[NPU] enhance accuracy for model kimi-vl-a3b-instruct (#17480)

Co-authored-by: cy <chenyang08056032@163.com>

python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py

@@ -103,6 +103,10 @@ def forward_mha_prepare_npu(
         k_pe = latent_cache[:, :, m.kv_lora_rank :]
         if m.rotary_emb is not None:
             q_pe, k_pe = m.rotary_emb(positions, q_pe, k_pe)
+        # this is for model kimi-vl-a3B-instruct
+        forward_batch.token_to_kv_pool.set_kv_buffer(
+            m, forward_batch.out_cache_loc, kv_a.unsqueeze(1), k_pe
+        )
 
     q[..., m.qk_nope_head_dim :] = q_pe
 

test/registered/ascend/vlm_models/test_ascend_kimi_vl_a3b_instruct.py

@@ -0,0 +1,28 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+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-4-npu-a3", nightly=True)
+
+
+class TestKimiVLA3BInstruct(GSM8KAscendMixin, CustomTestCase):
+    model = "/root/.cache/modelscope/hub/models/Kimi/Kimi-VL-A3B-Instruct"
+    accuracy = 0.66
+    other_args = [
+        "--trust-remote-code",
+        "--max-running-requests",
+        2048,
+        "--mem-fraction-static",
+        0.7,
+        "--attention-backend",
+        "ascend",
+        "--tp-size",
+        "4",
+        "--disable-cuda-graph",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()