ci: enhance xeon ci (#9395)

.github/workflows/pr-test-xeon.yml

@@ -28,6 +28,8 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     runs-on: xeon-gnr
+    env:
+      HF_HOME: /home/sdp/.cache/huggingface
     strategy:
       matrix:
         build_type: ['all']
@@ -46,6 +48,7 @@ jobs:
         run: |
           docker run -dt \
             -v ${{ github.workspace }}:/sglang-checkout/ --ipc=host \
+            -v ${HF_HOME}:/root/.cache/huggingface \
             --name ci_sglang_xeon \
             sglang_xeon
 
@@ -67,13 +70,13 @@ jobs:
 
       - name: Run unit tests
         if: steps.check_amx.outcome == 'success'
-        timeout-minutes: 20
+        timeout-minutes: 30
         run: |
           docker exec -w /sglang-checkout/ ci_sglang_xeon \
             bash -c "cd ./test/srt && python3 run_suite.py --suite per-commit-cpu"
 
       - name: Change permission
-        timeout-minutes: 20
+        timeout-minutes: 2
         run: |
           docker exec -u root ci_sglang_xeon bash -c "
             rm -rf /tmp/ci-home  &&

python/sglang/test/test_utils.py

@@ -61,6 +61,12 @@ DEFAULT_MODEL_NAME_FOR_DYNAMIC_QUANT_ACCURACY_TEST_FP8 = (
 DEFAULT_MODEL_NAME_FOR_MODELOPT_QUANT_ACCURACY_TEST_FP8 = (
     "nvidia/Llama-3.1-8B-Instruct-FP8"
 )
+DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8 = "Qwen/Qwen3-1.7B-FP8"
+DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE = "gaunernst/DeepSeek-V2-Lite-Chat-FP8"
+
+# W8A8 models
+DEFAULT_MODEL_NAME_FOR_TEST_W8A8 = "RedHatAI/Llama-3.2-3B-quantized.w8a8"
+DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE = "nytopop/Qwen3-30B-A3B.w8a8"
 
 # EAGLE
 DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST = "meta-llama/Llama-2-7b-chat-hf"

test/srt/test_intel_amx_attention_backend.py

@@ -3,13 +3,20 @@ Usage:
 python3 -m unittest test_intel_amx_attention_backend.TestIntelAMXAttnBackend.test_mmlu
 """
 
+import os
 import unittest
+from functools import wraps
 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_MODEL_NAME_FOR_TEST,
+    DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE,
+    DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8,
+    DEFAULT_MODEL_NAME_FOR_TEST_W8A8,
+    DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
@@ -19,28 +26,75 @@ from sglang.test.test_utils import (
 )
 
 
-class TestIntelAMXAttnBackend(CustomTestCase):
+def intel_amx_benchmark(extra_args=None, min_throughput=None):
-    def test_latency(self):
+    def decorator(test_func):
-        prefill_latency, decode_throughput, decode_latency = run_bench_one_batch(
+        @wraps(test_func)
-            DEFAULT_MLA_MODEL_NAME_FOR_TEST,
+        def wrapper(self):
-            [
+            common_args = [
                 "--attention-backend",
                 "intel_amx",
-                "--mem-fraction-static",
-                "0.05",
                 "--disable-radix",
                 "--trust-remote-code",
                 "--batch-size",
                 "4",
-            ],
+            ]
-        )
+            full_args = common_args + (extra_args or [])
 
-        print(f"{prefill_latency=}")
+            model = test_func(self)
-        print(f"{decode_throughput=}")
+            prefill_latency, decode_throughput, decode_latency = run_bench_one_batch(
-        print(f"{decode_latency=}")
+                model, full_args
+            )
 
-        if is_in_ci():
+            print(f"{model=}")
-            self.assertGreater(decode_throughput, 10)
+            print(f"{prefill_latency=}")
+            print(f"{decode_throughput=}")
+            print(f"{decode_latency=}")
+
+            if is_in_ci() and min_throughput is not None:
+                self.assertGreater(decode_throughput, min_throughput)
+
+        return wrapper
+
+    return decorator
+
+
+class TestIntelAMXAttnBackend(CustomTestCase):
+
+    @intel_amx_benchmark(min_throughput=10)
+    def test_latency_mla_model(self):
+        return DEFAULT_MLA_MODEL_NAME_FOR_TEST
+
+    @intel_amx_benchmark(min_throughput=40)
+    def test_latency_default_model(self):
+        return DEFAULT_MODEL_NAME_FOR_TEST
+
+    @intel_amx_benchmark(min_throughput=150)
+    def test_latency_fp8_qwen(self):
+        return DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8
+
+    @intel_amx_benchmark(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)
+    def test_latency_w8a8_default_model(self):
+        return DEFAULT_MODEL_NAME_FOR_TEST_W8A8
+
+    @intel_amx_benchmark(
+        extra_args=[
+            "--quantization",
+            "w8a8_int8",
+            "--mem-fraction-static",
+            "0.9",
+            "--max-total-tokens",
+            "65536",
+            "--tp",
+            "6",
+        ],
+        min_throughput=100,
+    )
+    def test_latency_w8a8_moe_model(self):
+        return DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE
 
     def test_mmlu(self):
         model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
@@ -68,9 +122,9 @@ class TestIntelAMXAttnBackend(CustomTestCase):
                 num_examples=64,
                 num_threads=32,
             )
-
             metrics = run_eval(args)
-            self.assertGreater(metrics["score"], 0.45)
+            if is_in_ci():
+                self.assertGreater(metrics["score"], 0.45)
         finally:
             kill_process_tree(process.pid)