[XPU] Integrate MoE and minor improvements in XPU attention backend (#13561)

test/srt/run_suite.py

@@ -74,6 +74,7 @@ suite_xeon = {
 suite_xpu = {
     "per-commit-xpu": [
         TestFile("xpu/test_intel_xpu_backend.py"),
+        TestFile("xpu/test_deepseek_ocr.py"),
     ],
 }
 

test/srt/xpu/test_deepseek_ocr.py

@@ -0,0 +1,121 @@
+"""
+python3 -m unittest test_deepseek_ocr.py
+"""
+
+import json
+import os
+import unittest
+
+import requests
+from transformers import AutoTokenizer
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+
+class TestDeepSeekOCR(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "deepseek-ai/DeepSeek-OCR"
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model, use_fast=False)
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.common_args = [
+            "--device",
+            "xpu",
+            "--attention-backend",
+            "intel_xpu",
+        ]
+        os.environ["SGLANG_USE_SGL_XPU"] = "1"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                *cls.common_args,
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        """Fixture that is run once after all tests in the class."""
+        kill_process_tree(cls.process.pid)
+
+    def get_request_json(self, max_new_tokens=32, n=1):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "<image>\n<|grounding|>Convert the document to pure text.",
+                "image_data": "../../examples/assets/example_image.png",
+                "sampling_params": {
+                    "temperature": 0 if n == 1 else 0.5,
+                    "max_new_tokens": max_new_tokens,
+                },
+            },
+        )
+        return response.json()
+
+    def run_decode(
+        self,
+        max_new_tokens=128,
+        n=1,
+    ):
+
+        ret = self.get_request_json(max_new_tokens=max_new_tokens, n=n)
+        print(json.dumps(ret, indent=2))
+
+        def assert_one_item(item):
+            if item["meta_info"]["finish_reason"]["type"] == "stop":
+                self.assertEqual(
+                    item["meta_info"]["finish_reason"]["matched"],
+                    self.tokenizer.eos_token_id,
+                )
+            elif item["meta_info"]["finish_reason"]["type"] == "length":
+                self.assertEqual(
+                    len(item["output_ids"]), item["meta_info"]["completion_tokens"]
+                )
+                self.assertEqual(len(item["output_ids"]), max_new_tokens)
+
+        # Determine whether to assert a single item or multiple items based on n
+        if n == 1:
+            assert_one_item(ret)
+        else:
+            self.assertEqual(len(ret), n)
+            for i in range(n):
+                assert_one_item(ret[i])
+
+        print("=" * 100)
+
+    def test_moe(self):
+        self.run_decode()
+
+
+class TestDeepSeekOCRTriton(TestDeepSeekOCR):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "deepseek-ai/DeepSeek-OCR"
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model, use_fast=False)
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.common_args = [
+            "--device",
+            "xpu",
+            "--attention-backend",
+            "intel_xpu",
+        ]
+        os.environ["SGLANG_USE_SGL_XPU"] = "0"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                *cls.common_args,
+            ],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()