migrate NPU PR/nightly test cases to a3-560T (#33346)

Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>

python/sglang/test/ascend/gsm8k_ascend_mixin.py

@@ -66,6 +66,14 @@ class GSM8KAscendMixin(ABC):
         kill_process_tree(cls.process.pid)
 
     def test_gsm8k(self):
+        from sglang.test.ascend.npu_eval_accuracy_kit import (
+            _is_pr_pipeline,
+            run_npu_pr_smoke,
+        )
+
+        if _is_pr_pipeline:
+            run_npu_pr_smoke(self.base_url)
+            return
         accuracy_threshold = getattr(self, "accuracy", 0.00)
         output_throughput_threshold = getattr(self, "output_throughput", 0.00)
 

python/sglang/test/ascend/npu_eval_accuracy_kit.py

@@ -16,16 +16,38 @@ Usage::
         gsm8k_accuracy_thres = 0.6
 """
 
+import os
+
+import requests
+
 from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
 
 _NPU_ACCURACY_TOLERANCE = 0.99
 _NPU_MAX_ACCURACY_ATTEMPTS = 3
 
+_is_pr_pipeline = os.environ.get("GITHUB_EVENT_NAME") == "pull_request"
+
+
+def run_npu_pr_smoke(base_url):
+    """Send a single inference request for PR pipeline smoke test."""
+    response = requests.post(
+        f"{base_url}/generate",
+        json={
+            "text": "The capital of France is",
+            "sampling_params": {"temperature": 0, "max_new_tokens": 32},
+        },
+    )
+    assert response.status_code == 200
+    assert "Paris" in response.text
+
 
 class NPUGSM8KMixin(GSM8KMixin):
     """NPU GSM8K accuracy mixin — 1% tolerance + up to 3 retries."""
 
     def test_gsm8k(self):
+        if _is_pr_pipeline:
+            run_npu_pr_smoke(self.base_url)
+            return
         threshold = self.gsm8k_score_threshold
         if threshold != threshold:  # NaN → legacy alias
             threshold = self.gsm8k_accuracy_thres