[NPU] enhance accuracy for model minimaxm2 from 16.5% to 95.5% (#17695)

python/sglang/srt/hardware_backend/npu/moe/topk.py

@@ -26,7 +26,7 @@ def fused_topk_npu(
     renormalize = topk_config.renormalize
     correction_bias = topk_config.correction_bias
 
-    if not use_grouped_topk:
+    if not use_grouped_topk and correction_bias is None:
         topk_weights, topk_ids, _ = torch.ops.npu.npu_moe_gating_top_k_softmax(
             router_logits,
             k=topk_config.top_k,

python/sglang/test/ascend/test_ascend_utils.py

@@ -148,6 +148,7 @@ STABLELM_2_1_6B_WEIGHTS_PATH = os.path.join(
     MODEL_WEIGHTS_DIR, "stabilityai/stablelm-2-1_6b"
 )
 XVERSE_MOE_A36B_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "xverse/XVERSE-MoE-A36B")
+MINIMAX_M2_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "cyankiwi/MiniMax-M2-BF16")
 
 # VLM model weights path
 DEEPSEEK_VL2_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "deepseek-ai/deepseek-vl2")

test/registered/ascend/llm_models/test_ascend_minimax_m2.py

@@ -0,0 +1,43 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import MINIMAX_M2_WEIGHTS_PATH
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_npu_ci(
+    est_time=400,
+    suite="nightly-8-npu-a3",
+    nightly=True,
+)
+
+
+class TestMiniMaxM2(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify that the inference accuracy of the cyankiwi/MiniMax-M2-BF16 model on the GSM8K dataset is no less than 0.9.
+
+    [Test Category] Model
+    [Test Target] cyankiwi/MiniMax-M2-BF16
+    """
+
+    model = MINIMAX_M2_WEIGHTS_PATH
+    accuracy = 0.9
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        "0.9",
+        "--attention-backend",
+        "ascend",
+        "--tp-size",
+        "8",
+        "--disable-cuda-graph",
+        "--disable-radix-cache",
+        "--disable-overlap-schedule",
+        "--max-running-requests",
+        "64",
+        "--chunked-prefill-size",
+        "-1",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()