[NPU] support GPTQ quantization on npu (#15203)

Signed-off-by: 22dimensions <waitingwind@foxmail.com>
2026-01-29 15:48:18 +08:00
parent cbf90d70ff
commit 7b79326751
5 changed files with 259 additions and 6 deletions
@@ -0,0 +1,73 @@
+import unittest
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+TEST_MODEL_MATRIX = {
+    "Qwen/Qwen3-1.7B-GPTQ-Int8": {
+        "accuracy": 0.65,
+    },
+}
+
+
+class TestAscendGPTQInt8(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = TEST_MODEL_MATRIX.keys()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.url = urlparse(DEFAULT_URL_FOR_TEST)
+        cls.common_args = [
+            "--trust-remote-code",
+            "--mem-fraction-static",
+            0.8,
+            "--attention-backend",
+            "ascend",
+            "--quantization",
+            "gptq",
+        ]
+
+    def test_a_gsm8k(self):
+        for model in self.models:
+            with self.subTest(model=model):
+                print(f"##=== Testing accuracy: {model} ===##")
+
+                process = popen_launch_server(
+                    model,
+                    self.base_url,
+                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                    other_args=[
+                        *self.common_args,
+                    ],
+                )
+
+                try:
+                    args = SimpleNamespace(
+                        num_shots=5,
+                        data_path=None,
+                        num_questions=1319,
+                        max_new_tokens=512,
+                        parallel=128,
+                        host=f"http://{self.url.hostname}",
+                        port=int(self.url.port),
+                    )
+
+                    metrics = run_eval_few_shot_gsm8k(args)
+                    self.assertGreaterEqual(
+                        metrics["accuracy"],
+                        TEST_MODEL_MATRIX[model]["accuracy"],
+                    )
+                finally:
+                    kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()