Add graph runner support with torch compile on CPU (#7843)

test/srt/run_suite.py

@@ -276,6 +276,7 @@ suite_xeon = {
         TestFile("cpu/test_shared_expert.py"),
         TestFile("cpu/test_topk.py"),
         TestFile("test_intel_amx_attention_backend.py"),
+        TestFile("test_cpu_graph.py"),
     ],
 }
 

test/srt/test_cpu_graph.py

@@ -0,0 +1,87 @@
+"""
+Usage:
+python3 -m unittest test_cpu_graph.TestCPUGraph.test_mmlu_torch_compile_cpu
+"""
+
+import copy
+import os
+import unittest
+from types import SimpleNamespace
+
+from test_intel_amx_attention_backend import intel_amx_benchmark
+
+from sglang.srt.utils import get_cpu_ids_by_node, 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,
+)
+
+
+class TestCPUGraph(CustomTestCase):
+
+    @intel_amx_benchmark(
+        extra_args=[
+            "--batch-size",
+            "1",
+            "--mem-fraction-static",
+            "0.05",
+            "--enable-torch-compile",
+            "--torch-compile-max-bs",
+            "1",
+        ],
+        min_throughput=10,
+    )
+    def test_latency_torch_compile_cpu(self):
+        return DEFAULT_MLA_MODEL_NAME_FOR_TEST
+
+    def test_mmlu_torch_compile_cpu(self):
+        model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
+        base_url = DEFAULT_URL_FOR_TEST
+        cpu_ids_by_node = get_cpu_ids_by_node()
+        n_numa_node = len(cpu_ids_by_node)
+        env = copy.deepcopy(os.environ)
+        env["SGLANG_CPU_OMP_THREADS_BIND"] = "all"
+        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",
+                "--enable-torch-compile",
+                "--torch-compile-max-bs",
+                "1",
+                "--tp",
+                f"{n_numa_node}",
+            ],
+            env=env,
+        )
+
+        try:
+            args = SimpleNamespace(
+                base_url=base_url,
+                model=model,
+                eval_name="mmlu",
+                num_examples=64,
+                num_threads=32,
+            )
+
+            metrics = run_eval(args)
+            if is_in_ci():
+                self.assertGreater(metrics["score"], 0.45)
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_intel_amx_attention_backend.py

@@ -3,7 +3,6 @@ Usage:
 python3 -m unittest test_intel_amx_attention_backend.TestIntelAMXAttnBackend.test_mmlu
 """
 
-import os
 import unittest
 from functools import wraps
 from types import SimpleNamespace
@@ -35,8 +34,6 @@ def intel_amx_benchmark(extra_args=None, min_throughput=None):
                 "intel_amx",
                 "--disable-radix",
                 "--trust-remote-code",
-                "--batch-size",
-                "4",
             ]
             full_args = common_args + (extra_args or [])
 
@@ -60,28 +57,33 @@ def intel_amx_benchmark(extra_args=None, min_throughput=None):
 
 class TestIntelAMXAttnBackend(CustomTestCase):
 
-    @intel_amx_benchmark(min_throughput=10)
+    @intel_amx_benchmark(extra_args=["--batch-size", "4"], min_throughput=10)
     def test_latency_mla_model(self):
         return DEFAULT_MLA_MODEL_NAME_FOR_TEST
 
-    @intel_amx_benchmark(min_throughput=40)
+    @intel_amx_benchmark(extra_args=["--batch-size", "4"], min_throughput=40)
     def test_latency_default_model(self):
         return DEFAULT_MODEL_NAME_FOR_TEST
 
-    @intel_amx_benchmark(min_throughput=150)
+    @intel_amx_benchmark(extra_args=["--batch-size", "4"], min_throughput=150)
     def test_latency_fp8_qwen(self):
         return DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8
 
-    @intel_amx_benchmark(min_throughput=50)
+    @intel_amx_benchmark(extra_args=["--batch-size", "4"], min_throughput=50)
     def test_latency_fp8_moe_model(self):
         return DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE
 
-    @intel_amx_benchmark(extra_args=["--quantization", "w8a8_int8"], min_throughput=100)
+    @intel_amx_benchmark(
+        extra_args=["--batch-size", "4", "--quantization", "w8a8_int8"],
+        min_throughput=100,
+    )
     def test_latency_w8a8_default_model(self):
         return DEFAULT_MODEL_NAME_FOR_TEST_W8A8
 
     @intel_amx_benchmark(
         extra_args=[
+            "--batch-size",
+            "4",
             "--quantization",
             "w8a8_int8",
             "--mem-fraction-static",