[AMD] Add AMD CI registration (1-gpu unit test) to nightly CI. (#16941)

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

@@ -0,0 +1,250 @@
+"""MI35x DeepSeek-R1 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-R1-0528 with basic configuration using few-shot completion
+benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-deepseek-r1 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-R1-0528 accuracy test (~30 min for basic only)
+register_amd_ci(
+    est_time=1800,
+    suite="nightly-amd-accuracy-8-gpu-mi35x-deepseek-r1",
+    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-R1 models for MI35x - only basic variant for faster CI
+# MTP, DP, and TC variants removed to reduce test time from ~2h to ~30min
+MI35X_DEEPSEEK_R1_MODELS = [
+    # DeepSeek-R1-0528 basic
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-R1-0528",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=3600,
+        variant="basic",
+        other_args=[
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--disable-radix-cache",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+            "--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 TestDeepSeekR1EvalMI35x(unittest.TestCase):
+    """DeepSeek-R1 GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = MI35X_DEEPSEEK_R1_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_deepseek_r1_accuracy(self):
+        """Test DeepSeek-R1 models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### DeepSeek-R1 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/mi35x/test_deepseek_r1_mxfp4_eval_mi35x.py

@@ -0,0 +1,300 @@
+"""MI35x DeepSeek-R1-MXFP4 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-R1-MXFP4 quantized model with multiple configurations
+(basic, MTP, DP, TC) using few-shot completion benchmark on MI35x.
+
+Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4 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-R1-MXFP4 accuracy tests (~120 min)
+register_amd_ci(
+    est_time=7200, suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4", nightly=True
+)
+
+INVALID = -9999999
+
+# Model path configuration for MI35x DeepSeek-R1-MXFP4
+# Priority: 1) env var, 2) local path, 3) HuggingFace model ID
+DEEPSEEK_R1_MXFP4_LOCAL_PATH = "/data2/models/amd-DeepSeek-R1-MXFP4-Preview"
+DEEPSEEK_R1_MXFP4_HF_MODEL_ID = "amd/DeepSeek-R1-MXFP4-Preview"
+
+
+def get_model_path() -> str:
+    """Get effective model path: env var > local path > HF model ID."""
+    env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
+    if env_path:
+        return env_path
+    if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
+        return DEEPSEEK_R1_MXFP4_LOCAL_PATH
+    return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
+
+
+@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
+
+
+def get_mxfp4_models() -> List[ModelConfig]:
+    """Get DeepSeek-R1-MXFP4 model configurations for MI35x."""
+    model_path = get_model_path()
+    return [
+        # DeepSeek-R1-MXFP4 basic
+        ModelConfig(
+            model_path=model_path,
+            tp_size=8,
+            accuracy_threshold=0.93,
+            timeout=3600,
+            variant="basic",
+            other_args=[
+                "--attention-backend",
+                "aiter",
+                "--chunked-prefill-size",
+                "131072",
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                "0.85",
+                "--trust-remote-code",
+            ],
+            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
+    ]
+
+
+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 TestDeepSeekR1MXFP4EvalMI35x(unittest.TestCase):
+    """DeepSeek-R1-MXFP4 GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = get_mxfp4_models()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_deepseek_r1_mxfp4_accuracy(self):
+        """Test DeepSeek-R1-MXFP4 models with GSM8K completion benchmark."""
+        # Check if model exists
+        model_path = get_model_path()
+        is_local_path = model_path.startswith("/")
+        if is_local_path and not os.path.exists(model_path):
+            print(f"\n⏭️ SKIPPING: Local model not found at {model_path}")
+            self.skipTest(f"Local model not found at {model_path}")
+            return
+
+        if is_local_path:
+            print(f"📁 Using local model: {model_path}")
+        else:
+            print(f"📥 Using HuggingFace model: {model_path}")
+
+        all_results = []
+        summary = "### DeepSeek-R1-MXFP4 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/mi35x/test_gpt_oss_eval_mi35x.py

@@ -0,0 +1,254 @@
+"""MI35x GPT-OSS GSM8K Completion Evaluation Test (8-GPU)
+
+Tests GPT-OSS models (openai/gpt-oss-20b, openai/gpt-oss-120b) using
+few-shot completion benchmark on MI35x.
+
+Note: MI35x uses openai/* paths, not lmsys/* paths like MI300X.
+
+Registry: nightly-amd-8-gpu-mi35x 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 GPT-OSS accuracy tests (~30 min)
+register_amd_ci(est_time=1800, suite="nightly-amd-8-gpu-mi35x", 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
+
+    def __post_init__(self):
+        if self.other_args is None:
+            self.other_args = []
+        if self.env_vars is None:
+            self.env_vars = {}
+
+
+# GPT-OSS models for MI35x (different paths from MI300X)
+MI35X_GPT_OSS_MODELS = [
+    ModelConfig(
+        model_path="openai/gpt-oss-20b",
+        tp_size=8,
+        accuracy_threshold=0.47,
+        other_args=[
+            "--chunked-prefill-size",
+            "130172",
+            "--max-running-requests",
+            "128",
+            "--mem-fraction-static",
+            "0.85",
+            "--attention-backend",
+            "triton",
+            "--trust-remote-code",
+        ],
+        env_vars={"SGLANG_USE_AITER": "1"},
+    ),
+    ModelConfig(
+        model_path="openai/gpt-oss-120b",
+        tp_size=8,
+        accuracy_threshold=0.79,
+        timeout=900,
+        other_args=[
+            "--chunked-prefill-size",
+            "130172",
+            "--max-running-requests",
+            "128",
+            "--mem-fraction-static",
+            "0.85",
+            "--attention-backend",
+            "triton",
+            "--trust-remote-code",
+        ],
+        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 TestGptOssEvalMI35x(unittest.TestCase):
+    """GPT-OSS GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = MI35X_GPT_OSS_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_gpt_oss_accuracy(self):
+        """Test GPT-OSS models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### GPT-OSS Models (MI35x)\n\n"
+        summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            with self.subTest(model=config.model_path):
+                print(f"\n{'='*60}")
+                print(f"Testing: {config.model_path}")
+                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": config.model_path,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {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.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": config.model_path,
+                            "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/mi35x/test_grok1_int4_eval_mi35x.py

@@ -0,0 +1,162 @@
+"""MI35x GROK1-INT4 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests Grok-1 INT4 (W4A8KV8) model using few-shot completion benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-grok1-int4 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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 - GROK1-INT4 accuracy tests on MI35x (~25 min)
+register_amd_ci(
+    est_time=1500, suite="nightly-amd-accuracy-8-gpu-mi35x-grok1-int4", nightly=True
+)
+
+INVALID = -9999999
+
+GROK1_INT4_MODEL_PATH = os.environ.get("GROK1_INT4_MODEL_PATH", "amd/grok-1-W4A8KV8")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestGrok1INT4EvalMI35x(unittest.TestCase):
+    """GROK1-INT4 GSM8K Completion Evaluation Test for MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.80
+
+    def test_grok1_int4_accuracy(self):
+        """Test Grok-1 INT4 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["RCCL_MSCCL_ENABLE"] = "0"
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_INT4_WEIGHT"] = "1"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--tokenizer-path",
+            GROK1_TOKENIZER_PATH,
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=GROK1_INT4_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### GROK1-INT4 (MI35x)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {GROK1_INT4_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -0,0 +1,164 @@
+"""MI35x GROK2 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests Grok-2 model using few-shot completion benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-grok2 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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 - GROK2 accuracy tests on MI35x (~25 min)
+register_amd_ci(
+    est_time=1500, suite="nightly-amd-accuracy-8-gpu-mi35x-grok2", nightly=True
+)
+
+INVALID = -9999999
+
+GROK2_MODEL_PATH = os.environ.get("GROK2_MODEL_PATH", "xai-org/grok-2")
+GROK2_TOKENIZER_PATH = os.environ.get(
+    "GROK2_TOKENIZER_PATH", "alvarobartt/grok-2-tokenizer"
+)
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestGrok2EvalMI35x(unittest.TestCase):
+    """GROK2 GSM8K Completion Evaluation Test for MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.915
+
+    def test_grok2_accuracy(self):
+        """Test Grok-2 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["RCCL_MSCCL_ENABLE"] = "0"
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_INT4_WEIGHT"] = "0"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--tokenizer-path",
+            GROK2_TOKENIZER_PATH,
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=GROK2_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### GROK2 (MI35x)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {GROK2_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_deepseek_r1_eval_amd.py

@@ -0,0 +1,309 @@
+"""AMD DeepSeek-R1 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-R1-0528 with multiple configurations (basic, MTP, DP, TC)
+using few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-deepseek-r1 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-R1 accuracy tests (~120 min)
+register_amd_ci(
+    est_time=7200, suite="nightly-amd-accuracy-8-gpu-deepseek-r1", 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-R1 models for MI300X
+DEEPSEEK_R1_MODELS = [
+    # DeepSeek-R1-0528 basic
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-R1-0528",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=3600,
+        variant="basic",
+        other_args=[
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--disable-radix-cache",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ],
+        env_vars={"SGLANG_USE_AITER": "1"},
+    ),
+    # DeepSeek-R1-0528 with MTP (EAGLE)
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-R1-0528",
+        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"},
+    ),
+    # DeepSeek-R1-0528 with DP attention
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-R1-0528",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=3600,
+        variant="DP",
+        other_args=[
+            "--chunked-prefill-size",
+            "131072",
+            "--dp-size",
+            "8",
+            "--enable-dp-attention",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ],
+        env_vars={
+            "SGLANG_USE_ROCM700A": "1",
+            "SGLANG_USE_AITER": "1",
+        },
+    ),
+    # DeepSeek-R1-0528 with torch compile
+    ModelConfig(
+        model_path="deepseek-ai/DeepSeek-R1-0528",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=7200,
+        variant="TC",
+        other_args=[
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.70",
+            "--cuda-graph-max-bs",
+            "8",
+            "--enable-torch-compile",
+            "--disable-cuda-graph",
+            "--trust-remote-code",
+        ],
+        env_vars={
+            "SGLANG_USE_ROCM700A": "1",
+            "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 TestDeepSeekR1EvalAMD(unittest.TestCase):
+    """DeepSeek-R1 GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = DEEPSEEK_R1_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_deepseek_r1_accuracy(self):
+        """Test DeepSeek-R1 models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### DeepSeek-R1 Models (MI300X)\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_deepseek_v31_eval_amd.py

@@ -0,0 +1,159 @@
+"""AMD DeepSeek-V3.1 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests DeepSeek-V3.1 model using few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-deepseek-v31 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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-V3.1 accuracy tests (~60 min)
+register_amd_ci(
+    est_time=3600, suite="nightly-amd-accuracy-8-gpu-deepseek-v31", nightly=True
+)
+
+INVALID = -9999999
+
+DEEPSEEK_V31_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V31_MODEL_PATH", "deepseek-ai/DeepSeek-V3-0324"
+)
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestDeepSeekV31EvalAMD(unittest.TestCase):
+    """DeepSeek-V3.1 GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.90
+
+    def test_deepseek_v31_accuracy(self):
+        """Test DeepSeek-V3.1 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--attention-backend",
+            "aiter",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=DEEPSEEK_V31_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### DeepSeek-V3.1 (MI300X)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {DEEPSEEK_V31_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_gpt_oss_eval_amd.py

@@ -0,0 +1,247 @@
+"""AMD GPT-OSS GSM8K Completion Evaluation Test (8-GPU)
+
+Tests GPT-OSS models (lmsys/gpt-oss-20b-bf16, lmsys/gpt-oss-120b-bf16) using
+few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu 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 - GPT-OSS accuracy tests (~30 min)
+register_amd_ci(est_time=1800, suite="nightly-amd-accuracy-8-gpu-gpt-oss", 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
+
+    def __post_init__(self):
+        if self.other_args is None:
+            self.other_args = []
+        if self.env_vars is None:
+            self.env_vars = {}
+
+
+# GPT-OSS models for MI300X
+GPT_OSS_MODELS = [
+    ModelConfig(
+        model_path="lmsys/gpt-oss-20b-bf16",
+        tp_size=8,
+        accuracy_threshold=0.47,
+        other_args=[
+            "--chunked-prefill-size",
+            "130172",
+            "--max-running-requests",
+            "128",
+            "--mem-fraction-static",
+            "0.85",
+            "--attention-backend",
+            "triton",
+            "--trust-remote-code",
+        ],
+        env_vars={"SGLANG_USE_AITER": "0"},
+    ),
+    ModelConfig(
+        model_path="lmsys/gpt-oss-120b-bf16",
+        tp_size=8,
+        accuracy_threshold=0.79,
+        timeout=900,
+        other_args=[
+            "--chunked-prefill-size",
+            "130172",
+            "--max-running-requests",
+            "128",
+            "--mem-fraction-static",
+            "0.85",
+            "--attention-backend",
+            "triton",
+            "--trust-remote-code",
+        ],
+        env_vars={"SGLANG_USE_AITER": "0"},
+    ),
+]
+
+
+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 TestGptOssEvalAMD(unittest.TestCase):
+    """GPT-OSS GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = GPT_OSS_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_gpt_oss_accuracy(self):
+        """Test GPT-OSS models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### GPT-OSS Models (MI300X)\n\n"
+        summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            with self.subTest(model=config.model_path):
+                print(f"\n{'='*60}")
+                print(f"Testing: {config.model_path}")
+                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": config.model_path,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {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.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": config.model_path,
+                            "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_grok1_fp8_eval_amd.py

@@ -0,0 +1,162 @@
+"""AMD GROK1-FP8 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests Grok-1 FP8 model using few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-grok1-fp8 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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 - GROK1-FP8 accuracy tests (~25 min)
+register_amd_ci(
+    est_time=1500, suite="nightly-amd-accuracy-8-gpu-grok1-fp8", nightly=True
+)
+
+INVALID = -9999999
+
+GROK1_FP8_MODEL_PATH = os.environ.get("GROK1_MODEL_PATH", "lmzheng/grok-1")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestGrok1FP8EvalAMD(unittest.TestCase):
+    """GROK1-FP8 GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.80
+
+    def test_grok1_fp8_accuracy(self):
+        """Test Grok-1 FP8 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["RCCL_MSCCL_ENABLE"] = "0"
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_INT4_WEIGHT"] = "0"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--tokenizer-path",
+            GROK1_TOKENIZER_PATH,
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=GROK1_FP8_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### GROK1-FP8 (MI300X)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {GROK1_FP8_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_grok1_int4_eval_amd.py

@@ -0,0 +1,162 @@
+"""AMD GROK1-INT4 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests Grok-1 INT4 (W4A8KV8) model using few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-grok1-int4 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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 - GROK1-INT4 accuracy tests (~25 min)
+register_amd_ci(
+    est_time=1500, suite="nightly-amd-accuracy-8-gpu-grok1-int4", nightly=True
+)
+
+INVALID = -9999999
+
+GROK1_INT4_MODEL_PATH = os.environ.get("GROK1_INT4_MODEL_PATH", "amd/grok-1-W4A8KV8")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestGrok1INT4EvalAMD(unittest.TestCase):
+    """GROK1-INT4 GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.80
+
+    def test_grok1_int4_accuracy(self):
+        """Test Grok-1 INT4 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["RCCL_MSCCL_ENABLE"] = "0"
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_INT4_WEIGHT"] = "1"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--tokenizer-path",
+            GROK1_TOKENIZER_PATH,
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=GROK1_INT4_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### GROK1-INT4 (MI300X)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {GROK1_INT4_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_grok2_eval_amd.py

@@ -0,0 +1,162 @@
+"""AMD GROK2 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests Grok-2 model using few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-grok2 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+
+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_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 - GROK2 accuracy tests (~25 min)
+register_amd_ci(est_time=1500, suite="nightly-amd-accuracy-8-gpu-grok2", nightly=True)
+
+INVALID = -9999999
+
+GROK2_MODEL_PATH = os.environ.get("GROK2_MODEL_PATH", "xai-org/grok-2")
+GROK2_TOKENIZER_PATH = os.environ.get(
+    "GROK2_TOKENIZER_PATH", "alvarobartt/grok-2-tokenizer"
+)
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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, num_questions=200, num_shots=5, parallel=64):
+    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"]))
+    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))
+    return float(acc), float(latency)
+
+
+class TestGrok2EvalAMD(unittest.TestCase):
+    """GROK2 GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+        cls.accuracy_threshold = 0.915
+
+    def test_grok2_accuracy(self):
+        """Test Grok-2 with GSM8K completion benchmark."""
+        env = os.environ.copy()
+        env["RCCL_MSCCL_ENABLE"] = "0"
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_INT4_WEIGHT"] = "0"
+
+        other_args = [
+            "--tp",
+            "8",
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--tokenizer-path",
+            GROK2_TOKENIZER_PATH,
+            "--trust-remote-code",
+        ]
+
+        process = popen_launch_server(
+            model=GROK2_MODEL_PATH,
+            base_url=self.base_url,
+            timeout=3600,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            acc, latency = run_gsm8k_benchmark(
+                self.base_url, num_questions=self.num_questions
+            )
+            passed = acc >= self.accuracy_threshold
+            status = "✅ PASS" if passed else "❌ FAIL"
+
+            summary = f"### GROK2 (MI300X)\n\n"
+            summary += f"| Model | Accuracy | Threshold | Status |\n"
+            summary += f"| ----- | -------- | --------- | ------ |\n"
+            summary += f"| {GROK2_MODEL_PATH} | {acc:.3f} | {self.accuracy_threshold} | {status} |\n"
+
+            if is_in_ci():
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                self.accuracy_threshold,
+                f"Accuracy {acc:.3f} below threshold {self.accuracy_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/test_grok_eval_amd.py

@@ -0,0 +1,287 @@
+"""AMD GROK GSM8K Completion Evaluation Test (8-GPU)
+
+Tests GROK models (Grok-1 FP8, Grok-1 INT4, Grok-2) using
+few-shot completion benchmark on MI300X.
+
+Registry: nightly-amd-8-gpu-grok 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
+
+# DISABLED: Split into individual files for each model variant
+# See: test_grok1_fp8_eval_amd.py, test_grok1_int4_eval_amd.py, test_grok2_eval_amd.py
+register_amd_ci(
+    est_time=2700,
+    suite="nightly-amd-8-gpu-grok",
+    nightly=True,
+    disabled="Split into test_grok1_fp8_eval_amd.py, test_grok1_int4_eval_amd.py, test_grok2_eval_amd.py",
+)
+
+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
+    tokenizer_path: Optional[str] = None
+    timeout: Optional[int] = None
+
+    def __post_init__(self):
+        if self.other_args is None:
+            self.other_args = []
+        if self.env_vars is None:
+            self.env_vars = {}
+
+
+# GROK models for MI300X
+GROK_MODELS = [
+    # GROK1-FP8
+    ModelConfig(
+        model_path="lmzheng/grok-1",
+        tp_size=8,
+        accuracy_threshold=0.80,
+        timeout=3600,
+        tokenizer_path="Xenova/grok-1-tokenizer",
+        other_args=[
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ],
+        env_vars={
+            "RCCL_MSCCL_ENABLE": "0",
+            "SGLANG_USE_AITER": "1",
+            "SGLANG_INT4_WEIGHT": "0",
+        },
+    ),
+    # GROK1-INT4
+    ModelConfig(
+        model_path="amd/grok-1-W4A8KV8",
+        tp_size=8,
+        accuracy_threshold=0.80,
+        timeout=3600,
+        tokenizer_path="Xenova/grok-1-tokenizer",
+        other_args=[
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ],
+        env_vars={
+            "RCCL_MSCCL_ENABLE": "0",
+            "SGLANG_USE_AITER": "1",
+            "SGLANG_INT4_WEIGHT": "1",
+        },
+    ),
+    # GROK2
+    ModelConfig(
+        model_path="xai-org/grok-2",
+        tp_size=8,
+        accuracy_threshold=0.915,
+        timeout=3600,
+        tokenizer_path="alvarobartt/grok-2-tokenizer",
+        other_args=[
+            "--quantization",
+            "fp8",
+            "--attention-backend",
+            "aiter",
+            "--mem-fraction-static",
+            "0.85",
+            "--trust-remote-code",
+        ],
+        env_vars={
+            "RCCL_MSCCL_ENABLE": "0",
+            "SGLANG_USE_AITER": "1",
+            "SGLANG_INT4_WEIGHT": "0",
+        },
+    ),
+]
+
+
+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 TestGrokEvalAMD(unittest.TestCase):
+    """GROK GSM8K Completion Evaluation Test for AMD MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = GROK_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_grok_accuracy(self):
+        """Test GROK models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### GROK Models (MI300X)\n\n"
+        summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            with self.subTest(model=config.model_path):
+                print(f"\n{'='*60}")
+                print(f"Testing: {config.model_path}")
+                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)])
+                if config.tokenizer_path:
+                    other_args.extend(["--tokenizer-path", config.tokenizer_path])
+                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": config.model_path,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {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.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": config.model_path,
+                            "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

@@ -0,0 +1,324 @@
+"""
+AMD VLM MMMU Evaluation Test - MI30x Only
+
+This test evaluates Vision-Language Models (VLMs) on the MMMU benchmark on AMD GPUs.
+Models are selected based on compatibility with AMD/ROCm platform.
+
+VLMs tested here:
+- Qwen2-VL series (Qwen2-VL-7B, Qwen2.5-VL-7B)
+- InternVL2 series
+- MiniCPM-v series
+- deepseek-vl2-small
+
+Note: Some VLMs from the Nvidia test are excluded due to AMD compatibility issues.
+Note: This test runs only on MI30x runners (linux-mi325-gpu-2), not on MI35x.
+
+Registry: nightly-amd-vlm suite (2-GPU VLM tests)
+"""
+
+import os
+import time
+import unittest
+import warnings
+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.run_eval import run_eval
+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,
+    write_results_to_json,
+)
+
+# Register for AMD CI - VLM MMMU evaluation tests (~120 min)
+register_amd_ci(est_time=7200, suite="nightly-amd-accuracy-2-gpu-vlm", nightly=True)
+
+# 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
+    {
+        "model_path": "Qwen/Qwen2-VL-7B-Instruct",
+        "tp_size": 1,
+        "accuracy_threshold": 0.30,
+        "extra_args": ["--trust-remote-code"],
+    },
+    {
+        "model_path": "Qwen/Qwen2.5-VL-7B-Instruct",
+        "tp_size": 1,
+        "accuracy_threshold": 0.33,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # InternVL2 - smaller model, good for testing
+    {
+        "model_path": "OpenGVLab/InternVL2_5-2B",
+        "tp_size": 1,
+        "accuracy_threshold": 0.29,
+        "extra_args": ["--trust-remote-code"],
+    },
+    # MiniCPM-v - lightweight VLM
+    {
+        "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": "deepseek-ai/deepseek-vl2-small",
+        "tp_size": 1,
+        "accuracy_threshold": 0.31,
+        "extra_args": ["--trust-remote-code"],
+    },
+]
+
+# Models that need special handling on AMD
+TRITON_ATTENTION_MODELS = {
+    "deepseek-ai/deepseek-vl2-small",  # MoE model
+}
+
+# Models known to fail on AMD - exclude from testing
+AMD_FAILING_VLM_MODELS = {
+    # Add models here as they are discovered to fail
+}
+
+
+def get_active_models():
+    """Get list of models to test, excluding known failures."""
+    return [m for m in AMD_VLM_MODELS if m["model_path"] not in AMD_FAILING_VLM_MODELS]
+
+
+class TestNightlyVLMMmmuEvalAMD(unittest.TestCase):
+    """AMD VLM MMMU Evaluation Test.
+
+    Tests Vision-Language Models on MMMU benchmark using AMD GPUs.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = get_active_models()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+    def test_mmmu_vlm_models(self):
+        """Test all configured VLM models on MMMU benchmark."""
+        warnings.filterwarnings(
+            "ignore", category=ResourceWarning, message="unclosed.*socket"
+        )
+        is_first = True
+        all_results = []
+        total_test_start = time.time()
+
+        print(f"\n{'='*60}")
+        print("AMD VLM MMMU Evaluation Test")
+        print(f"{'='*60}")
+        print(f"Benchmark: MMMU (100 samples)")
+        print(f"Models to test: {len(self.models)}")
+        for m in self.models:
+            print(f"  - {m['model_path']} (TP={m['tp_size']})")
+        print(f"{'='*60}\n")
+
+        for model_config in self.models:
+            model_path = model_config["model_path"]
+            tp_size = model_config["tp_size"]
+            accuracy_threshold = model_config["accuracy_threshold"]
+            extra_args = model_config.get("extra_args", [])
+            error_message = None
+
+            with self.subTest(model=model_path):
+                print(f"\n{'='*60}")
+                print(f"Testing: {model_path} (TP={tp_size})")
+                print(f"{'='*60}")
+
+                model_start = time.time()
+                startup_time = None
+                eval_time = None
+                score = None
+
+                # Set AMD-specific environment variables
+                if model_path in TRITON_ATTENTION_MODELS:
+                    os.environ["SGLANG_USE_AITER"] = "0"
+                else:
+                    os.environ["SGLANG_USE_AITER"] = "1"
+
+                # Build launch args
+                other_args = list(extra_args)
+                other_args.extend(["--log-level-http", "warning"])
+                if tp_size > 1:
+                    other_args.extend(["--tp", str(tp_size)])
+
+                # Launch server with timing
+                print(f"🚀 Launching server...")
+                server_start = time.time()
+                process = popen_launch_server(
+                    model=model_path,
+                    base_url=self.base_url,
+                    other_args=other_args,
+                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                )
+                startup_time = time.time() - server_start
+                print(f"⏱️  Server startup: {startup_time:.1f}s")
+
+                try:
+                    args = SimpleNamespace(
+                        base_url=self.base_url,
+                        model=model_path,
+                        eval_name="mmmu",
+                        num_examples=100,
+                        num_threads=64,
+                        max_tokens=30,
+                    )
+
+                    # Run evaluation with timing
+                    print(f"📊 Running MMMU evaluation (100 samples)...")
+                    eval_start = time.time()
+
+                    # Retry up to 3 times
+                    metrics = None
+                    for attempt in range(3):
+                        try:
+                            metrics = run_eval(args)
+                            score = metrics["score"]
+                            if score >= accuracy_threshold:
+                                break
+                        except Exception as e:
+                            print(f"   Attempt {attempt + 1} failed with error: {e}")
+                            if attempt == 2:
+                                raise
+
+                    eval_time = time.time() - eval_start
+                    total_time = time.time() - model_start
+
+                    # Print results
+                    print(f"\n📈 Results for {model_path}:")
+                    print(
+                        f"   Score: {score:.3f} (threshold: {accuracy_threshold:.2f})"
+                    )
+                    print(f"\n⏱️  Runtime breakdown:")
+                    print(f"   Server startup: {startup_time:.1f}s")
+                    print(f"   Evaluation: {eval_time:.1f}s")
+                    print(f"   Total: {total_time:.1f}s")
+
+                    passed = score >= accuracy_threshold
+                    if passed:
+                        print(f"\n   Status: ✅ PASSED")
+                    else:
+                        print(f"\n   Status: ❌ FAILED")
+
+                    write_results_to_json(model_path, metrics, "w" if is_first else "a")
+                    is_first = False
+
+                    all_results.append(
+                        {
+                            "model": model_path,
+                            "tp_size": tp_size,
+                            "score": score,
+                            "threshold": accuracy_threshold,
+                            "startup_time": startup_time,
+                            "eval_time": eval_time,
+                            "total_time": total_time,
+                            "passed": passed,
+                            "error": None,
+                        }
+                    )
+
+                except Exception as e:
+                    error_message = str(e)
+                    total_time = time.time() - model_start
+                    print(f"\n❌ Error evaluating {model_path}: {error_message}")
+                    all_results.append(
+                        {
+                            "model": model_path,
+                            "tp_size": tp_size,
+                            "score": None,
+                            "threshold": accuracy_threshold,
+                            "startup_time": startup_time,
+                            "eval_time": None,
+                            "total_time": total_time,
+                            "passed": False,
+                            "error": error_message,
+                        }
+                    )
+
+                finally:
+                    print(f"\n🛑 Stopping server...")
+                    kill_process_tree(process.pid)
+
+        # Calculate total test runtime
+        total_test_time = time.time() - total_test_start
+
+        # Generate summary
+        self._check_results(all_results, total_test_time)
+
+    def _check_results(self, results, total_test_time):
+        """Check results and generate summary."""
+        failed_models = []
+        passed_count = 0
+        failed_count = 0
+
+        summary = (
+            "| Model | TP | Score | Threshold | Startup | Eval | Total | Status |\n"
+        )
+        summary += (
+            "| ----- | -- | ----- | --------- | ------- | ---- | ----- | ------ |\n"
+        )
+
+        for result in results:
+            model = result["model"]
+            score = result["score"]
+            tp_size = result["tp_size"]
+            threshold = result["threshold"]
+            startup_time = result.get("startup_time")
+            eval_time = result.get("eval_time")
+            total_time = result.get("total_time")
+            error = result.get("error")
+
+            if error:
+                status = "❌ ERROR"
+                failed_count += 1
+                failed_models.append(f"- {model}: ERROR - {error[:100]}")
+            elif result["passed"]:
+                status = "✅ PASS"
+                passed_count += 1
+            else:
+                status = "❌ FAIL"
+                failed_count += 1
+                failed_models.append(
+                    f"- {model}: score={score:.4f}, threshold={threshold:.4f}"
+                )
+
+            # Format values
+            score_str = f"{score:.3f}" if score is not None else "N/A"
+            startup_str = f"{startup_time:.0f}s" if startup_time is not None else "N/A"
+            eval_str = f"{eval_time:.0f}s" if eval_time is not None else "N/A"
+            total_str = f"{total_time:.0f}s" if total_time is not None else "N/A"
+
+            summary += f"| {model} | {tp_size} | {score_str} | {threshold:.2f} | {startup_str} | {eval_str} | {total_str} | {status} |\n"
+
+        print(f"\n{'='*60}")
+        print("SUMMARY - AMD VLM MMMU Evaluation")
+        print(f"{'='*60}")
+        print(summary)
+        print(f"\n📊 Final Statistics:")
+        print(f"   Passed: {passed_count}")
+        print(f"   Failed: {failed_count}")
+        print(
+            f"\n⏱️  Total test runtime: {total_test_time:.1f}s ({total_test_time/60:.1f} min)"
+        )
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### TestNightlyVLMMmmuEvalAMD\n{summary}\n\n"
+                f"**Total Runtime:** {total_test_time:.1f}s ({total_test_time/60:.1f} min)"
+            )
+
+        if failed_models:
+            failure_msg = "\n".join(failed_models)
+            raise AssertionError(f"The following models failed:\n{failure_msg}")
+
+
+if __name__ == "__main__":
+    unittest.main()