[AMD] CI - Add MI35x nightly/PR tests for kv-cache-fp8 and allreduce-fusion (DeepSeek) (#19834)

Co-authored-by: bingxche <Bingxu.Chen@amd.com>

test/registered/amd/accuracy/mi35x/test_deepseek_r1_mxfp4_ar_fusion_eval_mi35x.py

@@ -0,0 +1,280 @@
+"""MI35x DeepSeek-R1-MXFP4 GSM8K Completion Evaluation Test with AIter AllReduce Fusion (8-GPU)
+
+Tests DeepSeek-R1-MXFP4 quantized model with --enable-aiter-allreduce-fusion
+using few-shot completion benchmark on MI35x.
+
+Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion 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 AllReduce Fusion accuracy test (~60 min)
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion",
+    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 with AllReduce Fusion."""
+    model_path = get_model_path()
+    return [
+        ModelConfig(
+            model_path=model_path,
+            tp_size=8,
+            accuracy_threshold=0.93,
+            timeout=3600,
+            variant="ar-fusion",
+            other_args=[
+                "--attention-backend",
+                "aiter",
+                "--chunked-prefill-size",
+                "131072",
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                "0.85",
+                "--trust-remote-code",
+                "--enable-aiter-allreduce-fusion",
+            ],
+            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 TestDeepSeekR1MXFP4ArFusionEvalMI35x(unittest.TestCase):
+    """DeepSeek-R1-MXFP4 GSM8K Evaluation with AllReduce Fusion 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_ar_fusion_accuracy(self):
+        """Test DeepSeek-R1-MXFP4 models with AllReduce Fusion on GSM8K."""
+        # 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 AllReduce Fusion 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"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": display_name,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": display_name,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi35x/test_deepseek_r1_mxfp4_kv_fp8_eval_mi35x.py

@@ -0,0 +1,281 @@
+"""MI35x DeepSeek-R1-MXFP4 GSM8K Completion Evaluation Test with KV Cache FP8 (8-GPU)
+
+Tests DeepSeek-R1-MXFP4 quantized model with --kv-cache-dtype fp8_e4m3
+using few-shot completion benchmark on MI35x.
+
+Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8 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 KV FP8 accuracy test (~60 min)
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8",
+    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 with KV cache FP8."""
+    model_path = get_model_path()
+    return [
+        ModelConfig(
+            model_path=model_path,
+            tp_size=8,
+            accuracy_threshold=0.93,
+            timeout=3600,
+            variant="kv-fp8",
+            other_args=[
+                "--attention-backend",
+                "aiter",
+                "--chunked-prefill-size",
+                "131072",
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                "0.85",
+                "--trust-remote-code",
+                "--kv-cache-dtype",
+                "fp8_e4m3",
+            ],
+            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 TestDeepSeekR1MXFP4KvFp8EvalMI35x(unittest.TestCase):
+    """DeepSeek-R1-MXFP4 GSM8K Evaluation with KV Cache FP8 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_kv_fp8_accuracy(self):
+        """Test DeepSeek-R1-MXFP4 models with KV cache FP8 on GSM8K."""
+        # 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 KV FP8 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"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": display_name,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": display_name,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/perf/mi35x/test_deepseek_r1_mxfp4_ar_fusion_perf_mi35x.py

@@ -0,0 +1,177 @@
+"""MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 model with AIter AllReduce Fusion.
+
+This test benchmarks the DeepSeek-R1-MXFP4 quantized model on MI35x with 8 GPUs
+using --enable-aiter-allreduce-fusion.
+
+The model path can be configured via DEEPSEEK_R1_MXFP4_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion suite
+
+Example usage:
+    DEEPSEEK_R1_MXFP4_MODEL_PATH=/data2/models/amd-DeepSeek-R1-MXFP4-Preview python -m pytest test_deepseek_r1_mxfp4_ar_fusion_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 AllReduce Fusion benchmark on MI35x (~300 min)
+register_amd_ci(
+    est_time=18000,
+    suite="nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion",
+    nightly=True,
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    # Skip first result if it's a warmup (same batch_size as second result)
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path 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_ar_fusion_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 TestDeepseekR1MXFP4ArFusionPerfMI35x(unittest.TestCase):
+    """MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 with AllReduce Fusion.
+
+    Tests the DeepSeek-R1-MXFP4 quantized model on TP=8 with --enable-aiter-allreduce-fusion.
+    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, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.variants = [
+            {
+                "name": "ar-fusion",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--chunked-prefill-size",
+                    "131072",
+                    "--disable-radix-cache",
+                    "--mem-fraction-static",
+                    "0.85",
+                    "--enable-aiter-allreduce-fusion",
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark across all configured variants."""
+        failed_variants = []
+
+        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
+
+        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"],
+                        enable_profile=False,
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    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()

test/registered/amd/perf/mi35x/test_deepseek_r1_mxfp4_kv_fp8_perf_mi35x.py

@@ -0,0 +1,178 @@
+"""MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 model with KV Cache FP8.
+
+This test benchmarks the DeepSeek-R1-MXFP4 quantized model on MI35x with 8 GPUs
+using --kv-cache-dtype fp8_e4m3.
+
+The model path can be configured via DEEPSEEK_R1_MXFP4_MODEL_PATH environment variable.
+
+Registry: nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8 suite
+
+Example usage:
+    DEEPSEEK_R1_MXFP4_MODEL_PATH=/data2/models/amd-DeepSeek-R1-MXFP4-Preview python -m pytest test_deepseek_r1_mxfp4_kv_fp8_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 KV FP8 benchmark on MI35x (~300 min)
+register_amd_ci(
+    est_time=18000,
+    suite="nightly-perf-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8",
+    nightly=True,
+)
+
+
+def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
+    """Generate a simplified markdown report without traces and cost columns.
+
+    Skips the first result if it's a warmup run (duplicate batch_size).
+    """
+    model_header = results[0].model_path
+    if results[0].run_name and results[0].run_name != "default":
+        model_header += f" ({results[0].run_name})"
+
+    gpu_config = os.getenv("GPU_CONFIG", "MI35x")
+    if gpu_config:
+        model_header += f" [{gpu_config}]"
+
+    summary = f"### {model_header}\n"
+    summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
+    summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
+
+    # Skip first result if it's a warmup (same batch_size as second result)
+    report_results = (
+        results[1:]
+        if len(results) > 1 and results[0].batch_size == results[1].batch_size
+        else results
+    )
+
+    for result in report_results:
+        itl = 1 / (result.output_throughput / result.batch_size) * 1000
+        summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
+
+    return summary
+
+
+# Model path 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_kv_fp8_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 TestDeepseekR1MXFP4KvFp8PerfMI35x(unittest.TestCase):
+    """MI35x Nightly performance benchmark for DeepSeek-R1-MXFP4 with KV Cache FP8.
+
+    Tests the DeepSeek-R1-MXFP4 quantized model on TP=8 with --kv-cache-dtype fp8_e4m3.
+    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, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        cls.variants = [
+            {
+                "name": "kv-fp8",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--chunked-prefill-size",
+                    "131072",
+                    "--disable-radix-cache",
+                    "--mem-fraction-static",
+                    "0.85",
+                    "--kv-cache-dtype",
+                    "fp8_e4m3",
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+        cls.runner.full_report = f"## {cls.__name__}\n"
+
+    def test_bench_one_batch(self):
+        """Run benchmark across all configured variants."""
+        failed_variants = []
+
+        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
+
+        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"],
+                        enable_profile=False,
+                    )
+                    results = result_tuple[0]
+                    success = result_tuple[1]
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    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()

test/registered/amd/test_deepseek_v3_basic_kv_fp8.py

@@ -0,0 +1,86 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_amd_ci,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+register_amd_ci(
+    est_time=1200, suite="nightly-amd-8-gpu-deepseek-v3-kv-fp8", nightly=True
+)
+
+FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
+
+
+class TestDeepseekV3BasicKvFp8(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = FULL_DEEPSEEK_V3_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true, "num_threads": 64}',
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        args = SimpleNamespace(
+            num_shots=8,
+            data_path=None,
+            num_questions=1400,
+            parallel=1400,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3 kv-fp8)\n" f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], 0.93)
+
+    def test_bs_1_speed(self):
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v3 kv-fp8)\n" f"{speed=:.2f} token/s\n"
+            )
+            if is_in_amd_ci():
+                self.assertGreater(speed, 40)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/test_deepseek_v3_mtp_kv_fp8.py

@@ -0,0 +1,116 @@
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_amd_ci,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+register_amd_ci(
+    est_time=1200, suite="nightly-amd-8-gpu-deepseek-v3-kv-fp8", nightly=True
+)
+
+FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
+
+
+class TestDeepseekV3MTPKvFp8(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = FULL_DEEPSEEK_V3_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "8",
+            "--trust-remote-code",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true, "num_threads": 64}',
+        ]
+        if not is_in_amd_ci():
+            other_args += ["--mem-frac", "0.7"]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
+        print(f"{avg_spec_accept_length=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3 mtp kv-fp8)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+            self.assertGreater(metrics["accuracy"], 0.93)
+            if is_in_amd_ci():
+                self.assertGreater(avg_spec_accept_length, 2.8)
+
+    def test_bs_1_speed(self):
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{acc_length=:.2f} {speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v3 mtp kv-fp8)\n"
+                f"{acc_length=:.2f}\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            if is_in_amd_ci():
+                self.assertGreater(acc_length, 2.8)
+            else:
+                self.assertGreater(acc_length, 2.9)
+            if is_in_amd_ci():
+                self.assertGreater(speed, 90)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ops/test_aiter_allreduce_fusion_amd.py

@@ -10,8 +10,7 @@ import torch
 
 from sglang.test.ci.ci_register import register_amd_ci
 
-# Dedicated AMD 8-GPU suite for AITER fused allreduce+rmsnorm validation.
-register_amd_ci(est_time=240, suite="stage-c-test-aiter-fusion-8-gpu-amd")
+register_amd_ci(est_time=240, suite="stage-c-test-large-8-gpu-amd")
 
 
 class TestAiterAllreduceFusionAmd(unittest.TestCase):

test/run_suite.py

@@ -26,7 +26,7 @@ PER_COMMIT_SUITES = {
         "stage-b-test-large-8-gpu-35x-disaggregation-amd",
         "stage-b-test-large-1-gpu-amd",
         "stage-b-test-large-2-gpu-amd",
-        "stage-c-test-aiter-fusion-8-gpu-amd",
+        "stage-c-test-large-8-gpu-amd",
         "stage-c-test-large-8-gpu-amd-mi35x",
     ],
     HWBackend.CUDA: [