minor: update gsm8k eval (#2091)

test/srt/test_nightly_gsm8k_eval.py

+import json
+import os
 import unittest
+from datetime import datetime
 from types import SimpleNamespace
 
 from sglang.srt.utils import kill_child_process
@@ -14,6 +17,26 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+MODEL_SCORE_THRESHOLDS = {
+    "meta-llama/Llama-3.1-8B-Instruct": 0.8316,
+    "mistralai/Mistral-7B-Instruct-v0.3": 0.5861,
+    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.8672,
+    "google/gemma-2-27b-it": 0.9227,
+    "meta-llama/Llama-3.1-70B-Instruct": 0.9623,
+    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.6415,
+    "Qwen/Qwen2-57B-A14B-Instruct": 0.8791,
+    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.8672,
+    "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.5544,
+    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8": 0.8356,
+    "neuralmagic/gemma-2-2b-it-FP8": 0.6059,
+    "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8": 0.9504,
+    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.6138,
+    "neuralmagic/Qwen2-72B-Instruct-FP8": 0.9504,
+    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.8197,
+    "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4": 0.8395,
+    "hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4": 0.8435,
+}
+
 
 def parse_models(model_string):
     return [model.strip() for model in model_string.split(",") if model.strip()]
@@ -23,10 +46,8 @@ def launch_server(base_url, model, is_fp8, is_tp2):
     other_args = ["--log-level-http", "warning", "--trust-remote-code"]
     if is_fp8:
         if "Llama-3" in model or "gemma-2" in model:
-            # compressed-tensors
             other_args.extend(["--kv-cache-dtype", "fp8_e5m2"])
         elif "Qwen2-72B-Instruct-FP8" in model:
-            # bug
             other_args.extend(["--quantization", "fp8"])
         else:
             other_args.extend(["--quantization", "fp8", "--kv-cache-dtype", "fp8_e5m2"])
@@ -48,6 +69,49 @@ def launch_server(base_url, model, is_fp8, is_tp2):
     return process
 
 
+def write_results_to_json(model, metrics, mode="a"):
+    result = {
+        "timestamp": datetime.now().isoformat(),
+        "model": model,
+        "metrics": metrics,
+        "score": metrics["score"],
+    }
+
+    existing_results = []
+    if mode == "a" and os.path.exists("results.json"):
+        try:
+            with open("results.json", "r") as f:
+                existing_results = json.load(f)
+        except json.JSONDecodeError:
+            existing_results = []
+
+    if isinstance(existing_results, list):
+        existing_results.append(result)
+    else:
+        existing_results = [result]
+
+    with open("results.json", "w") as f:
+        json.dump(existing_results, f, indent=2)
+
+
+def check_model_scores(results):
+    failed_models = []
+    for model, score in results:
+        threshold = MODEL_SCORE_THRESHOLDS.get(model)
+        if threshold is None:
+            print(f"Warning: No threshold defined for model {model}")
+            continue
+
+        if score < threshold:
+            failed_models.append(
+                f"\nScore Check Failed: {model}\n"
+                f"Model {model} score ({score:.4f}) is below threshold ({threshold:.4f})"
+            )
+
+    if failed_models:
+        raise AssertionError("\n".join(failed_models))
+
+
 class TestEvalAccuracyLarge(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
@@ -68,6 +132,9 @@ class TestEvalAccuracyLarge(unittest.TestCase):
             kill_child_process(self.process.pid, include_self=True)
 
     def test_mgsm_en_all_models(self):
+        is_first = True
+        all_results = []
+
         for model_group, is_fp8, is_tp2 in self.model_groups:
             for model in model_group:
                 with self.subTest(model=model):
@@ -85,11 +152,24 @@ class TestEvalAccuracyLarge(unittest.TestCase):
                     print(
                         f"{'=' * 42}\n{model} - metrics={metrics} score={metrics['score']}\n{'=' * 42}\n"
                     )
-                    # loosely threshold
-                    assert metrics["score"] > 0.5, f"score={metrics['score']} <= 0.5"
+
+                    write_results_to_json(model, metrics, "w" if is_first else "a")
+                    is_first = False
+
+                    all_results.append((model, metrics["score"]))
 
                     self.tearDown()
 
+        try:
+            with open("results.json", "r") as f:
+                print("\nFinal Results from results.json:")
+                print(json.dumps(json.load(f), indent=2))
+        except Exception as e:
+            print(f"Error reading results.json: {e}")
+
+        # Check all scores after collecting all results
+        check_model_scores(all_results)
+
 
 if __name__ == "__main__":
     unittest.main()