[AMD] Add Kimi-K2, DeepSeek-V3.2 tests to nightly CI (#17523)

Co-authored-by: YC Tseng <yctseng@amd.com>

.github/workflows/nightly-test-amd.yml

@@ -25,11 +25,13 @@ on:
           - 'nightly-perf-2-gpu-text'
           - 'nightly-perf-2-gpu-vlm'
           - 'nightly-accuracy-8-gpu'
-          - 'nightly-accuracy-8-gpu-deepseek-r1'
           # MI30x Accuracy + Performance Tests (combined)
           - 'nightly-8-gpu-grok1-int4'
           - 'nightly-8-gpu-grok2'
           - 'nightly-8-gpu-deepseek-v31'
+          - 'nightly-8-gpu-deepseek-v32'
+          - 'nightly-8-gpu-deepseek-v32-mtp'
+          - 'nightly-8-gpu-kimi-k2'
           # MI35x jobs
           - 'nightly-test-1-gpu-mi35x'
           - 'nightly-accuracy-8-gpu-mi35x'
@@ -37,6 +39,7 @@ on:
           - 'nightly-8-gpu-mi35x-grok2'
           - 'nightly-8-gpu-mi35x-deepseek-r1-mxfp4'
           - 'nightly-accuracy-8-gpu-mi35x-deepseek-v32'
+          - 'nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp'
           - 'nightly-perf-8-gpu-mi35x-deepseek-v32-basic'
           - 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp'
   workflow_call:
@@ -248,35 +251,6 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
-  # 8-GPU DeepSeek-R1 Accuracy Test (separate job due to long loading time)
-  nightly-accuracy-8-gpu-deepseek-r1:
-    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-accuracy-8-gpu-deepseek-r1')
-    runs-on: linux-mi325-gpu-8
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ inputs.ref || github.ref }}
-
-      - name: Setup docker
-        run: |
-          touch github_summary.md
-          bash scripts/ci/amd/amd_ci_start_container.sh
-        env:
-          GITHUB_WORKSPACE: ${{ github.workspace }}
-
-      - name: Install dependencies
-        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
-
-      - name: Accuracy Test (8-GPU DeepSeek-R1)
-        timeout-minutes: 240
-        run: |
-          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-r1 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
-          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
-          exit ${TEST_EXIT_CODE:-0}
-
   # ============================================== MI30x Combined Accuracy + Performance Tests ==============================================
   # 8-GPU Grok1-INT4 (Accuracy + Performance combined)
   nightly-8-gpu-grok1-int4:
@@ -407,6 +381,118 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # 8-GPU DeepSeek-V3.2 (Basic Accuracy + Perf)
+  nightly-8-gpu-deepseek-v32:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-deepseek-v32')
+    runs-on: linux-mi325-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Accuracy Test (8-GPU DeepSeek-V3.2 Basic)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v32 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+      - name: Performance Test (8-GPU DeepSeek-V3.2 Basic)
+        timeout-minutes: 150
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-deepseek-v32-basic --nightly --timeout-per-file 5400 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  # 8-GPU DeepSeek-V3.2 MTP (MTP Accuracy + Perf)
+  nightly-8-gpu-deepseek-v32-mtp:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-deepseek-v32-mtp')
+    runs-on: linux-mi325-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Accuracy Test (8-GPU DeepSeek-V3.2 MTP)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v32-mtp --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+      - name: Performance Test (8-GPU DeepSeek-V3.2 MTP)
+        timeout-minutes: 180
+        continue-on-error: true  # Perf test failure doesn't fail the job if accuracy passed
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-deepseek-v32-mtp --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  # 8-GPU Kimi-K2 (Accuracy + Speed)
+  nightly-8-gpu-kimi-k2:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-kimi-k2')
+    runs-on: linux-mi325-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Accuracy Test (8-GPU Kimi-K2)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-kimi-k2 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # ============================================== MI35x Tests ==============================================
   # MI35x 1-GPU tests - platform-agnostic tests that may work on CDNA4 (gfx950)
   nightly-test-1-gpu-mi35x:
@@ -641,6 +727,39 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # MI35x 8-GPU DeepSeek-V3.2 TP+MTP Accuracy Test
+  nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp')
+    runs-on: linux-mi35x-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd/amd_ci_install_dependency.sh
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x (8-GPU DeepSeek-V3.2 TP+MTP)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # MI35x 8-GPU DeepSeek-V3.2 Performance Test (Basic)
   nightly-perf-8-gpu-mi35x-deepseek-v32-basic:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-perf-8-gpu-mi35x-deepseek-v32-basic')
@@ -698,12 +817,12 @@ jobs:
           bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
 
       - name: Performance Test MI35x (8-GPU DeepSeek-V3.2 MTP)
-        timeout-minutes: 150
+        timeout-minutes: 180
         run: |
           > github_summary.md  # Clear summary file
           bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
             -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 5400 || TEST_EXIT_CODE=$?
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
@@ -719,11 +838,13 @@ jobs:
       - nightly-perf-2-gpu-text
       - nightly-perf-2-gpu-vlm
       - nightly-accuracy-8-gpu
-      - nightly-accuracy-8-gpu-deepseek-r1
       # MI30x Combined Accuracy + Performance Tests
       - nightly-8-gpu-grok1-int4
       - nightly-8-gpu-grok2
       - nightly-8-gpu-deepseek-v31
+      - nightly-8-gpu-deepseek-v32
+      - nightly-8-gpu-deepseek-v32-mtp
+      - nightly-8-gpu-kimi-k2
       # MI35x jobs
       - nightly-test-1-gpu-mi35x
       - nightly-accuracy-8-gpu-mi35x
@@ -731,8 +852,10 @@ jobs:
       - nightly-8-gpu-mi35x-grok2
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32
-      - nightly-perf-8-gpu-mi35x-deepseek-v32-basic
+      - nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp
-      - nightly-perf-8-gpu-mi35x-deepseek-v32-mtp
+      # MI35x perf jobs excluded from check - perf failures don't block CI
+      # - nightly-perf-8-gpu-mi35x-deepseek-v32-basic
+      # - nightly-perf-8-gpu-mi35x-deepseek-v32-mtp
     runs-on: ubuntu-latest
     steps:
       - name: Check if any job failed

test/registered/amd/accuracy/mi35x/test_deepseek_r1_eval_mi35x.py

@@ -214,6 +214,9 @@ class TestDeepSeekR1EvalMI35x(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/mi35x/test_deepseek_r1_mxfp4_eval_mi35x.py

@@ -239,6 +239,9 @@ class TestDeepSeekR1MXFP4EvalMI35x(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/mi35x/test_deepseek_v32_dp_eval_mi35x.py

@@ -0,0 +1,119 @@
+"""MI35x DeepSeek-V3.2 DP GSM8K Accuracy Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with DP=8 + TP=8 + dp-attention using few-shot
+completion benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-dp suite
+"""
+
+import os
+
+# Set HF cache for MI35x
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - MI35x DeepSeek-V3.2 DP accuracy test
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-dp",
+    nightly=True,
+)
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+# Accuracy threshold
+GSM8K_ACCURACY_THRESHOLD = 0.935
+
+
+class TestDeepseekV32DP(CustomTestCase):
+    """Test DeepSeek V3.2 with DP=8 + TP=8 + dp-attention.
+
+    This test runs GSM8K evaluation and measures accuracy on MI35x.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--dp",
+            "8",
+            "--enable-dp-attention",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(self):
+        """GSM8K evaluation for DP configuration.
+
+        Named with 'a' prefix to run first (alphabetically) to warm up the server.
+        """
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v32 DP MI35x)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], GSM8K_ACCURACY_THRESHOLD)
+
+    def test_bs_1_speed(self):
+        """Single batch speed test for DP configuration."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v32 DP MI35x)\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(speed, 10)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi35x/test_deepseek_v32_eval_mi35x.py

@@ -215,6 +215,9 @@ class TestDeepSeekV32EvalMI35x(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/mi35x/test_deepseek_v32_mtp_eval_mi35x.py

@@ -0,0 +1,142 @@
+"""MI35x DeepSeek-V3.2 TP+MTP GSM8K Accuracy Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with TP=8 + MTP (EAGLE speculative decoding) using few-shot
+completion benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-mtp suite
+"""
+
+import os
+
+# Set HF cache for MI35x
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - MI35x DeepSeek-V3.2 TP+MTP accuracy test
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-mtp",
+    nightly=True,
+)
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+# Accuracy and performance thresholds
+GSM8K_ACCURACY_THRESHOLD = 0.94
+AVG_SPEC_ACCEPT_LENGTH_THRESHOLD = 2.7
+
+
+class TestDeepseekV32TPMTP(CustomTestCase):
+    """Test DeepSeek V3.2 with TP=8 + MTP (EAGLE speculative decoding).
+
+    This test runs GSM8K evaluation and measures both accuracy and
+    speculative decoding acceptance length on MI35x.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+            "--mem-frac",
+            "0.7",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(self):
+        """GSM8K evaluation for TP+MTP configuration.
+
+        Named with 'a' prefix to run first (alphabetically) to warm up the server.
+        """
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
+        print(f"{avg_spec_accept_length=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v32 TP+MTP MI35x)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+            self.assertGreater(metrics["accuracy"], GSM8K_ACCURACY_THRESHOLD)
+            self.assertGreater(avg_spec_accept_length, AVG_SPEC_ACCEPT_LENGTH_THRESHOLD)
+
+    def test_bs_1_speed(self):
+        """Single batch speed test for TP+MTP configuration."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{acc_length=:.2f} {speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v32 TP+MTP MI35x)\n"
+                f"{acc_length=:.2f}\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(acc_length, AVG_SPEC_ACCEPT_LENGTH_THRESHOLD)
+            self.assertGreater(speed, 55)  # Lowered from 60 for AMD MI35x
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi35x/test_gpt_oss_eval_mi35x.py

@@ -218,6 +218,9 @@ class TestGptOssEvalMI35x(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/mi35x/test_grok1_int4_eval_mi35x.py

@@ -140,6 +140,7 @@ class TestGrok1INT4EvalMI35x(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### GROK1-INT4 (MI35x)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/mi35x/test_grok2_eval_mi35x.py

@@ -142,6 +142,7 @@ class TestGrok2EvalMI35x(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### GROK2 (MI35x)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/test_deepseek_r1_eval_amd.py

@@ -273,6 +273,9 @@ class TestDeepSeekR1EvalAMD(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/test_deepseek_v31_eval_amd.py

@@ -137,6 +137,7 @@ class TestDeepSeekV31EvalAMD(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### DeepSeek-V3.1 (MI300X)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/test_deepseek_v32_dp_eval_amd.py

@@ -0,0 +1,122 @@
+"""AMD DeepSeek-V3.2 DP GSM8K Accuracy Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with DP=8 + TP=8 + dp-attention using few-shot
+completion benchmark on MI325/MI300X.
+
+Registry: nightly-amd-accuracy-8-gpu-deepseek-v32-dp suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - DeepSeek-V3.2 DP accuracy test
+register_amd_ci(
+    est_time=5400,
+    suite="nightly-amd-accuracy-8-gpu-deepseek-v32-dp",
+    nightly=True,
+)
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+# Accuracy threshold
+GSM8K_ACCURACY_THRESHOLD = 0.935
+
+
+class TestDeepseekV32DP(CustomTestCase):
+    """Test DeepSeek V3.2 with DP=8 + TP=8 + dp-attention.
+
+    This test runs GSM8K evaluation and measures accuracy on MI325/MI300X.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--dp",
+            "8",
+            "--enable-dp-attention",
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.85",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_USE_ROCM700A"] = "1"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(self):
+        """GSM8K evaluation for DP configuration.
+
+        Named with 'a' prefix to run first (alphabetically) to warm up the server.
+        """
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v32 DP MI325)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], GSM8K_ACCURACY_THRESHOLD)
+
+    def test_bs_1_speed(self):
+        """Single batch speed test for DP configuration."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v32 DP MI325)\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(speed, 10)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_deepseek_v32_eval_amd.py

@@ -0,0 +1,248 @@
+"""AMD DeepSeek-V3.2 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with basic configuration using few-shot completion
+benchmark on MI325/MI300X.
+
+Registry: nightly-amd-accuracy-8-gpu-deepseek-v32 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+# Register for AMD CI - DeepSeek-V3.2 accuracy test (~60 min for basic only)
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-accuracy-8-gpu-deepseek-v32",
+    nightly=True,
+)
+
+INVALID = -9999999
+
+
+@dataclass
+class ModelConfig:
+    """Configuration for a model to test."""
+
+    model_path: str
+    tp_size: int = 8
+    accuracy_threshold: float = 0.50
+    other_args: Optional[List[str]] = None
+    env_vars: Optional[dict] = None
+    timeout: Optional[int] = None
+    variant: Optional[str] = None
+
+    def __post_init__(self):
+        if self.other_args is None:
+            self.other_args = []
+        if self.env_vars is None:
+            self.env_vars = {}
+
+    def get_display_name(self) -> str:
+        if self.variant:
+            return f"{self.model_path} ({self.variant})"
+        return self.model_path
+
+
+# DeepSeek-V3.2 models for MI325/MI300X - basic variant
+DEEPSEEK_V32_MODELS = [
+    # DeepSeek-V3.2 basic (TP=8 only)
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-V3.2",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=3600,
+        variant="basic",
+        other_args=[
+            "--trust-remote-code",
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.85",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",  # 20 minutes for weight loading
+        ],
+        env_vars={"SGLANG_USE_AITER": "1"},
+    ),
+]
+
+
+def get_one_example(lines, i, include_answer):
+    """Format a single GSM8K example."""
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    """Get k few-shot examples for prompting."""
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    """Extract numerical answer from response."""
+    answer_str = answer_str.replace(",", "")
+    numbers = re.findall(r"\d+", answer_str)
+    if len(numbers) < 1:
+        return INVALID
+    try:
+        return ast.literal_eval(numbers[-1])
+    except SyntaxError:
+        return INVALID
+
+
+def run_gsm8k_benchmark(
+    base_url: str,
+    num_questions: int = 200,
+    num_shots: int = 5,
+    parallel: int = 64,
+) -> Tuple[float, float, float]:
+    """Run GSM8K few-shot completion benchmark."""
+    import sglang as sgl
+    from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
+
+    url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
+    data_path = download_and_cache_file(url)
+    lines = list(read_jsonl(data_path))
+
+    few_shot_examples = get_few_shot_examples(lines, num_shots)
+
+    questions = []
+    labels = []
+    for i in range(len(lines[:num_questions])):
+        questions.append(get_one_example(lines, i, False))
+        labels.append(get_answer_value(lines[i]["answer"]))
+    assert all(l != INVALID for l in labels)
+    arguments = [{"question": q} for q in questions]
+
+    @sgl.function
+    def few_shot_gsm8k(s, question):
+        s += few_shot_examples + question
+        s += sgl.gen(
+            "answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
+        )
+
+    backend = RuntimeEndpoint(base_url)
+    sgl.set_default_backend(backend)
+
+    tic = time.perf_counter()
+    states = few_shot_gsm8k.run_batch(
+        arguments, temperature=0, num_threads=parallel, progress_bar=True
+    )
+    latency = time.perf_counter() - tic
+
+    preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
+    acc = np.mean(np.array(preds) == np.array(labels))
+    invalid = np.mean(np.array(preds) == INVALID)
+
+    return float(acc), float(invalid), float(latency)
+
+
+class TestDeepSeekV32EvalAMD(unittest.TestCase):
+    """DeepSeek-V3.2 GSM8K Completion Evaluation Test for AMD MI325/MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = DEEPSEEK_V32_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_deepseek_v32_accuracy(self):
+        """Test DeepSeek-V3.2 models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### DeepSeek-V3.2 Models (MI325)\n\n"
+        summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            display_name = config.get_display_name()
+            with self.subTest(model=display_name):
+                print(f"\n{'='*60}")
+                print(f"Testing: {display_name}")
+                print(f"{'='*60}")
+
+                env = os.environ.copy()
+                for key, value in config.env_vars.items():
+                    env[key] = value
+
+                other_args = list(config.other_args)
+                other_args.extend(["--tp", str(config.tp_size)])
+                timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+                try:
+                    process = popen_launch_server(
+                        model=config.model_path,
+                        base_url=self.base_url,
+                        timeout=timeout,
+                        other_args=other_args,
+                        env=env,
+                    )
+
+                    try:
+                        acc, invalid, latency = run_gsm8k_benchmark(
+                            self.base_url, num_questions=self.num_questions
+                        )
+                        passed = acc >= config.accuracy_threshold
+                        status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": display_name,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": display_name,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_deepseek_v32_mtp_eval_amd.py

@@ -0,0 +1,142 @@
+"""AMD DeepSeek-V3.2 TP+MTP GSM8K Accuracy Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with TP=8 + MTP (EAGLE speculative decoding) using few-shot
+completion benchmark on MI325/MI300X.
+
+Registry: nightly-amd-accuracy-8-gpu-deepseek-v32-mtp suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - DeepSeek-V3.2 TP+MTP accuracy test
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-accuracy-8-gpu-deepseek-v32-mtp",
+    nightly=True,
+)
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+# Accuracy and performance thresholds
+GSM8K_ACCURACY_THRESHOLD = 0.94
+AVG_SPEC_ACCEPT_LENGTH_THRESHOLD = 2.7
+
+
+class TestDeepseekV32TPMTP(CustomTestCase):
+    """Test DeepSeek V3.2 with TP=8 + MTP (EAGLE speculative decoding).
+
+    This test runs GSM8K evaluation and measures both accuracy and
+    speculative decoding acceptance length on MI325/MI300X.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+            "--mem-fraction-static",
+            "0.7",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(self):
+        """GSM8K evaluation for TP+MTP configuration.
+
+        Named with 'a' prefix to run first (alphabetically) to warm up the server.
+        """
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
+        print(f"{avg_spec_accept_length=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v32 TP+MTP MI325)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+            self.assertGreater(metrics["accuracy"], GSM8K_ACCURACY_THRESHOLD)
+            self.assertGreater(avg_spec_accept_length, AVG_SPEC_ACCEPT_LENGTH_THRESHOLD)
+
+    def test_bs_1_speed(self):
+        """Single batch speed test for TP+MTP configuration."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{acc_length=:.2f} {speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v32 TP+MTP MI325)\n"
+                f"{acc_length=:.2f}\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(acc_length, AVG_SPEC_ACCEPT_LENGTH_THRESHOLD)
+            self.assertGreater(speed, 55)  # Lowered from 60 for AMD MI325
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_deepseek_v32_tc_eval_amd.py

@@ -0,0 +1,123 @@
+"""AMD DeepSeek-V3.2 TC GSM8K Accuracy Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with Torch Compile configuration using few-shot
+completion benchmark on MI325/MI300X.
+
+Registry: nightly-amd-accuracy-8-gpu-deepseek-v32-tc suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - DeepSeek-V3.2 TC accuracy test
+register_amd_ci(
+    est_time=7200,
+    suite="nightly-amd-accuracy-8-gpu-deepseek-v32-tc",
+    nightly=True,
+)
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+# Accuracy threshold
+GSM8K_ACCURACY_THRESHOLD = 0.935
+
+
+class TestDeepseekV32TC(CustomTestCase):
+    """Test DeepSeek V3.2 with Torch Compile.
+
+    This test runs GSM8K evaluation and measures accuracy on MI325/MI300X.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.70",
+            "--cuda-graph-max-bs",
+            "8",
+            "--enable-torch-compile",
+            "--disable-cuda-graph",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_USE_ROCM700A"] = "1"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(self):
+        """GSM8K evaluation for TC configuration.
+
+        Named with 'a' prefix to run first (alphabetically) to warm up the server.
+        """
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v32 TC MI325)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], GSM8K_ACCURACY_THRESHOLD)
+
+    def test_bs_1_speed(self):
+        """Single batch speed test for TC configuration."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v32 TC MI325)\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(speed, 10)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_gpt_oss_eval_amd.py

@@ -56,7 +56,7 @@ GPT_OSS_MODELS = [
     ModelConfig(
         model_path="lmsys/gpt-oss-20b-bf16",
         tp_size=8,
-        accuracy_threshold=0.47,
+        accuracy_threshold=0.45,
         other_args=[
             "--chunked-prefill-size",
             "130172",
@@ -211,6 +211,9 @@ class TestGptOssEvalAMD(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/test_grok1_fp8_eval_amd.py

@@ -140,6 +140,7 @@ class TestGrok1FP8EvalAMD(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### GROK1-FP8 (MI300X)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/test_grok1_int4_eval_amd.py

@@ -140,6 +140,7 @@ class TestGrok1INT4EvalAMD(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### GROK1-INT4 (MI300X)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/test_grok2_eval_amd.py

@@ -140,6 +140,7 @@ class TestGrok2EvalAMD(unittest.TestCase):
             )
             passed = acc >= self.accuracy_threshold
             status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={self.accuracy_threshold} {status}")
 
             summary = f"### GROK2 (MI300X)\n\n"
             summary += f"| Model | Accuracy | Threshold | Status |\n"

test/registered/amd/accuracy/test_grok_eval_amd.py

@@ -251,6 +251,9 @@ class TestGrokEvalAMD(unittest.TestCase):
                         )
                         passed = acc >= config.accuracy_threshold
                         status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
 
                         all_results.append(
                             {

test/registered/amd/accuracy/test_gsm8k_eval_amd.py

@@ -59,9 +59,9 @@ MODEL_SCORE_THRESHOLDS = {
     "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.8,
     "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.54,
     "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8": 0.94,
-    "neuralmagic/Qwen2-72B-Instruct-FP8": 0.94,
+    "neuralmagic/Qwen2-72B-Instruct-FP8": 0.92,
-    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.86,
+    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.81,
-    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.62,
+    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.57,
     "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8": 0.84,
 }
 

test/registered/amd/accuracy/test_kimi_k2_eval_amd.py

@@ -0,0 +1,101 @@
+"""AMD Kimi-K2 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests moonshotai/Kimi-K2-Instruct-0905 with GSM8K few-shot benchmark on MI325.
+
+Registry: nightly-amd-accuracy-8-gpu-kimi-k2 suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - Kimi K2 accuracy test (~60 min)
+register_amd_ci(est_time=3600, suite="nightly-amd-accuracy-8-gpu-kimi-k2", nightly=True)
+
+KIMI_K2_MODEL_PATH = "moonshotai/Kimi-K2-Instruct-0905"
+SERVER_LAUNCH_TIMEOUT = 3600
+ACCURACY_THRESHOLD = 0.94
+
+
+class TestKimiK2EvalAMD(CustomTestCase):
+    """Kimi-K2 GSM8K Completion Evaluation Test for AMD MI325."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = KIMI_K2_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "8",
+            "--decode-attention-backend",
+            "triton",
+            "--prefill-attention-backend",
+            "aiter",
+            "--trust-remote-code",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_ROCM_FUSED_DECODE_MLA"] = "0"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_kimi_k2_gsm8k_accuracy(self):
+        """Test Kimi-K2 with GSM8K few-shot completion benchmark."""
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=8,
+            data_path=None,
+            num_questions=1319,
+            parallel=1319,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        acc = metrics["accuracy"]
+
+        passed = acc >= ACCURACY_THRESHOLD
+        status = "✅ PASS" if passed else "❌ FAIL"
+        print(f"  accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")
+
+        if is_in_ci():
+            summary = "### Kimi-K2 Model (MI325)\n\n"
+            summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+            summary += "| ----- | -- | -------- | --------- | ------ |\n"
+            summary += f"| {KIMI_K2_MODEL_PATH} | 8 | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
+            write_github_step_summary(summary)
+
+        self.assertGreaterEqual(
+            acc,
+            ACCURACY_THRESHOLD,
+            f"Kimi-K2 accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_vlms_mmmu_eval_amd.py

@@ -127,7 +127,8 @@ TRITON_ATTENTION_MODELS = {
 
 # Models known to fail on AMD - exclude from testing
 AMD_FAILING_VLM_MODELS = {
-    # Add models here as they are discovered to fail
+    # GLM-4.1V processor not registered yet (Glm4vForConditionalGeneration)
+    "zai-org/GLM-4.1V-9B-Thinking",
 }
 
 

test/registered/amd/perf/mi35x/test_deepseek_v32_basic_perf_mi35x.py

@@ -93,6 +93,8 @@ class TestNightlyDeepseekV32BasicPerformance(unittest.TestCase):
                 "0.85",
                 "--model-loader-extra-config",
                 '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
             ],
         }
 

test/registered/amd/perf/mi35x/test_deepseek_v32_mtp_perf_mi35x.py

@@ -13,12 +13,17 @@ Example usage:
 
 import os
 import unittest
-from typing import List
+from typing import List, Optional, Tuple
 
+from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_amd_ci
 from sglang.test.nightly_bench_utils import BenchmarkResult
 from sglang.test.nightly_utils import NightlyBenchmarkRunner
-from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    _parse_int_list_env,
+    popen_launch_server,
+)
 
 # Register for AMD CI - DeepSeek-V3.2 MTP benchmark (~90 min)
 register_amd_ci(
@@ -57,11 +62,58 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     return summary
 
 
+def _run_benchmark_with_timeout(
+    runner: NightlyBenchmarkRunner,
+    model_path: str,
+    batch_sizes: List[int],
+    input_lens: Tuple[int, ...],
+    output_lens: Tuple[int, ...],
+    other_args: List[str],
+    variant: str,
+    extra_bench_args: Optional[List[str]],
+    timeout: int,
+) -> Tuple[List[BenchmarkResult], bool, Optional[float]]:
+    """Run benchmark with a custom server launch timeout."""
+    model_description = f"{model_path}" + (f" ({variant})" if variant else "")
+    process = popen_launch_server(
+        model=model_path,
+        base_url=runner.base_url,
+        other_args=other_args,
+        timeout=timeout,
+    )
+    try:
+        profile_path_prefix, json_output_file = runner.generate_profile_filename(
+            model_path, variant
+        )
+        bench_args = list(extra_bench_args) if extra_bench_args else []
+        if variant:
+            bench_args.extend(["--run-name", variant])
+        command = runner.build_benchmark_command(
+            model_path,
+            batch_sizes,
+            input_lens,
+            output_lens,
+            profile_path_prefix,
+            json_output_file,
+            extra_args=bench_args,
+        )
+        _, cmd_success = runner.run_benchmark_command(command, model_description)
+        if not cmd_success:
+            return [], False, None
+        benchmark_results, load_success = runner.load_benchmark_results(
+            json_output_file, model_description
+        )
+        return benchmark_results, load_success, None
+    finally:
+        kill_process_tree(process.pid)
+
+
 # Model path can be overridden via environment variable
 DEEPSEEK_V32_MODEL_PATH = os.environ.get(
     "DEEPSEEK_V32_MODEL_PATH", "deepseek-ai/DeepSeek-V3.2"
 )
 PROFILE_DIR = "performance_profiles_deepseek_v32_mtp"
+SERVER_LAUNCH_TIMEOUT = 5400
 
 
 class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
@@ -102,6 +154,8 @@ class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
                 "0.7",
                 "--model-loader-extra-config",
                 '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
             ],
         }
 
@@ -113,7 +167,8 @@ class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
     def test_bench_one_batch(self):
         """Run benchmark for MTP variant."""
         try:
-            result_tuple = self.runner.run_benchmark_for_model(
+            result_tuple = _run_benchmark_with_timeout(
+                runner=self.runner,
                 model_path=self.model,
                 batch_sizes=self.batch_sizes,
                 input_lens=self.input_lens,
@@ -121,6 +176,7 @@ class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
                 other_args=self.variant_config["other_args"],
                 variant=self.variant_config["name"],
                 extra_bench_args=["--trust-remote-code"],
+                timeout=SERVER_LAUNCH_TIMEOUT,
             )
             results = result_tuple[0]
             success = result_tuple[1]

test/registered/amd/perf/test_deepseek_v32_basic_perf_amd.py

@@ -0,0 +1,142 @@
+"""AMD Nightly performance benchmark for DeepSeek-V3.2 model (basic variant).
+
+This test benchmarks the DeepSeek-V3.2 model with basic TP=8 configuration on 8 GPUs.
+
+The model path can be configured via DEEPSEEK_V32_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-deepseek-v32-basic suite
+
+Example usage:
+    DEEPSEEK_V32_MODEL_PATH=deepseek-ai/DeepSeek-V3.2 python -m pytest test_deepseek_v32_basic_perf_amd.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - DeepSeek-V3.2 basic benchmark (~90 min)
+register_amd_ci(
+    est_time=5400, suite="nightly-perf-8-gpu-deepseek-v32-basic", nightly=True
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI325")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    # Skip first result if it's a warmup (same batch_size as second result)
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path can be overridden via environment variable
+DEEPSEEK_V32_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V32_MODEL_PATH", "deepseek-ai/DeepSeek-V3.2"
+)
+PROFILE_DIR = "performance_profiles_deepseek_v32_basic_mi325"
+
+
+class TestNightlyDeepseekV32BasicPerformance(unittest.TestCase):
+    """AMD Nightly performance benchmark for DeepSeek-V3.2 model (basic variant).
+
+    Tests the DeepSeek-V3.2 model with basic TP=8 configuration on MI325/MI300X.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Basic variant configuration for DeepSeek-V3.2
+        # MI325 uses aiter attention backend
+        cls.variant_config = {
+            "name": "basic",
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--attention-backend",
+                "aiter",
+                "--chunked-prefill-size",
+                "131072",
+                "--mem-fraction-static",
+                "0.85",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
+            ],
+            "env_vars": {"SGLANG_USE_AITER": "1"},
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        # Override full_report to remove traces help text
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark for basic variant."""
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model,
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.variant_config["other_args"],
+                variant=self.variant_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+            avg_spec_accept_length = result_tuple[2] if len(result_tuple) > 2 else None
+
+            # Log speculative decoding accept length
+            if avg_spec_accept_length is not None:
+                print(f"  avg_spec_accept_length={avg_spec_accept_length:.2f}")
+
+            # Use simplified report format without traces
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            if not success:
+                raise AssertionError(
+                    f"Benchmark failed for {self.model} (basic variant)"
+                )
+        finally:
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/perf/test_deepseek_v32_mtp_perf_amd.py

@@ -0,0 +1,149 @@
+"""AMD Nightly performance benchmark for DeepSeek-V3.2 model (MTP variant).
+
+This test benchmarks the DeepSeek-V3.2 model with MTP (EAGLE speculative decoding)
+configuration on 8 GPUs.
+
+The model path can be configured via DEEPSEEK_V32_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-deepseek-v32-mtp suite
+
+Example usage:
+    DEEPSEEK_V32_MODEL_PATH=deepseek-ai/DeepSeek-V3.2 python -m pytest test_deepseek_v32_mtp_perf_amd.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - DeepSeek-V3.2 MTP benchmark (~120 min)
+register_amd_ci(
+    est_time=7200, suite="nightly-perf-8-gpu-deepseek-v32-mtp", nightly=True
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI325")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    # Skip first result if it's a warmup (same batch_size as second result)
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path can be overridden via environment variable
+DEEPSEEK_V32_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V32_MODEL_PATH", "deepseek-ai/DeepSeek-V3.2"
+)
+PROFILE_DIR = "performance_profiles_deepseek_v32_mtp_mi325"
+
+
+class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
+    """AMD Nightly performance benchmark for DeepSeek-V3.2 model (MTP variant).
+
+    Tests the DeepSeek-V3.2 model with MTP (EAGLE speculative decoding) on MI325/MI300X.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # MTP variant configuration for DeepSeek-V3.2
+        # MI325 uses aiter attention backend + EAGLE speculative decoding
+        cls.variant_config = {
+            "name": "mtp",
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--attention-backend",
+                "aiter",
+                "--chunked-prefill-size",
+                "131072",
+                "--speculative-algorithm",
+                "EAGLE",
+                "--speculative-num-steps",
+                "3",
+                "--speculative-eagle-topk",
+                "1",
+                "--speculative-num-draft-tokens",
+                "4",
+                "--mem-fraction-static",
+                "0.7",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+                "--watchdog-timeout",
+                "1200",
+            ],
+            "env_vars": {"SGLANG_USE_AITER": "1"},
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        # Override full_report to remove traces help text
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark for MTP variant."""
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model,
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.variant_config["other_args"],
+                variant=self.variant_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+            avg_spec_accept_length = result_tuple[2] if len(result_tuple) > 2 else None
+
+            # Log speculative decoding accept length
+            if avg_spec_accept_length is not None:
+                print(f"  avg_spec_accept_length={avg_spec_accept_length:.2f}")
+
+            # Use simplified report format without traces
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            if not success:
+                raise AssertionError(f"Benchmark failed for {self.model} (MTP variant)")
+        finally:
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()