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

2026-01-13 23:41:45 -08:00
parent 030496eb06
commit b025cff441
39 changed files with 3701 additions and 2180 deletions
@@ -0,0 +1,168 @@
+"""MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 model.
+
+This test benchmarks the DeepSeek-R1-MXFP4 quantized model on MI35x with 8 GPUs.
+
+The model path can be configured via DEEPSEEK_R1_MXFP4_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4 suite
+
+Example usage:
+    DEEPSEEK_R1_MXFP4_MODEL_PATH=/data2/models/amd-DeepSeek-R1-MXFP4-Preview python -m pytest test_deepseek_r1_mxfp4_perf_mi35x.py -v
+"""
+
+import os
+
+# Set HF cache to /data2/models/ for MI35x so HF models download there
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - DeepSeek-R1-MXFP4 benchmark on MI35x (~300 min)
+register_amd_ci(
+    est_time=18000, suite="nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4", nightly=True
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path 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"
+PROFILE_DIR = "performance_profiles_deepseek_r1_mxfp4_mi35x"
+
+
+def get_model_path() -> str:
+    """Get effective model path: env var > local path > HF model ID."""
+    # Check env var first
+    env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
+    if env_path:
+        return env_path
+    # Check local path
+    if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
+        return DEEPSEEK_R1_MXFP4_LOCAL_PATH
+    # Fall back to HF model ID
+    return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
+
+
+class TestDeepseekR1MXFP4PerfMI35x(unittest.TestCase):
+    """MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 model.
+
+    Tests the DeepSeek-R1-MXFP4 quantized model on TP=8 with DP=8.
+    Uses local path if available, otherwise downloads from HuggingFace.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = get_model_path()
+        print(f"Using model path: {cls.model}")
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Define variant configurations for DeepSeek-R1-MXFP4 on MI35x
+        # Only run basic variant for perf (DP/TC/MTP covered in accuracy tests)
+        cls.variants = [
+            {
+                "name": "basic",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--chunked-prefill-size",
+                    "131072",
+                    "--disable-radix-cache",
+                    "--mem-fraction-static",
+                    "0.85",
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        # Override full_report to remove traces help text
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark across all configured variants."""
+        failed_variants = []
+
+        # For local paths, check if exists. HF model IDs will download automatically.
+        is_local_path = self.model.startswith("/")
+        if is_local_path and not os.path.exists(self.model):
+            print(f"\n⏭️ SKIPPING: Local model not found at {self.model}")
+            self.runner.full_report += (
+                f"\n⏭️ Test skipped: Local model not found at {self.model}\n"
+            )
+            self.runner.write_final_report()
+            return
+
+        # Log model source
+        if is_local_path:
+            print(f"📁 Using local model: {self.model}")
+        else:
+            print(
+                f"📥 Using HuggingFace model: {self.model} (will download if not cached)"
+            )
+
+        try:
+            for variant_config in self.variants:
+                with self.subTest(variant=variant_config["name"]):
+                    result_tuple = self.runner.run_benchmark_for_model(
+                        model_path=self.model,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=variant_config["other_args"],
+                        variant=variant_config["name"],
+                        extra_bench_args=["--trust-remote-code"],
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    # Use simplified report format without traces
+                    if results:
+                        self.runner.full_report += (
+                            generate_simple_markdown_report(results) + "\n"
+                        )
+        finally:
+            self.runner.write_final_report()
+
+        if failed_variants:
+            raise AssertionError(
+                f"Benchmark failed for {self.model} with the following variants: "
+                f"{', '.join(failed_variants)}"
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,126 @@
+"""MI35x Nightly performance benchmark for Grok-1 INT4 (W4A8KV8).
+
+This test benchmarks Grok-1 (314B MOE) with INT4 weight quantization on 8 GPUs.
+
+Registry: nightly-perf-8-gpu-mi35x-grok1-int4 suite
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - Grok-1 INT4 benchmark on MI35x (~25 min)
+register_amd_ci(
+    est_time=1500, suite="nightly-perf-8-gpu-mi35x-grok1-int4", nightly=True
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model and tokenizer paths can be overridden via environment variables
+GROK1_MODEL_PATH = os.environ.get("GROK1_MODEL_PATH", "amd/grok-1-W4A8KV8")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+PROFILE_DIR = "performance_profiles_grok1_int4_mi35x"
+
+
+class TestGrok1INT4PerfMI35x(unittest.TestCase):
+    """Test suite for Grok-1 INT4 performance benchmarks on MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "grok1-int4-mi35x",
+            "model_path": GROK1_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--quantization",
+                "fp8",
+                "--mem-fraction-static",
+                "0.85",
+                "--tokenizer-path",
+                GROK1_TOKENIZER_PATH,
+                "--attention-backend",
+                "aiter",
+            ],
+            "env_vars": {
+                "RCCL_MSCCL_ENABLE": "0",
+                "SGLANG_USE_AITER": "1",
+                "SGLANG_INT4_WEIGHT": "1",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_grok1_int4_perf(self):
+        """Run Grok-1 INT4 performance benchmark on MI35x."""
+        # Set environment variables
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for Grok-1 INT4 on MI35x")
+        finally:
+            # Restore original environment
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,126 @@
+"""MI35x Nightly performance benchmark for Grok-2.
+
+This test benchmarks Grok-2 with FP8 quantization on 8 GPUs.
+
+Registry: nightly-perf-8-gpu-mi35x-grok2 suite
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - Grok-2 benchmark on MI35x (~25 min)
+register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-mi35x-grok2", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model and tokenizer paths can be overridden via environment variables
+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"
+)
+PROFILE_DIR = "performance_profiles_grok2_mi35x"
+
+
+class TestGrok2PerfMI35x(unittest.TestCase):
+    """Test suite for Grok-2 performance benchmarks on MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "grok2-mi35x",
+            "model_path": GROK2_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--quantization",
+                "fp8",
+                "--mem-fraction-static",
+                "0.85",
+                "--tokenizer-path",
+                GROK2_TOKENIZER_PATH,
+                "--attention-backend",
+                "aiter",
+            ],
+            "env_vars": {
+                "RCCL_MSCCL_ENABLE": "0",
+                "SGLANG_USE_AITER": "1",
+                "SGLANG_INT4_WEIGHT": "0",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_grok2_perf(self):
+        """Run Grok-2 performance benchmark on MI35x."""
+        # Set environment variables
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for Grok-2 on MI35x")
+        finally:
+            # Restore original environment
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,145 @@
+"""Nightly performance benchmark for DeepSeek-V3.1 model.
+
+This test benchmarks the DeepSeek-V3.1 model with basic and MTP configurations on 8 GPUs.
+
+The model path can be configured via DEEPSEEK_V31_MODEL_PATH environment variable.
+
+Example usage:
+    DEEPSEEK_V31_MODEL_PATH=deepseek-ai/DeepSeek-V3.1 python -m pytest test_deepseek_v31_perf.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - DeepSeek-V3.1 benchmark (basic + MTP, ~300 min)
+register_amd_ci(est_time=18000, suite="nightly-perf-8-gpu-deepseek-v31", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path can be overridden via environment variable
+DEEPSEEK_V31_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V31_MODEL_PATH", "deepseek-ai/DeepSeek-V3.1"
+)
+PROFILE_DIR = "performance_profiles_deepseek_v31"
+
+
+class TestNightlyDeepseekV31Performance(unittest.TestCase):
+    """Nightly performance benchmark for DeepSeek-V3.1 model.
+
+    Tests the DeepSeek-V3.1 model with both basic and MTP configurations on TP=8.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V31_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Define variant configurations for DeepSeek-V3.1
+        cls.variants = [
+            {
+                "name": "basic",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--mem-fraction-static",
+                    "0.85",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "mtp",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-num-steps",
+                    "3",
+                    "--speculative-eagle-topk",
+                    "1",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                    "--mem-fraction-static",
+                    "0.7",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        # Override full_report to remove traces help text
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark across all configured variants."""
+        failed_variants = []
+
+        try:
+            for variant_config in self.variants:
+                with self.subTest(variant=variant_config["name"]):
+                    result_tuple = self.runner.run_benchmark_for_model(
+                        model_path=self.model,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=variant_config["other_args"],
+                        variant=variant_config["name"],
+                        extra_bench_args=["--trust-remote-code"],
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    # Use simplified report format without traces
+                    if results:
+                        self.runner.full_report += (
+                            generate_simple_markdown_report(results) + "\n"
+                        )
+        finally:
+            self.runner.write_final_report()
+
+        if failed_variants:
+            raise AssertionError(
+                f"Benchmark failed for {self.model} with the following variants: "
+                f"{', '.join(failed_variants)}"
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,145 @@
+"""Nightly performance benchmark for DeepSeek-V3 model.
+
+This test benchmarks the DeepSeek-V3 model with basic and MTP configurations on 8 GPUs.
+
+The model path can be configured via DEEPSEEK_V3_MODEL_PATH environment variable.
+
+Example usage:
+    DEEPSEEK_V3_MODEL_PATH=deepseek-ai/DeepSeek-V3-0324 python -m pytest test_deepseek_v3_perf.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - DeepSeek-V3 benchmark (basic + MTP, ~300 min)
+register_amd_ci(est_time=18000, suite="nightly-perf-8-gpu-deepseek-v3", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path can be overridden via environment variable
+DEEPSEEK_V3_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V3_MODEL_PATH", "deepseek-ai/DeepSeek-V3-0324"
+)
+PROFILE_DIR = "performance_profiles_deepseek_v3"
+
+
+class TestNightlyDeepseekV3Performance(unittest.TestCase):
+    """Nightly performance benchmark for DeepSeek-V3 model.
+
+    Tests the DeepSeek-V3 model with both basic and MTP configurations on TP=8.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V3_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Define variant configurations for DeepSeek-V3
+        cls.variants = [
+            {
+                "name": "basic",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--mem-fraction-static",
+                    "0.85",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "mtp",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-num-steps",
+                    "3",
+                    "--speculative-eagle-topk",
+                    "1",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                    "--mem-fraction-static",
+                    "0.7",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        # Override full_report to remove traces help text
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark across all configured variants."""
+        failed_variants = []
+
+        try:
+            for variant_config in self.variants:
+                with self.subTest(variant=variant_config["name"]):
+                    result_tuple = self.runner.run_benchmark_for_model(
+                        model_path=self.model,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=variant_config["other_args"],
+                        variant=variant_config["name"],
+                        extra_bench_args=["--trust-remote-code"],
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    # Use simplified report format without traces
+                    if results:
+                        self.runner.full_report += (
+                            generate_simple_markdown_report(results) + "\n"
+                        )
+        finally:
+            self.runner.write_final_report()
+
+        if failed_variants:
+            raise AssertionError(
+                f"Benchmark failed for {self.model} with the following variants: "
+                f"{', '.join(failed_variants)}"
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,133 @@
+"""Nightly performance benchmark for Grok-1 FP8.
+
+This test benchmarks Grok-1 (314B MOE) with FP8 quantization on 8 GPUs.
+
+Model paths can be configured via environment variables:
+- GROK1_MODEL_PATH: Path to Grok-1 model (default: lmzheng/grok-1)
+- GROK1_TOKENIZER_PATH: Path to Grok-1 tokenizer (default: Xenova/grok-1-tokenizer)
+
+Example usage:
+    python -m pytest test_grok1_fp8_perf.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - Grok-1 FP8 benchmark (~25 min)
+register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-grok1-fp8", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model and tokenizer paths can be overridden via environment variables
+GROK1_MODEL_PATH = os.environ.get("GROK1_MODEL_PATH", "lmzheng/grok-1")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+PROFILE_DIR = "performance_profiles_grok1_fp8"
+
+
+class TestNightlyGrok1FP8Performance(unittest.TestCase):
+    """Nightly performance benchmark for Grok-1 FP8.
+
+    Tests Grok-1 (314B MOE) with FP8 quantization on TP=8.
+    Runtime: ~25 minutes
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "grok1-fp8",
+            "model_path": GROK1_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--quantization",
+                "fp8",
+                "--mem-fraction-static",
+                "0.85",
+                "--tokenizer-path",
+                GROK1_TOKENIZER_PATH,
+                "--attention-backend",
+                "aiter",
+            ],
+            "env_vars": {
+                "RCCL_MSCCL_ENABLE": "0",
+                "SGLANG_USE_AITER": "1",
+                "SGLANG_INT4_WEIGHT": "0",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_grok1_fp8(self):
+        """Run benchmark for Grok-1 FP8."""
+        # Set environment variables
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for Grok-1 FP8")
+        finally:
+            # Restore original environment
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,133 @@
+"""Nightly performance benchmark for Grok-1 INT4 (W4A8KV8).
+
+This test benchmarks Grok-1 (314B MOE) with INT4 weight quantization on 8 GPUs.
+
+Model paths can be configured via environment variables:
+- GROK1_MODEL_PATH: Path to Grok-1 INT4 model (default: amd/grok-1-W4A8KV8)
+- GROK1_TOKENIZER_PATH: Path to Grok-1 tokenizer (default: Xenova/grok-1-tokenizer)
+
+Example usage:
+    python -m pytest test_grok1_int4_perf.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - Grok-1 INT4 benchmark (~25 min)
+register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-grok1-int4", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model and tokenizer paths can be overridden via environment variables
+GROK1_MODEL_PATH = os.environ.get("GROK1_MODEL_PATH", "amd/grok-1-W4A8KV8")
+GROK1_TOKENIZER_PATH = os.environ.get("GROK1_TOKENIZER_PATH", "Xenova/grok-1-tokenizer")
+PROFILE_DIR = "performance_profiles_grok1_int4"
+
+
+class TestNightlyGrok1INT4Performance(unittest.TestCase):
+    """Nightly performance benchmark for Grok-1 INT4 (W4A8KV8).
+
+    Tests Grok-1 (314B MOE) with INT4 weight quantization on TP=8.
+    Runtime: ~25 minutes
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "grok1-int4",
+            "model_path": GROK1_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--quantization",
+                "fp8",
+                "--mem-fraction-static",
+                "0.85",
+                "--tokenizer-path",
+                GROK1_TOKENIZER_PATH,
+                "--attention-backend",
+                "aiter",
+            ],
+            "env_vars": {
+                "RCCL_MSCCL_ENABLE": "0",
+                "SGLANG_USE_AITER": "1",
+                "SGLANG_INT4_WEIGHT": "1",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_grok1_int4(self):
+        """Run benchmark for Grok-1 INT4."""
+        # Set environment variables
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for Grok-1 INT4")
+        finally:
+            # Restore original environment
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,135 @@
+"""Nightly performance benchmark for Grok-2.
+
+This test benchmarks Grok-2 with FP8 quantization on 8 GPUs.
+
+Model paths can be configured via environment variables:
+- GROK2_MODEL_PATH: Path to Grok-2 model (default: xai-org/grok-2)
+- GROK2_TOKENIZER_PATH: Path to Grok-2 tokenizer (default: alvarobartt/grok-2-tokenizer)
+
+Example usage:
+    python -m pytest test_grok2_perf.py -v
+"""
+
+import os
+import unittest
+from typing import List
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.nightly_bench_utils import BenchmarkResult
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+# Register for AMD CI - Grok-2 benchmark (~25 min)
+register_amd_ci(est_time=1500, suite="nightly-perf-8-gpu-grok2", nightly=True)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns."""
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    for result in results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model and tokenizer paths can be overridden via environment variables
+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"
+)
+PROFILE_DIR = "performance_profiles_grok2"
+
+
+class TestNightlyGrok2Performance(unittest.TestCase):
+    """Nightly performance benchmark for Grok-2.
+
+    Tests Grok-2 with FP8 quantization on TP=8.
+    Runtime: ~25 minutes
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.model_config = {
+            "name": "grok2",
+            "model_path": GROK2_MODEL_PATH,
+            "other_args": [
+                "--trust-remote-code",
+                "--tp",
+                "8",
+                "--quantization",
+                "fp8",
+                "--mem-fraction-static",
+                "0.85",
+                "--tokenizer-path",
+                GROK2_TOKENIZER_PATH,
+                "--attention-backend",
+                "aiter",
+            ],
+            "env_vars": {
+                "RCCL_MSCCL_ENABLE": "0",
+                "SGLANG_USE_AITER": "1",
+                "SGLANG_INT4_WEIGHT": "0",
+            },
+        }
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_grok2(self):
+        """Run benchmark for Grok-2."""
+        # Set environment variables
+        old_env = {}
+        for key, value in self.model_config.get("env_vars", {}).items():
+            old_env[key] = os.environ.get(key)
+            os.environ[key] = value
+            print(f"Setting env: {key}={value}")
+
+        try:
+            result_tuple = self.runner.run_benchmark_for_model(
+                model_path=self.model_config["model_path"],
+                batch_sizes=self.batch_sizes,
+                input_lens=self.input_lens,
+                output_lens=self.output_lens,
+                other_args=self.model_config["other_args"],
+                variant=self.model_config["name"],
+                extra_bench_args=["--trust-remote-code"],
+            )
+            results = result_tuple[0]
+            success = result_tuple[1]
+
+            if results:
+                self.runner.full_report += (
+                    generate_simple_markdown_report(results) + "\n"
+                )
+
+            self.assertTrue(success, "Benchmark failed for Grok-2")
+        finally:
+            # Restore original environment
+            for key, value in old_env.items():
+                if value is None:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = value
+            self.runner.write_final_report()
+
+
+if __name__ == "__main__":
+    unittest.main()