From d44584e8d8b2eac5876f0d92028e59af8981d18a Mon Sep 17 00:00:00 2001
From: Jacob0226 <jacchang@amd.com>
Date: Wed, 27 May 2026 16:49:48 +0800
Subject: [PATCH] [AMD] [CI] Add GLM-5.1 MXFP4 TP2 accuracy gate (#26396)

---
 .../mi35x/test_glm51_mxfp4_tp2_gsm8k_mi35x.py | 144 ++++++++++++++++++
 1 file changed, 144 insertions(+)
 create mode 100644 test/registered/amd/accuracy/mi35x/test_glm51_mxfp4_tp2_gsm8k_mi35x.py

diff --git a/test/registered/amd/accuracy/mi35x/test_glm51_mxfp4_tp2_gsm8k_mi35x.py b/test/registered/amd/accuracy/mi35x/test_glm51_mxfp4_tp2_gsm8k_mi35x.py
new file mode 100644
index 000000000..b2c5ae956
--- /dev/null
+++ b/test/registered/amd/accuracy/mi35x/test_glm51_mxfp4_tp2_gsm8k_mi35x.py
@@ -0,0 +1,144 @@
+"""MI355X GLM-5.1-MXFP4 TP=2 GSM8K accuracy gate.
+
+This is a PR Test (AMD) regression test for the GLM-5.1-MXFP4 TP=2
+accuracy drop seen on MI355X/gfx950 when aiter selected a bad BF16 GEMM path.
+
+Registry: stage-c-test-large-8-gpu-amd-mi35x suite
+"""
+
+import os
+import resource
+import unittest
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+
+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_gsm8k_eval
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+register_amd_ci(
+    est_time=3600,
+    suite="stage-c-test-large-8-gpu-amd-mi35x",
+)
+
+GLM51_MXFP4_MODEL_ID = "amd/GLM-5.1-MXFP4"
+GLM51_MXFP4_LOCAL_PATHS = (
+    "/data2/models/amd-GLM-5.1-MXFP4",
+    "/data/huggingface/hub/amd/GLM-5.1-MXFP4",
+)
+
+GSM8K_ACCURACY_THRESHOLD = 0.92
+GSM8K_INVALID_THRESHOLD = 0.02
+DEFAULT_NUM_QUESTIONS = 1200
+DEFAULT_PARALLEL = 1200
+
+
+def _raise_nofile_limit() -> None:
+    """GSM8K with high parallelism can exceed the default soft nofile=1024."""
+    soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+    target = 65535 if hard == resource.RLIM_INFINITY else min(hard, 65535)
+    if soft < target:
+        resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))
+
+
+def _get_model_path() -> str:
+    env_path = os.environ.get("GLM51_MXFP4_MODEL_PATH")
+    if env_path:
+        return env_path
+    for path in GLM51_MXFP4_LOCAL_PATHS:
+        if os.path.exists(path):
+            return path
+    return GLM51_MXFP4_MODEL_ID
+
+
+class TestGLM51MXFP4TP2GSM8KMI35x(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        _raise_nofile_limit()
+        cls.model = _get_model_path()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            model=cls.model,
+            base_url=cls.base_url,
+            timeout=5400,
+            other_args=[
+                "--tp",
+                "2",
+                "--trust-remote-code",
+                "--reasoning-parser",
+                "glm45",
+                "--tool-call-parser",
+                "glm47",
+                "--watchdog-timeout",
+                "1200",
+                "--mem-fraction-static",
+                "0.85",
+                "--kv-cache-dtype",
+                "fp8_e4m3",
+                "--disable-radix-cache",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true, "num_threads": 8}',
+                "--dsa-prefill-backend",
+                "tilelang",
+                "--dsa-decode-backend",
+                "tilelang",
+                "--tokenizer-worker-num",
+                "4",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k_accuracy(self):
+        num_questions = int(
+            os.environ.get("GLM51_MXFP4_GSM8K_NUM_QUESTIONS", DEFAULT_NUM_QUESTIONS)
+        )
+        parallel = int(os.environ.get("GLM51_MXFP4_GSM8K_PARALLEL", DEFAULT_PARALLEL))
+        url = urlparse(self.base_url)
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=num_questions,
+            max_new_tokens=512,
+            parallel=parallel,
+            host=url.hostname or "127.0.0.1",
+            port=url.port or 30000,
+            temperature=0.0,
+        )
+
+        metrics = run_gsm8k_eval(args)
+        accuracy = metrics["accuracy"]
+        invalid = metrics["invalid"]
+        summary = (
+            "### GLM-5.1-MXFP4 TP=2 GSM8K (MI355X)\n\n"
+            "| Model | TP | Questions | Accuracy | Invalid | Threshold | Status |\n"
+            "| ----- | -- | --------- | -------- | ------- | --------- | ------ |\n"
+        )
+        passed = (
+            accuracy >= GSM8K_ACCURACY_THRESHOLD and invalid <= GSM8K_INVALID_THRESHOLD
+        )
+        status = "PASS" if passed else "FAIL"
+        summary += (
+            f"| {self.model} | 2 | {num_questions} | {accuracy:.3f} | "
+            f"{invalid:.3f} | accuracy >= {GSM8K_ACCURACY_THRESHOLD:.2f} | {status} |\n"
+        )
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        self.assertGreaterEqual(accuracy, GSM8K_ACCURACY_THRESHOLD)
+        self.assertLessEqual(invalid, GSM8K_INVALID_THRESHOLD)
+
+
+if __name__ == "__main__":
+    unittest.main()