Add V2-lite model test

python/sglang/test/test_utils.py

@@ -5,6 +5,7 @@
 import logging
 import os
 import random
+import re
 import subprocess
 import threading
 import time
@@ -840,12 +841,23 @@ def run_bench_one_batch(model, other_args):
         print(f"Output: {output}", flush=True)
         print(f"Error: {error}", flush=True)
 
-        lastline = output.split("\n")[-3]
-        output_throughput = float(lastline.split(" ")[-2])
+        # Return prefill_latency, decode_throughput, decode_latency
+        prefill_line = output.split("\n")[-9]
+        decode_line = output.split("\n")[-3]
+        pattern = (
+            r"latency: (?P<latency>\d+\.\d+).*?throughput:\s*(?P<throughput>\d+\.\d+)"
+        )
+        match = re.search(pattern, prefill_line)
+        if match:
+            prefill_latency = float(match.group("latency"))
+        match = re.search(pattern, decode_line)
+        if match:
+            decode_latency = float(match.group("latency"))
+            decode_throughput = float(match.group("throughput"))
     finally:
         kill_process_tree(process.pid)
 
-    return output_throughput
+    return prefill_latency, decode_throughput, decode_latency
 
 
 def run_bench_offline_throughput(model, other_args):

test/srt/models/test_dummy_grok_models.py

@@ -6,7 +6,7 @@
 class TestDummyGrok1(CustomTestCase):
 
     def test_dummy_grok_1(self):
-        output_throughput = run_bench_one_batch(
+        _, output_throughput, _ = run_bench_one_batch(
             None,
             [
                 "--model",

test/srt/run_suite.py

@@ -198,6 +198,7 @@ class TestFile:
         TestFile("cpu/test_rope.py"),
         TestFile("cpu/test_shared_expert.py"),
         TestFile("cpu/test_topk.py"),
+        TestFile("test_intel_amx_attention_backend.py"),
     ],
     "nightly": [
         TestFile("test_nightly_gsm8k_eval.py"),

test/srt/test_bench_one_batch.py

@@ -20,7 +20,7 @@
 class TestBenchOneBatch(CustomTestCase):
 
     def test_bs1_small(self):
-        output_throughput = run_bench_one_batch(
+        _, output_throughput, _ = run_bench_one_batch(
             DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
         )
         self.assertGreater(output_throughput, 50)

test/srt/test_flashmla.py

@@ -67,7 +67,7 @@ def test_gsm8k(self):
 
 class TestFlashMLAAttnLatency(unittest.TestCase):
     def test_latency(self):
-        output_throughput = run_bench_one_batch(
+        _, output_throughput, _ = run_bench_one_batch(
             DEFAULT_MODEL_NAME_FOR_TEST_MLA,
             [
                 "--attention-backend",

test/srt/test_intel_amx_attention_backend.py

@@ -0,0 +1,79 @@
+"""
+Usage:
+python3 -m unittest test_intel_amx_attention_backend.TestIntelAMXAttnBackend.test_mmlu
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MLA_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    run_bench_one_batch,
+)
+
+
+class TestIntelAMXAttnBackend(CustomTestCase):
+    def test_latency(self):
+        prefill_latency, decode_throughput, decode_latency = run_bench_one_batch(
+            DEFAULT_MLA_MODEL_NAME_FOR_TEST,
+            [
+                "--attention-backend",
+                "intel_amx",
+                "--mem-fraction-static",
+                "0.05",
+                "--disable-radix",
+                "--trust-remote-code",
+                "--batch-size",
+                "4",
+            ],
+        )
+
+        print(f"{prefill_latency=}")
+        print(f"{decode_throughput=}")
+        print(f"{decode_latency=}")
+
+        if is_in_ci():
+            self.assertGreater(decode_throughput, 10)
+
+    def test_mmlu(self):
+        model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
+        base_url = DEFAULT_URL_FOR_TEST
+        process = popen_launch_server(
+            model,
+            base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--attention-backend",
+                "intel_amx",
+                "--mem-fraction-static",
+                "0.05",
+                "--disable-radix",
+                "--trust-remote-code",
+                "--disable-overlap-schedule",
+            ],
+        )
+
+        try:
+            args = SimpleNamespace(
+                base_url=base_url,
+                model=model,
+                eval_name="mmlu",
+                num_examples=64,
+                num_threads=32,
+            )
+
+            metrics = run_eval(args)
+            self.assertGreater(metrics["score"], 0.5)
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()