diff --git a/.github/workflows/pr-test-amd-rocm720.yml b/.github/workflows/pr-test-amd-rocm720.yml
index e9d83d2e1..7df6e3251 100644
--- a/.github/workflows/pr-test-amd-rocm720.yml
+++ b/.github/workflows/pr-test-amd-rocm720.yml
@@ -898,7 +898,7 @@ jobs:
       fail-fast: false
       matrix:
         runner: [linux-mi325-8gpu-sglang]
-        part: [0, 1, 2]
+        part: [0, 1, 2, 3]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -925,7 +925,7 @@ jobs:
       - name: Run test
         timeout-minutes: 120
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 5400 ${{ 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-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 4 --timeout-per-file 5400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }}
 
   stage-c-test-large-8-gpu-amd-mi35x-rocm720:
     needs: [check-changes]
diff --git a/.github/workflows/pr-test-amd.yml b/.github/workflows/pr-test-amd.yml
index 08f67b88c..69123f488 100644
--- a/.github/workflows/pr-test-amd.yml
+++ b/.github/workflows/pr-test-amd.yml
@@ -997,7 +997,7 @@ jobs:
       - name: Run test
         timeout-minutes: 120
         run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 5400 ${{ 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-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 4 --timeout-per-file 5400 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
 
   stage-c-test-large-8-gpu-amd-mi35x:
     needs: [check-changes, call-gate, wait-for-stage-b-amd]
diff --git a/python/sglang/srt/layers/attention/triton_backend.py b/python/sglang/srt/layers/attention/triton_backend.py
index 01c27f3e8..f280f8b9f 100644
--- a/python/sglang/srt/layers/attention/triton_backend.py
+++ b/python/sglang/srt/layers/attention/triton_backend.py
@@ -22,10 +22,12 @@ from sglang.srt.utils import (
     get_device_core_count,
     get_int_env_var,
     is_cuda,
+    is_gfx942_supported,
     next_power_of_2,
 )
 
 _is_cuda = is_cuda()
+_is_gfx942 = is_gfx942_supported()
 
 if _is_cuda:
     from sgl_kernel.utils import is_arch_support_pdl
@@ -163,6 +165,15 @@ class TritonAttnBackend(AttentionBackend):
                 self.device_core_count,
                 self.max_context_len,
             )
+            if _is_gfx942:
+                # gfx942 (MI300X / MI325X) has 304 CUs, so #20479's next_power_of_2(sm_count)
+                # rounds up to 512 — twice MI355X's natural cap of 256 — and the persistent
+                # cuda_graph_attn_logits fp32 buffer hits ~4 GiB on Kimi-K2.6 (v_head_dim=512),
+                # faulting in ROCm CUDA graph replay
+                # (https://github.com/sgl-project/sglang/actions/runs/25513282022/job/74877480809).
+                # Pin the cap at 256 so gfx942 matches the gfx950 (MI355X) behavior that we
+                # already validated end-to-end.
+                self.max_kv_splits = min(self.max_kv_splits, 256)
         if _is_cuda:
             self.use_pdl = is_arch_support_pdl()
         else:
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index bdf496316..85f576ac2 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -3656,6 +3656,18 @@ def is_gfx95_supported():
         return False
 
 
+@lru_cache(maxsize=1)
+def is_gfx942_supported():
+    """
+    Returns whether the current platform is AMD CDNA3 (gfx942 — MI300X / MI325X).
+    """
+    if torch.version.hip:
+        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
+        return any(gfx in gcn_arch for gfx in ["gfx942"])
+    else:
+        return False
+
+
 def get_hip_version():
     if torch.version.hip:
         return tuple(map(int, torch.version.hip.split("-")[0].split(".")))
diff --git a/test/registered/amd/test_kimi_k2_instruct.py b/test/registered/amd/test_kimi_k2_instruct.py
index 896d3b154..bb2c57c3d 100644
--- a/test/registered/amd/test_kimi_k2_instruct.py
+++ b/test/registered/amd/test_kimi_k2_instruct.py
@@ -63,7 +63,7 @@ class TestKimiK2Instruct0905(CustomTestCase):
             num_shots=8,
             data_path=None,
             num_questions=1319,
-            parallel=1319,
+            parallel=512,
             max_new_tokens=512,
             host="http://127.0.0.1",
             port=int(self.base_url.split(":")[-1]),