[NPU][Bugfix] Fix qwen3 error when enable-dp-lm-head (#16115)

python/sglang/srt/layers/rotary_embedding.py

@@ -294,8 +294,8 @@ class RotaryEmbedding(MultiPlatformOp):
         assert (
             fused_set_kv_buffer_arg is None
         ), "fused_set_kv_buffer_arg is not supported for npu implementation"
-
-        rotary_mode = "half"
+        if query.dtype == torch.bfloat16 and self.cos_sin_cache.dtype == torch.float:
+            return self.forward_native(positions, query, key, offsets)
         if self.is_neox_style:
             rotary_mode = "half"
         else:

python/sglang/srt/layers/vocab_parallel_embedding.py

@@ -31,6 +31,7 @@ from sglang.srt.utils import (
     cpu_has_amx_support,
     get_compiler_backend,
     is_cpu,
+    is_npu,
     set_weight_attrs,
 )
 
@@ -38,6 +39,7 @@ DEFAULT_VOCAB_PADDING_SIZE = 64
 
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
+_is_npu = is_npu()
 
 logger = logging.getLogger(__name__)
 
@@ -123,7 +125,7 @@ class VocabParallelEmbeddingShardIndices:
         assert self.num_added_elements <= self.num_added_elements_padded
 
 
-@torch.compile(dynamic=True, backend=get_compiler_backend())
+@torch.compile(dynamic=True, backend=get_compiler_backend(), disable=_is_npu)
 def get_masked_input_and_mask(
     input_: torch.Tensor,
     org_vocab_start_index: int,