[NPU] Add PR test cases (#32392)

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

python/sglang/test/ascend/npu_eval_accuracy_kit.py

@@ -0,0 +1,49 @@
+"""NPU accuracy test mixin.
+
+Wraps :class:`~sglang.test.kits.eval_accuracy_kit.GSM8KMixin` with:
+
+* 1% accuracy tolerance (threshold * 0.99)
+* Automatic retry (up to 3 attempts)
+
+Delegates to ``super().test_gsm8k()`` so community-side changes to eval
+setup, backend selection, or parameters are inherited automatically.
+
+Usage::
+
+    from sglang.test.ascend.npu_eval_accuracy_kit import NPUGSM8KMixin
+
+    class TestFoo(CustomTestCase, NPUGSM8KMixin):
+        gsm8k_accuracy_thres = 0.6
+"""
+
+from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
+
+_NPU_ACCURACY_TOLERANCE = 0.99
+_NPU_MAX_ACCURACY_ATTEMPTS = 3
+
+
+class NPUGSM8KMixin(GSM8KMixin):
+    """NPU GSM8K accuracy mixin — 1% tolerance + up to 3 retries."""
+
+    def test_gsm8k(self):
+        threshold = self.gsm8k_score_threshold
+        if threshold != threshold:  # NaN → legacy alias
+            threshold = self.gsm8k_accuracy_thres
+        relaxed = threshold * _NPU_ACCURACY_TOLERANCE
+
+        # Patch both so the parent picks up the relaxed threshold.
+        self.gsm8k_score_threshold = relaxed
+        self.gsm8k_accuracy_thres = relaxed
+
+        for attempt in range(_NPU_MAX_ACCURACY_ATTEMPTS):
+            try:
+                super().test_gsm8k()
+                return
+            except AssertionError:
+                if attempt == _NPU_MAX_ACCURACY_ATTEMPTS - 1:
+                    raise
+                print(
+                    f"[{type(self).__name__}] GSM8K attempt "
+                    f"{attempt + 1}/{_NPU_MAX_ACCURACY_ATTEMPTS} failed, "
+                    f"retrying..."
+                )

python/sglang/test/ascend/test_ascend_utils.py

@@ -212,7 +212,7 @@ QWEN3_NEXT_80B_A3B_INSTRUCT_WEIGHTS_PATH = os.path.join(
     MODEL_WEIGHTS_DIR, "Qwen/Qwen3-Next-80B-A3B-Instruct"
 )
 QWEN3_32B_EAGLE3_WEIGHTS_PATH = os.path.join(
-    MODEL_WEIGHTS_DIR, "Zjcxy-SmartAI/Qwen3-32B-Eagle3"
+    MODEL_WEIGHTS_DIR, "Zjcxy-SmartAI/Eagle3-Qwen3-32B-zh"
 )
 QWEN3_32B_W8A8_MINDIE_WEIGHTS_PATH = os.path.join(
     MODEL_WEIGHTS_DIR, "aleoyang/Qwen3-32B-w8a8-MindIE"