[AMD] Add DeepSeek-V3.2 and VLMs model in nightly tests (#17179)

Co-authored-by: michaelzhang-ai <michaelzhang-ai@users.noreply.github.com>
Co-authored-by: YC Tseng <yctseng@amd.com>
Co-authored-by: Bingxu Chen <bingxche@amd.com>

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

@@ -22,6 +22,8 @@ on:
           # MI30x Accuracy Tests (GSM8K / MMMU)
           - 'nightly-accuracy-2-gpu'
           - 'nightly-accuracy-2-gpu-vlm'
+          - '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)
@@ -31,10 +33,12 @@ on:
           # MI35x jobs
           - 'nightly-test-1-gpu-mi35x'
           - 'nightly-accuracy-8-gpu-mi35x'
-          - 'nightly-accuracy-8-gpu-mi35x-deepseek-r1'
           - 'nightly-8-gpu-mi35x-grok1-int4'
           - '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-perf-8-gpu-mi35x-deepseek-v32-mtp'
   workflow_call:
     inputs:
       ref:
@@ -106,6 +110,7 @@ jobs:
 
       - name: Nightly Test (2-GPU)
         run: |
+          > github_summary.md  # Clear summary file
           bash scripts/ci/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 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
@@ -135,12 +140,75 @@ jobs:
       - name: Nightly Accuracy Test (2-GPU VLM MMMU)
         timeout-minutes: 180
         run: |
+          > github_summary.md  # Clear summary file
           bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
             -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
             python3 run_suite.py --hw amd --suite nightly-amd-accuracy-2-gpu-vlm --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # 2-GPU Text Models Performance Tests
+  nightly-perf-2-gpu-text:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-perf-2-gpu-text')
+    runs-on: linux-mi325-gpu-2
+    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_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd_ci_install_dependency.sh
+
+      - name: Performance Test (2-GPU Text Models)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-perf-text-2-gpu --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  # 2-GPU VLM Performance Tests
+  nightly-perf-2-gpu-vlm:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-perf-2-gpu-vlm')
+    runs-on: linux-mi325-gpu-2
+    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_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd_ci_install_dependency.sh
+
+      - name: Performance Test (2-GPU VLM Models)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_USE_AITER=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-perf-vlm-2-gpu --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # 8-GPU Accuracy Tests - GPT-OSS, Grok1-FP8 (accuracy only)
   nightly-accuracy-8-gpu:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-accuracy-8-gpu')
@@ -404,38 +472,6 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
-  # MI35x 8-GPU DeepSeek-R1-0528 Accuracy Test (separate job due to long loading time)
-  nightly-accuracy-8-gpu-mi35x-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-mi35x-deepseek-r1')
-    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_ci_start_container.sh
-        env:
-          GITHUB_WORKSPACE: ${{ github.workspace }}
-
-      - name: Install dependencies
-        run: |
-          bash scripts/ci/amd_ci_install_dependency.sh
-          # Install tabulate for run_suite.py (missing in MI35x container)
-          bash scripts/ci/amd_ci_exec.sh pip install tabulate
-
-      - name: Accuracy Test MI35x (8-GPU DeepSeek-R1-0528)
-        timeout-minutes: 240
-        run: |
-          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
-            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-mi35x-deepseek-r1 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
-          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
-          exit ${TEST_EXIT_CODE:-0}
-
   # MI35x 8-GPU Grok1-INT4 (Accuracy + Performance combined)
   nightly-8-gpu-mi35x-grok1-int4:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-mi35x-grok1-int4')
@@ -572,6 +608,105 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # MI35x 8-GPU DeepSeek-V3.2 Accuracy Test
+  nightly-accuracy-8-gpu-mi35x-deepseek-v32:
+    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')
+    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_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd_ci_install_dependency.sh
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x (8-GPU DeepSeek-V3.2)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v32 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          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')
+    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_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd_ci_install_dependency.sh
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd_ci_exec.sh pip install tabulate
+
+      - name: Performance Test MI35x (8-GPU DeepSeek-V3.2 Basic)
+        timeout-minutes: 150
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-basic --nightly --timeout-per-file 5400 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  # MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP)
+  nightly-perf-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-perf-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_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd_ci_install_dependency.sh
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd_ci_exec.sh pip install tabulate
+
+      - name: Performance Test MI35x (8-GPU DeepSeek-V3.2 MTP)
+        timeout-minutes: 150
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 5400 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   check-all-jobs:
     if: always() && (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request' || github.event_name == 'workflow_dispatch')
     needs:
@@ -580,6 +715,9 @@ jobs:
       # MI30x Accuracy Tests
       - nightly-accuracy-2-gpu
       - nightly-accuracy-2-gpu-vlm
+      # MI30x Performance Tests
+      - 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
@@ -589,10 +727,12 @@ jobs:
       # MI35x jobs
       - nightly-test-1-gpu-mi35x
       - nightly-accuracy-8-gpu-mi35x
-      - nightly-accuracy-8-gpu-mi35x-deepseek-r1
       - nightly-8-gpu-mi35x-grok1-int4
       - 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-perf-8-gpu-mi35x-deepseek-v32-mtp
     runs-on: ubuntu-latest
     steps:
       - name: Check if any job failed

scripts/ci/amd_ci_exec.sh

@@ -53,8 +53,11 @@ for key in "${!ENV_MAP[@]}"; do
   ENV_ARGS+=("-e" "$key=${ENV_MAP[$key]}")
 done
 
-# Run docker exec with retry logic for HF network issues
+# Run docker exec with retry logic for HuggingFace network/download issues
-# First attempt: normal mode
+# When HF model downloads fail due to network timeouts or rate limits,
+# retrying with HF_HUB_OFFLINE=1 uses cached models from previous downloads.
+#
+# First attempt: normal mode (allows HF downloads)
 if docker exec \
   -w "$WORKDIR" \
   "${ENV_ARGS[@]}" \
@@ -65,7 +68,18 @@ else
 fi
 
 echo "First attempt failed with exit code $FIRST_EXIT_CODE"
-echo "Retrying with HF_HUB_OFFLINE=1 (offline mode)..."
+
+# Skip retry for test failures that won't be fixed by offline mode:
+#   - Exit 1: Test assertion failures (accuracy below threshold)
+#   - Exit 137 (128+9): Process killed by OOM
+#   - Exit 255: Test suite completed with test errors
+# Only retry for other exit codes (e.g., network timeouts, HF download failures)
+if [[ "$FIRST_EXIT_CODE" -eq 1 || "$FIRST_EXIT_CODE" -eq 137 || "$FIRST_EXIT_CODE" -eq 255 ]]; then
+  echo "Exit code $FIRST_EXIT_CODE indicates test failure (not network issue), not retrying"
+  exit $FIRST_EXIT_CODE
+fi
+
+echo "Retrying with HF_HUB_OFFLINE=1 (offline mode to use cached models)..."
 
 # Second attempt: force HF offline mode to avoid network timeouts
 docker exec \

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

@@ -1,7 +1,7 @@
 """MI35x DeepSeek-R1-MXFP4 GSM8K Completion Evaluation Test (8-GPU)
 
-Tests DeepSeek-R1-MXFP4 quantized model with multiple configurations
+Tests DeepSeek-R1-MXFP4 quantized model with basic configuration
-(basic, MTP, DP, TC) using few-shot completion benchmark on MI35x.
+using few-shot completion benchmark on MI35x.
 
 Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4 suite
 """
@@ -32,9 +32,9 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import download_and_cache_file, read_jsonl
 
-# Register for AMD CI - MI35x DeepSeek-R1-MXFP4 accuracy tests (~120 min)
+# Register for AMD CI - MI35x DeepSeek-R1-MXFP4 accuracy test (~60 min, basic only)
 register_amd_ci(
-    est_time=7200, suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4", nightly=True
+    est_time=3600, suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4", nightly=True
 )
 
 INVALID = -9999999
@@ -83,7 +83,7 @@ def get_mxfp4_models() -> List[ModelConfig]:
     """Get DeepSeek-R1-MXFP4 model configurations for MI35x."""
     model_path = get_model_path()
     return [
-        # DeepSeek-R1-MXFP4 basic
+        # DeepSeek-R1-MXFP4 basic only (MTP tested in perf job)
         ModelConfig(
             model_path=model_path,
             tp_size=8,
@@ -102,31 +102,6 @@ def get_mxfp4_models() -> List[ModelConfig]:
             ],
             env_vars={"SGLANG_USE_AITER": "1"},
         ),
-        # DeepSeek-R1-MXFP4 with MTP (EAGLE)
-        ModelConfig(
-            model_path=model_path,
-            tp_size=8,
-            accuracy_threshold=0.93,
-            timeout=3600,
-            variant="MTP",
-            other_args=[
-                "--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",
-                "--trust-remote-code",
-            ],
-            env_vars={"SGLANG_USE_AITER": "1"},
-        ),
-        # Note: DP and TC variants are not supported for MXFP4 on MI35x
     ]
 
 

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

@@ -0,0 +1,251 @@
+"""MI35x DeepSeek-V3.2 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.2 with basic configuration using few-shot completion
+benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32 suite
+"""
+
+import ast
+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 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 - MI35x DeepSeek-V3.2 accuracy test (~60 min for basic only)
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-8-gpu-mi35x-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 MI35x - only basic variant for nightly
+# DP variant removed due to barrier deadlock during model loading
+MI35X_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",
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
+            "--mem-fraction-static",
+            "0.85",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",  # 20 minutes for weight loading
+        ],
+        env_vars={},
+    ),
+]
+
+
+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 TestDeepSeekV32EvalMI35x(unittest.TestCase):
+    """DeepSeek-V3.2 GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = MI35X_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 (MI35x)\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"
+
+                        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_vlms_mmmu_eval_amd.py

@@ -5,15 +5,18 @@ This test evaluates Vision-Language Models (VLMs) on the MMMU benchmark on AMD G
 Models are selected based on compatibility with AMD/ROCm platform.
 
 VLMs tested here:
-- Qwen2-VL series (Qwen2-VL-7B, Qwen2.5-VL-7B)
+- Qwen VL series (Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-30B)
-- InternVL2 series
+- InternVL2 series (InternVL2_5-2B)
-- MiniCPM-v series
+- MiniCPM series (MiniCPM-v-2_6, MiniCPM-o-2_6)
-- deepseek-vl2-small
+- DeepSeek VL series (deepseek-vl2-small, Janus-Pro-7B)
+- Kimi VL (Kimi-VL-A3B-Instruct)
+- MiMo VL (MiMo-VL-7B-RL)
+- GLM VL (GLM-4.1V-9B-Thinking)
 
-Note: Some VLMs from the Nvidia test are excluded due to AMD compatibility issues.
+Note: NVILA models are excluded (NVIDIA-specific).
 Note: This test runs only on MI30x runners (linux-mi325-gpu-2), not on MI35x.
 
-Registry: nightly-amd-vlm suite (2-GPU VLM tests)
+Registry: nightly-amd-accuracy-2-gpu-vlm suite (2-GPU VLM tests)
 """
 
 import os
@@ -40,7 +43,7 @@ register_amd_ci(est_time=7200, suite="nightly-amd-accuracy-2-gpu-vlm", nightly=T
 # AMD-verified VLM models with conservative thresholds on 100 MMMU samples
 # Format: (model_path, tp_size, accuracy_threshold, extra_args)
 AMD_VLM_MODELS = [
-    # Qwen2-VL series - well supported on AMD
+    # Qwen VL series - well supported on AMD
     {
         "model_path": "Qwen/Qwen2-VL-7B-Instruct",
         "tp_size": 1,
@@ -53,6 +56,12 @@ AMD_VLM_MODELS = [
         "accuracy_threshold": 0.33,
         "extra_args": ["--trust-remote-code"],
     },
+    {
+        "model_path": "Qwen/Qwen3-VL-30B-A3B-Instruct",
+        "tp_size": 2,
+        "accuracy_threshold": 0.29,
+        "extra_args": ["--trust-remote-code"],
+    },
     # InternVL2 - smaller model, good for testing
     {
         "model_path": "OpenGVLab/InternVL2_5-2B",
@@ -60,25 +69,60 @@ AMD_VLM_MODELS = [
         "accuracy_threshold": 0.29,
         "extra_args": ["--trust-remote-code"],
     },
-    # MiniCPM-v - lightweight VLM
+    # MiniCPM series
     {
         "model_path": "openbmb/MiniCPM-v-2_6",
         "tp_size": 1,
         "accuracy_threshold": 0.25,
         "extra_args": ["--trust-remote-code"],
     },
-    # DeepSeek VL2 small - MoE VLM
+    {
+        "model_path": "openbmb/MiniCPM-o-2_6",
+        "tp_size": 1,
+        "accuracy_threshold": 0.32,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # DeepSeek VL series
     {
         "model_path": "deepseek-ai/deepseek-vl2-small",
         "tp_size": 1,
         "accuracy_threshold": 0.31,
         "extra_args": ["--trust-remote-code"],
     },
+    {
+        "model_path": "deepseek-ai/Janus-Pro-7B",
+        "tp_size": 1,
+        "accuracy_threshold": 0.28,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # Kimi VL - MoE
+    {
+        "model_path": "moonshotai/Kimi-VL-A3B-Instruct",
+        "tp_size": 1,
+        "accuracy_threshold": 0.26,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # MiMo VL
+    {
+        "model_path": "XiaomiMiMo/MiMo-VL-7B-RL",
+        "tp_size": 1,
+        "accuracy_threshold": 0.27,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # GLM VL
+    {
+        "model_path": "zai-org/GLM-4.1V-9B-Thinking",
+        "tp_size": 1,
+        "accuracy_threshold": 0.27,
+        "extra_args": ["--trust-remote-code"],
+    },
 ]
 
-# Models that need special handling on AMD
+# Models that need special handling on AMD (MoE models)
 TRITON_ATTENTION_MODELS = {
-    "deepseek-ai/deepseek-vl2-small",  # MoE model
+    "deepseek-ai/deepseek-vl2-small",
+    "Qwen/Qwen3-VL-30B-A3B-Instruct",
+    "moonshotai/Kimi-VL-A3B-Instruct",
 }
 
 # Models known to fail on AMD - exclude from testing

test/registered/amd/perf/mi35x/test_deepseek_r1_mxfp4_perf_mi35x.py

@@ -30,7 +30,10 @@ register_amd_ci(
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -43,7 +46,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -82,7 +92,7 @@ class TestDeepseekR1MXFP4PerfMI35x(unittest.TestCase):
         cls.model = get_model_path()
         print(f"Using model path: {cls.model}")
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        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"))
 

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

@@ -0,0 +1,134 @@
+"""MI35x 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-mi35x-deepseek-v32-basic suite
+
+Example usage:
+    DEEPSEEK_V32_MODEL_PATH=deepseek-ai/DeepSeek-V3.2 python -m pytest test_deepseek_v32_basic_perf_mi35x.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-mi35x-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", "MI35x")
+    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"
+
+
+class TestNightlyDeepseekV32BasicPerformance(unittest.TestCase):
+    """MI35x Nightly performance benchmark for DeepSeek-V3.2 model (basic variant).
+
+    Tests the DeepSeek-V3.2 model with basic TP=8 configuration.
+    """
+
+    @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
+        # MI35x uses tilelang NSA backends
+        cls.variant_config = {
+            "name": "basic",
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--nsa-prefill-backend",
+                "tilelang",
+                "--nsa-decode-backend",
+                "tilelang",
+                "--mem-fraction-static",
+                "0.85",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+            ],
+        }
+
+        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]
+
+            # 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/mi35x/test_deepseek_v32_mtp_perf_mi35x.py

@@ -0,0 +1,146 @@
+"""MI35x 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-mi35x-deepseek-v32-mtp suite
+
+Example usage:
+    DEEPSEEK_V32_MODEL_PATH=deepseek-ai/DeepSeek-V3.2 python -m pytest test_deepseek_v32_mtp_perf_mi35x.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 (~90 min)
+register_amd_ci(
+    est_time=5400, suite="nightly-perf-8-gpu-mi35x-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", "MI35x")
+    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"
+
+
+class TestNightlyDeepseekV32MTPPerformance(unittest.TestCase):
+    """MI35x Nightly performance benchmark for DeepSeek-V3.2 model (MTP variant).
+
+    Tests the DeepSeek-V3.2 model with MTP (EAGLE speculative decoding) on TP=8.
+    """
+
+    @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
+        # MI35x uses tilelang NSA backends + EAGLE speculative decoding
+        cls.variant_config = {
+            "name": "mtp",
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--nsa-prefill-backend",
+                "tilelang",
+                "--nsa-decode-backend",
+                "tilelang",
+                "--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}',
+            ],
+        }
+
+        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()

test/registered/amd/perf/mi35x/test_grok1_int4_perf_mi35x.py

@@ -21,7 +21,10 @@ register_amd_ci(
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -34,7 +37,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -53,7 +63,7 @@ class TestGrok1INT4PerfMI35x(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.batch_sizes = [1, 8, 16, 64]
         cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
         cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
 

test/registered/amd/perf/mi35x/test_grok2_perf_mi35x.py

@@ -19,7 +19,10 @@ register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-mi35x-grok2", nightly=T
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -32,7 +35,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -53,7 +63,7 @@ class TestGrok2PerfMI35x(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.batch_sizes = [1, 8, 16, 64]
         cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
         cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
 

test/registered/amd/perf/test_deepseek_v31_perf.py

@@ -22,7 +22,10 @@ register_amd_ci(est_time=18000, suite="nightly-perf-8-gpu-deepseek-v31", nightly
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -35,7 +38,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -59,7 +69,7 @@ class TestNightlyDeepseekV31Performance(unittest.TestCase):
     def setUpClass(cls):
         cls.model = DEEPSEEK_V31_MODEL_PATH
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        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"))
 

test/registered/amd/perf/test_grok1_int4_perf.py

@@ -24,7 +24,10 @@ register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-grok1-int4", nightly=Tr
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -37,7 +40,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -60,7 +70,7 @@ class TestNightlyGrok1INT4Performance(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.batch_sizes = [1, 8, 16, 64]
         cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
         cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
 

test/registered/amd/perf/test_grok2_perf.py

@@ -24,7 +24,10 @@ register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-grok2", nightly=True)
 
 
 def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
-    """Generate a simplified markdown report without traces and cost columns."""
+    """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})"
@@ -37,7 +40,14 @@ def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
     summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
     summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
 
-    for result in results:
+    # 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"
 
@@ -62,7 +72,7 @@ class TestNightlyGrok2Performance(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.batch_sizes = [1, 8, 16, 64]
         cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
         cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
 

test/registered/amd/perf/test_text_models_perf_amd.py

@@ -0,0 +1,132 @@
+"""AMD Nightly performance benchmark for text models (2-GPU).
+
+This test benchmarks text models on AMD MI30x/MI35x with 2 GPUs.
+
+Registry: nightly-amd-perf-text-2-gpu suite
+
+Example usage:
+    python -m pytest test_text_models_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,
+    ModelLaunchSettings,
+    _parse_int_list_env,
+    parse_models,
+)
+
+# Register for AMD CI - Text models benchmark (~60 min)
+register_amd_ci(est_time=3600, suite="nightly-amd-perf-text-2-gpu", nightly=True)
+
+PROFILE_DIR = "performance_profiles_text_models_amd"
+
+
+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", "AMD")
+    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
+
+
+class TestNightlyTextModelsPerfAMD(unittest.TestCase):
+    """AMD Nightly performance benchmark for text models (2-GPU)."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = []
+        # Llama-3.1-8B on TP=1
+        for model_path in parse_models("meta-llama/Llama-3.1-8B-Instruct"):
+            cls.models.append(
+                ModelLaunchSettings(
+                    model_path,
+                    tp_size=1,
+                    extra_args=["--attention-backend", "aiter"],
+                )
+            )
+        # Qwen2-57B MoE on TP=2
+        for model_path in parse_models("Qwen/Qwen2-57B-A14B-Instruct"):
+            cls.models.append(
+                ModelLaunchSettings(
+                    model_path,
+                    tp_size=2,
+                    extra_args=["--attention-backend", "aiter"],
+                )
+            )
+
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        # First batch_size=1 is warmup (standalone job, no accuracy test to warm up)
+        cls.batch_sizes = [1, 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"))
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark for all configured text models."""
+        all_model_succeed = True
+
+        try:
+            for model_setup in self.models:
+                with self.subTest(model=model_setup.model_path):
+                    other_args = list(model_setup.extra_args or [])
+                    if model_setup.tp_size and model_setup.tp_size > 1:
+                        other_args.extend(["--tp", str(model_setup.tp_size)])
+
+                    result_tuple = self.runner.run_benchmark_for_model(
+                        model_path=model_setup.model_path,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=other_args,
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        all_model_succeed = False
+
+                    if results:
+                        self.runner.full_report += (
+                            generate_simple_markdown_report(results) + "\n"
+                        )
+        finally:
+            self.runner.write_final_report()
+
+        if not all_model_succeed:
+            raise AssertionError("Some models failed the perf tests.")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/perf/test_vlms_perf_amd.py

@@ -0,0 +1,145 @@
+"""AMD Nightly performance benchmark for VLM models (2-GPU).
+
+This test benchmarks Vision-Language Models on AMD MI30x/MI35x with 2 GPUs.
+
+Registry: nightly-amd-perf-vlm-2-gpu suite
+
+Example usage:
+    python -m pytest test_vlms_perf_amd.py -v
+"""
+
+import os
+import unittest
+import warnings
+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,
+    ModelLaunchSettings,
+    _parse_int_list_env,
+    parse_models,
+)
+
+# Register for AMD CI - VLM models benchmark (~120 min)
+register_amd_ci(est_time=7200, suite="nightly-amd-perf-vlm-2-gpu", nightly=True)
+
+PROFILE_DIR = "performance_profiles_vlms_amd"
+
+# VLM models suitable for AMD
+MODEL_DEFAULTS = [
+    ModelLaunchSettings(
+        "Qwen/Qwen2.5-VL-7B-Instruct",
+        extra_args=["--mem-fraction-static=0.7"],
+    ),
+    ModelLaunchSettings(
+        "Qwen/Qwen3-VL-30B-A3B-Instruct",
+        tp_size=2,
+    ),
+]
+
+
+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", "AMD")
+    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
+
+
+class TestNightlyVLMsPerfAMD(unittest.TestCase):
+    """AMD Nightly performance benchmark for VLM models (2-GPU)."""
+
+    @classmethod
+    def setUpClass(cls):
+        warnings.filterwarnings(
+            "ignore", category=ResourceWarning, message="unclosed.*socket"
+        )
+
+        nightly_vlm_models_str = os.environ.get("NIGHTLY_VLM_MODELS")
+        if nightly_vlm_models_str:
+            cls.models = []
+            model_paths = parse_models(nightly_vlm_models_str)
+            for model_path in model_paths:
+                cls.models.append(ModelLaunchSettings(model_path))
+        else:
+            cls.models = MODEL_DEFAULTS
+
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        # First batch_size=1 is warmup (standalone job, no accuracy test to warm up)
+        cls.batch_sizes = _parse_int_list_env("NIGHTLY_VLM_BATCH_SIZES", "1,1,2,8,16")
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_OUTPUT_LENS", "512"))
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark for all configured VLM models."""
+        all_model_succeed = True
+
+        try:
+            for model_setup in self.models:
+                with self.subTest(model=model_setup.model_path):
+                    other_args = list(model_setup.extra_args or [])
+                    if model_setup.tp_size and model_setup.tp_size > 1:
+                        other_args.extend(["--tp", str(model_setup.tp_size)])
+
+                    # VLMs need additional benchmark args for dataset and trust-remote-code
+                    extra_bench_args = [
+                        "--trust-remote-code",
+                        "--dataset-name=mmmu",
+                    ]
+
+                    result_tuple = self.runner.run_benchmark_for_model(
+                        model_path=model_setup.model_path,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=other_args,
+                        extra_bench_args=extra_bench_args,
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        all_model_succeed = False
+
+                    if results:
+                        self.runner.full_report += (
+                            generate_simple_markdown_report(results) + "\n"
+                        )
+        finally:
+            self.runner.write_final_report()
+
+        if not all_model_succeed:
+            raise AssertionError("Some models failed the perf tests.")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/test_deepseek_v32_basic.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
-DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2-Exp"
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 
 class TestDeepseekV32DP(CustomTestCase):

test/registered/amd/test_deepseek_v32_mtp.py

@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
 )
 
 register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
-FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2-Exp"
+FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 
 class TestDeepseekV32DPMTP(CustomTestCase):