
+        description: 'AMD runner pool to dispatch GPU jobs to'
+        required: false
+        type: string
+        default: mi300
       run_all_tests:
         description: "Run all tests (for releasing or testing purpose)"
         required: false
@@ -217,13 +222,14 @@ jobs:
     uses: ./.github/workflows/pr-test-amd-extra.yml
     with:
       ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
+      runner_arch: ${{ inputs.runner_arch || 'mi300' }}
       aiter_ref: ${{ inputs.aiter_ref }}
       continue_on_error: true
     secrets: inherit
 
   # =============================================== sgl-kernel ====================================================
   sgl-kernel-unit-test-amd:
-    name: ${{ format('sgl-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('sgl-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -235,7 +241,7 @@ jobs:
           needs.check-changes.outputs.sgl_kernel == 'true'
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -281,7 +287,7 @@ jobs:
           exit $failures
 
   sgl-kernel-unit-test-2-gpu-amd:
-    name: ${{ format('sgl-kernel-unit-test-2-gpu-amd (linux-{0}-2gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('sgl-kernel-unit-test-2-gpu-amd (linux-{0}-2gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -293,7 +299,7 @@ jobs:
           needs.check-changes.outputs.sgl_kernel == 'true'
         )
       )
-    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -322,7 +328,7 @@ jobs:
   # =============================================== primary ====================================================
 
   stage-a-test-1-gpu-small-amd:
-    name: ${{ format('stage-a-test-1-gpu-small-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('stage-a-test-1-gpu-small-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -334,7 +340,7 @@ jobs:
           ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -359,7 +365,7 @@ jobs:
           bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-a-test-1-gpu-small-amd ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   jit-kernel-unit-test-amd:
-    name: ${{ format('jit-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('jit-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -371,7 +377,7 @@ jobs:
           needs.check-changes.outputs.jit_kernel == 'true'
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -396,7 +402,7 @@ jobs:
           bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite jit-kernel-unit-test-amd ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   jit-kernel-benchmark-test-amd:
-    name: ${{ format('jit-kernel-benchmark-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('jit-kernel-benchmark-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -408,7 +414,7 @@ jobs:
           needs.check-changes.outputs.jit_kernel == 'true'
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -459,7 +465,7 @@ jobs:
           max-wait-minutes: '240'
 
   stage-b-test-1-gpu-small-amd:
-    name: ${{ format('stage-b-test-1-gpu-small-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('stage-b-test-1-gpu-small-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, wait-for-stage-a-amd]
     if: |
       always() &&
@@ -473,9 +479,10 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 4
       matrix:
         part: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -494,12 +501,12 @@ jobs:
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 60
+        timeout-minutes: ${{ (inputs.runner_arch || 'mi300') == 'mi300' && 75 || 60 }}
         run: |
           bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 14 --timeout-per-file 2400 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-b-test-1-gpu-small-amd-nondeterministic:
-    name: ${{ format('stage-b-test-1-gpu-small-amd-nondeterministic (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('stage-b-test-1-gpu-small-amd-nondeterministic (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, wait-for-stage-a-amd]
     if: |
       always() &&
@@ -511,7 +518,7 @@ jobs:
           ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -574,7 +581,7 @@ jobs:
           bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd-mi35x ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-b-test-1-gpu-large-amd:
-    name: ${{ format('stage-b-test-1-gpu-large-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('stage-b-test-1-gpu-large-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, wait-for-stage-a-amd]
     if: |
       always() &&
@@ -588,9 +595,10 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         part: [0, 1, 2]
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -609,12 +617,12 @@ jobs:
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 45
+        timeout-minutes: ${{ (inputs.runner_arch || 'mi300') == 'mi300' && 60 || 45 }}
         run: |
           bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 2700 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-b-test-2-gpu-large-amd:
-    name: ${{ format('stage-b-test-2-gpu-large-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('stage-b-test-2-gpu-large-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, wait-for-stage-a-amd]
     if: |
       always() &&
@@ -628,9 +636,10 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         part: [0, 1]
-    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -649,12 +658,12 @@ jobs:
         run: bash scripts/ci/amd/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 45
+        timeout-minutes: ${{ (inputs.runner_arch || 'mi300') == 'mi300' && 120 || 45 }}
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 2700 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
+          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file ${{ (inputs.runner_arch || 'mi300') == 'mi300' && 5400 || 2700 }} ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   multimodal-gen-test-1-gpu-amd:
-    name: ${{ format('multimodal-gen-test-1-gpu-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('multimodal-gen-test-1-gpu-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -668,9 +677,10 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         part: [0, 1, 2, 3]
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -699,16 +709,7 @@ jobs:
 
       - name: Setup kernel caches
         run: |
-          # Use the persistent /sgl-data directory (mounted from /home/runner/sgl-data)
-          # This directory persists across container restarts on the self-hosted runner
-          docker exec ci_sglang mkdir -p /sgl-data/aiter-kernels /sgl-data/miopen-cache /sgl-data/hf-cache/hub
-
-          # Clear pre-built AITER kernels from Docker image to avoid segfaults
-          # The image may have stale/incompatible kernels at /sgl-workspace/aiter/aiter/jit/
-          echo "Clearing pre-built AITER kernels from Docker image..."
-          docker exec ci_sglang rm -rf /sgl-workspace/aiter/aiter/jit/*.so 2>/dev/null || true
-          docker exec ci_sglang rm -rf /sgl-data/aiter-kernels/*.so 2>/dev/null || true
-          echo "AITER kernels cleared - will be rebuilt on first use"
+          docker exec ci_sglang mkdir -p /sgl-data/miopen-cache /sgl-data/hf-cache/hub
 
           # Create persistent cache marker if /sgl-data is a real mount (not ephemeral)
           # This tells the test cleanup code to NOT delete downloaded models
@@ -729,7 +730,7 @@ jobs:
           free -h
           echo ""
           echo "=== Disk Space ==="
-          df -h /home/runner/sgl-data 2>/dev/null || df -h
+          df -h /home/runner/sglang-data 2>/dev/null || df -h
           echo ""
           echo "=== HF Cache Directory Structure ==="
           docker exec ci_sglang ls -la /sgl-data/hf-cache/ 2>/dev/null || echo "HF cache dir not found"
@@ -768,7 +769,6 @@ jobs:
             -e SGLANG_SKIP_CONSISTENCY=1 \
             -e SGLANG_TEST_NUM_INFERENCE_STEPS=5 \
             -e SGLANG_DIFFUSION_ARTIFACT_DIR=/sglang-checkout/diffusion-failures \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -e HF_HUB_ENABLE_HF_TRANSFER=1 \
             -e HF_HUB_DISABLE_SYMLINKS_WARNING=1 \
@@ -794,7 +794,7 @@ jobs:
           retention-days: 7
 
   multimodal-gen-test-2-gpu-amd:
-    name: ${{ format('multimodal-gen-test-2-gpu-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('multimodal-gen-test-2-gpu-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -808,9 +808,10 @@ jobs:
       )
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         part: [0, 1, 2]  # 3 partitions: 2 parametrized + 1 standalone (single_test_file/test_disagg_server.py)
-    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -839,15 +840,7 @@ jobs:
 
       - name: Setup kernel caches
         run: |
-          # Use the persistent /sgl-data directory (mounted from /home/runner/sgl-data)
-          docker exec ci_sglang mkdir -p /sgl-data/aiter-kernels /sgl-data/miopen-cache /sgl-data/hf-cache/hub
-
-          # Clear pre-built AITER kernels from Docker image to avoid segfaults
-          # The image may have stale/incompatible kernels at /sgl-workspace/aiter/aiter/jit/
-          echo "Clearing pre-built AITER kernels from Docker image..."
-          docker exec ci_sglang rm -rf /sgl-workspace/aiter/aiter/jit/*.so 2>/dev/null || true
-          docker exec ci_sglang rm -rf /sgl-data/aiter-kernels/*.so 2>/dev/null || true
-          echo "AITER kernels cleared - will be rebuilt on first use"
+          docker exec ci_sglang mkdir -p /sgl-data/miopen-cache /sgl-data/hf-cache/hub
 
           # Create persistent cache marker if /sgl-data is a real mount (not ephemeral)
           # This tells the test cleanup code to NOT delete downloaded models
@@ -868,7 +861,7 @@ jobs:
           free -h
           echo ""
           echo "=== Disk Space ==="
-          df -h /home/runner/sgl-data 2>/dev/null || df -h
+          df -h /home/runner/sglang-data 2>/dev/null || df -h
           echo ""
           echo "=== HF Cache Directory Structure ==="
           docker exec ci_sglang ls -la /sgl-data/hf-cache/ 2>/dev/null || echo "HF cache dir not found"
@@ -890,7 +883,7 @@ jobs:
           docker exec ci_sglang rocm-smi --showmeminfo vram 2>/dev/null || echo "rocm-smi not available"
 
       - name: Run diffusion server tests (2-GPU)
-        timeout-minutes: 150
+        timeout-minutes: ${{ (inputs.runner_arch || 'mi300') == 'mi300' && 180 || 150 }}
         run: |
           # AMD CI: All 2-GPU tests including LoRA
           # Tests: T2V, T2I, I2V, LoRA
@@ -907,7 +900,6 @@ jobs:
             -e SGLANG_SKIP_CONSISTENCY=1 \
             -e SGLANG_TEST_NUM_INFERENCE_STEPS=5 \
             -e SGLANG_DIFFUSION_ARTIFACT_DIR=/sglang-checkout/diffusion-failures \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -e HF_HUB_ENABLE_HF_TRANSFER=1 \
             -e HF_HUB_DISABLE_SYMLINKS_WARNING=1 \
@@ -937,7 +929,7 @@ jobs:
   # that don't require NVIDIA hardware, so they should run on AMD too. This
   # closes the AMD coverage gap the dashboard surfaces for these tests.
   multimodal-gen-unit-test-amd:
-    name: ${{ format('multimodal-gen-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
+    name: ${{ format('multimodal-gen-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi300') }}
     needs: [check-changes, call-gate]
     if: |
       always() && !cancelled() &&
@@ -949,7 +941,7 @@ jobs:
           needs.check-changes.outputs.multimodal_gen == 'true'
         )
       )
-    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -983,7 +975,6 @@ jobs:
           # memory-format behavior that the ROCm conv path doesn't reproduce
           # (CUDA-specific). The rest of the unit suite is portable.
           docker exec \
-            -e AITER_JIT_DIR=/sgl-data/aiter-kernels \
             -e MIOPEN_USER_DB_PATH=/sgl-data/miopen-cache \
             -w /sglang-checkout/python \
             ci_sglang python3 sglang/multimodal_gen/test/run_suite.py \
@@ -1019,7 +1010,7 @@ jobs:
           max-wait-minutes: '480'
 
   stage-c-test-4-gpu-amd:
-    name: ${{ format('stage-c-test-4-gpu-amd (linux-{0}-4gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('stage-c-test-4-gpu-amd (linux-{0}-4gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, call-gate, wait-for-stage-b-amd]
     if: |
       always() &&
@@ -1035,7 +1026,7 @@ jobs:
       fail-fast: false
       matrix:
         part: [0]
-    runs-on: ${{ format('linux-{0}-4gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-4gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -1075,7 +1066,7 @@ jobs:
               ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-c-test-large-8-gpu-amd:
-    name: ${{ format('stage-c-test-large-8-gpu-amd (linux-{0}-8gpu-sglang, {1})', inputs.runner_arch || 'mi325', matrix.part) }}
+    name: ${{ format('stage-c-test-large-8-gpu-amd (linux-{0}-8gpu-sglang, {1})', inputs.runner_arch || 'mi300', matrix.part) }}
     needs: [check-changes, call-gate, wait-for-stage-b-amd]
     if: |
       always() &&
@@ -1088,12 +1079,13 @@ jobs:
         )
       )
     env:
-      RUNNER_LABELS: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || 'mi325') }}
+      RUNNER_LABELS: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || 'mi300') }}
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         part: [0, 1, 2, 3]
-    runs-on: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || 'mi325') }}
+    runs-on: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || 'mi300') }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
diff --git a/scripts/ci/amd/amd_ci_start_container.sh b/scripts/ci/amd/amd_ci_start_container.sh
index 21bdd07a5..68ee717d2 100755
--- a/scripts/ci/amd/amd_ci_start_container.sh
+++ b/scripts/ci/amd/amd_ci_start_container.sh
@@ -57,7 +57,7 @@ while [[ $# -gt 0 ]]; do
       echo ""
       echo "Environment:"
       echo "  ENABLE_CACHE_HOST=1|0"
-      echo "      Mount /home/runner/sglang-data to /sgl-data. Defaults to 0."
+      echo "      Mount /home/runner/sglang-data to /sgl-data. Defaults to 1 when RUNNER_NAME contains 300 or 35x, otherwise 0. Missing host cache falls back to container-local /sgl-data."
       exit 0
       ;;
     *) echo "Unknown option $1"; exit 1;;
@@ -288,15 +288,24 @@ else
 fi
 
 CACHE_HOST=/home/runner/sglang-data
-ENABLE_CACHE_HOST="${ENABLE_CACHE_HOST:-0}"
+if [[ -z "${ENABLE_CACHE_HOST:-}" ]]; then
+  RUNNER_NAME_LOWER="${RUNNER_NAME:-}"
+  RUNNER_NAME_LOWER="${RUNNER_NAME_LOWER,,}"
+  if [[ "${RUNNER_NAME_LOWER}" == *300* || "${RUNNER_NAME_LOWER}" == *35x* ]]; then
+    ENABLE_CACHE_HOST="1"
+  else
+    ENABLE_CACHE_HOST="0"
+  fi
+fi
 case "${ENABLE_CACHE_HOST,,}" in
   1|true|yes|on|pvc|persistent)
-    if [[ ! -d "$CACHE_HOST" ]]; then
-      echo "Error: ENABLE_CACHE_HOST=1 but ${CACHE_HOST} does not exist." >&2
-      exit 1
+    if [[ -d "$CACHE_HOST" ]]; then
+      CACHE_VOLUME="-v $CACHE_HOST:/sgl-data"
+      echo "Mounting persistent CI data: ${CACHE_HOST} -> /sgl-data"
+    else
+      CACHE_VOLUME=""
+      echo "Warning: ${CACHE_HOST} does not exist; using container-local /sgl-data." >&2
     fi
-    CACHE_VOLUME="-v $CACHE_HOST:/sgl-data"
-    echo "Mounting persistent CI data: ${CACHE_HOST} -> /sgl-data"
     ;;
   0|false|no|off|"")
     CACHE_VOLUME=""
@@ -332,8 +341,7 @@ docker run -dt --user root --device=/dev/kfd ${DEVICE_FLAG} \
 docker exec ci_sglang mkdir -p \
   /sgl-data/hf-cache/hub \
   /sgl-data/pip-cache \
-  /sgl-data/miopen-cache \
-  /sgl-data/aiter-kernels
+  /sgl-data/miopen-cache
 
 # The checkout is owned by the runner (non-root) but the container runs as
 # root.  Git >= 2.35.2 rejects cross-user repos; mark the mount as safe so
diff --git a/scripts/ci/amd/amd_ci_start_container_disagg.sh b/scripts/ci/amd/amd_ci_start_container_disagg.sh
index 7d404619b..e46c2a43d 100755
--- a/scripts/ci/amd/amd_ci_start_container_disagg.sh
+++ b/scripts/ci/amd/amd_ci_start_container_disagg.sh
@@ -228,13 +228,36 @@ else
   retry_with_backoff 6 docker pull "${IMAGE}"
 fi
 
-# CACHE_HOST=/home/runner/sgl-data
-CACHE_HOST=/home/runner/temp-sglang-data
-if [[ -d "$CACHE_HOST" ]]; then
-    CACHE_VOLUME="-v $CACHE_HOST:/sgl-data"
-else
-    CACHE_VOLUME=""
+CACHE_HOST=/home/runner/sglang-data
+if [[ -z "${ENABLE_CACHE_HOST:-}" ]]; then
+  RUNNER_NAME_LOWER="${RUNNER_NAME:-}"
+  RUNNER_NAME_LOWER="${RUNNER_NAME_LOWER,,}"
+  if [[ "${RUNNER_NAME_LOWER}" == *300* || "${RUNNER_NAME_LOWER}" == *35x* ]]; then
+    ENABLE_CACHE_HOST="1"
+  else
+    ENABLE_CACHE_HOST="0"
+  fi
 fi
+case "${ENABLE_CACHE_HOST,,}" in
+  1|true|yes|on|pvc|persistent)
+    if [[ -d "$CACHE_HOST" ]]; then
+      CACHE_VOLUME="-v $CACHE_HOST:/sgl-data"
+      echo "Mounting persistent CI data: ${CACHE_HOST} -> /sgl-data"
+    else
+      CACHE_VOLUME=""
+      echo "Warning: ${CACHE_HOST} does not exist; using container-local /sgl-data." >&2
+    fi
+    ;;
+  0|false|no|off|"")
+    CACHE_VOLUME=""
+    echo "Not mounting ${CACHE_HOST}; /sgl-data will be container-local."
+    ;;
+  *)
+    echo "Error: unsupported ENABLE_CACHE_HOST='${ENABLE_CACHE_HOST}'" >&2
+    echo "Use 1/true/pvc/persistent or 0/false/off." >&2
+    exit 1
+    ;;
+esac
 
 # Detect libionic library for RDMA support
 LIBIONIC_MOUNT=""
@@ -325,6 +348,11 @@ docker run -dt --user root \
   --name ci_sglang \
   "${IMAGE}"
 
+docker exec ci_sglang mkdir -p \
+  /sgl-data/hf-cache/hub \
+  /sgl-data/pip-cache \
+  /sgl-data/miopen-cache
+
 # The checkout is owned by the runner (non-root) but the container runs as
 # root.  Git >= 2.35.2 rejects cross-user repos; mark the mount as safe so
 # setuptools-scm / vcs_versioning can resolve the package version.
diff --git a/test/registered/amd/accuracy/mi30x/test_minimax_m27_eval_amd.py b/test/registered/amd/accuracy/mi30x/test_minimax_m27_eval_amd.py
index b272986ac..c33e93eb8 100644
--- a/test/registered/amd/accuracy/mi30x/test_minimax_m27_eval_amd.py
+++ b/test/registered/amd/accuracy/mi30x/test_minimax_m27_eval_amd.py
@@ -167,7 +167,7 @@ class TestMiniMaxM27EvalAMD(unittest.TestCase):
     def setUpClass(cls):
         cls.models = MINIMAX_M27_MODELS
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "1319"))
 
     def test_minimax_m27_accuracy(self):
         """Test MiniMax-M2.7 with GSM8K completion benchmark."""
diff --git a/test/registered/perf/test_bench_serving_1gpu_part1.py b/test/registered/perf/test_bench_serving_1gpu_part1.py
index ef257e4b7..09bc9905f 100644
--- a/test/registered/perf/test_bench_serving_1gpu_part1.py
+++ b/test/registered/perf/test_bench_serving_1gpu_part1.py
@@ -117,7 +117,7 @@ class TestBenchServing1GPUPart1(CustomTestCase):
                 f"Output throughput: {res['output_throughput']:.2f} token/s\n"
             )
             if is_in_amd_ci():
-                self.assertGreater(res["output_throughput"], 3500)
+                self.assertGreater(res["output_throughput"], 2700)
             else:
                 self.assertGreater(res["output_throughput"], 3700)
 
@@ -150,7 +150,10 @@ class TestBenchServing1GPUPart1(CustomTestCase):
                 f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
                 f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
             )
-            self.assertLess(res["median_e2e_latency_ms"], 2400)
+            if is_in_amd_ci():
+                self.assertLess(res["median_e2e_latency_ms"], 3320)
+            else:
+                self.assertLess(res["median_e2e_latency_ms"], 2400)
             # relax for mi300x (LoRA TTFT ~2x slower than mi325)
             if is_in_amd_ci():
                 self.assertLess(res["median_ttft_ms"], 100)
@@ -166,7 +169,10 @@ class TestBenchServing1GPUPart1(CustomTestCase):
                 f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
                 f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
             )
-            self.assertLess(res["median_e2e_latency_ms"], 4000)
+            if is_in_amd_ci():
+                self.assertLess(res["median_e2e_latency_ms"], 6000)
+            else:
+                self.assertLess(res["median_e2e_latency_ms"], 4000)
             # relax for mi300x (LoRA TTFT ~2x slower than mi325)
             if is_in_amd_ci():
                 self.assertLess(res["median_ttft_ms"], 130)