[NPU] bugfix with Kimi-k2 and bge-reranker-v2 model (#17478)

Co-authored-by: amote-i <49533125+amote-i@users.noreply.github.com>
Co-authored-by: cy <chenyang08056032@163.com>

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1648,7 +1648,7 @@ class NPUCompressedTensorsW4A16Int4DynamicMoEMethod(CompressedTensorsMoEMethod):
         self.num_experts = num_experts
         if (
             extra_weight_attrs.get(
-                "intermediate_size_full", intermediate_size_per_partition
+                "moe_intermediate_size", intermediate_size_per_partition
             )
             // intermediate_size_per_partition
             > 1

python/sglang/srt/layers/rotary_embedding.py

@@ -115,9 +115,10 @@ class RotaryEmbedding(MultiPlatformOp):
             cache = cache.to(dtype)
 
         if (
-            (not (_is_cuda or _is_npu) or self.head_size not in [64, 128, 256, 512])
+            (not (_is_cuda) or self.head_size not in [64, 128, 256, 512])
             and not (_is_cpu)
             and not (_is_xpu)
+            and not (_is_npu)
         ):
             if _is_cuda or _is_hip:
                 from sgl_kernel import rotary_embedding