fix: fix the bug blocking completion template application (#17010)

Co-authored-by: xdtbynd <supercluster@vip.qq.com>
Co-authored-by: cy <chenyang08056032@163.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>

python/sglang/srt/managers/template_manager.py

@@ -30,6 +30,7 @@ from sglang.srt.parser.code_completion_parser import (
     FimPosition,
     completion_template_exists,
     register_completion_template,
+    set_completion_template,
 )
 from sglang.srt.parser.conversation import (
     Conversation,
@@ -199,6 +200,8 @@ class TemplateManager:
         else:
             self._completion_template_name = completion_template_arg
 
+        set_completion_template(self._completion_template_name)
+
     def initialize_templates(
         self,
         tokenizer_manager: TokenizerManager,

python/sglang/srt/parser/code_completion_parser.py

@@ -15,15 +15,16 @@
 
 import dataclasses
 import logging
-from enum import auto
+from enum import Enum, auto
+from typing import Optional
 
 from sglang.srt.entrypoints.openai.protocol import CompletionRequest
 
 logger = logging.getLogger(__name__)
-completion_template_name = None
+completion_template_name: Optional[str] = None
 
 
-class FimPosition:
+class FimPosition(Enum):
     """Position of fim middle token."""
 
     MIDDLE = auto()
@@ -68,6 +69,12 @@ def completion_template_exists(template_name: str) -> bool:
     return template_name in completion_templates
 
 
+def set_completion_template(template_name: str) -> None:
+    global completion_template_name
+    if completion_template_name is None:
+        completion_template_name = template_name
+
+
 def is_completion_template_defined() -> bool:
     global completion_template_name
     return completion_template_name is not None

test/registered/ascend/basic_function/deepseek_coder.json

@@ -0,0 +1,7 @@
+{
+  "name": "deepseek_coder",
+  "fim_begin_token": "<｜fim▁begin｜>",
+  "fim_middle_token": "<｜fim▁hole｜>",
+  "fim_end_token": "<｜fim▁end｜>",
+  "fim_position": "MIDDLE"
+}

test/registered/ascend/basic_function/test_ascend_fim_completion.py

@@ -0,0 +1,99 @@
+import unittest
+
+import openai
+
+from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils.hf_transformers_utils import get_tokenizer
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+
+
+class TestFimCompletion(CustomTestCase):
+    """Testcase：Verify set --completion-template, the model's FIM (Fill-in-the-Middle) completion function work correctly.
+
+    [Test Category] Parameter
+    [Test Target] --completion-template
+    """
+
+    model = "/root/.cache/modelscope/hub/models/deepseek-ai/deepseek-coder-1.3b-base"
+    other_args = [
+        "--completion-template",
+        "deepseek_coder",
+        "--attention-backend",
+        "ascend",
+        "--disable-cuda-graph",
+        "--mem-fraction-static",
+        0.8,
+    ]
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=cls.other_args,
+        )
+        cls.base_url += "/v1"
+        cls.tokenizer = get_tokenizer(cls.model)
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def run_fim_completion(self, number_of_completion):
+        client = openai.Client(api_key=self.api_key, base_url=self.base_url)
+        prompt = "function sum(a: number, b: number): number{\n"
+        suffix = "}"
+
+        prompt_input = self.tokenizer.encode(prompt) + self.tokenizer.encode(suffix)
+        num_prompt_tokens = len(prompt_input) + 2
+
+        response = client.completions.create(
+            model=self.model,
+            prompt=prompt,
+            suffix=suffix,
+            temperature=0.3,
+            max_tokens=32,
+            stream=False,
+            n=number_of_completion,
+        )
+        assert len(response.choices) == number_of_completion
+        assert response.id
+        assert response.created
+        assert response.object == "text_completion"
+        assert (
+            response.usage.prompt_tokens == num_prompt_tokens
+        ), f"{response.usage.prompt_tokens} vs {num_prompt_tokens}"
+        assert response.usage.completion_tokens > 0
+        assert response.usage.total_tokens > 0
+
+    def test_fim_completion(self):
+        for number_of_completion in [1, 3]:
+            self.run_fim_completion(number_of_completion)
+
+
+class TestFimCompletionJson(TestFimCompletion):
+    other_args = [
+        "--completion-template",
+        "./deepseek_coder.json",
+        "--attention-backend",
+        "ascend",
+        "--disable-cuda-graph",
+        "--mem-fraction-static",
+        0.8,
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()