[310p]: add rmsnorm gated fallback and unit test (#7424)

### What this PR does / why we need it?
RFC #7394
310P cannot use the fused `rmsnormgated` operator and must fall back to
the native implementation.

### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
ut
- vLLM version: v0.17.0
- vLLM main:
4497431df6

---------

Signed-off-by: Tflowers-0129 <2906339855@qq.com>

vllm_ascend/_310p/ops/layernorm.py

@@ -1,5 +1,6 @@
 import torch
 import torch_npu
+from vllm.model_executor.layers.layernorm import RMSNormGated
 
 from vllm_ascend.ops.layernorm import AscendGemmaRMSNorm, AscendRMSNorm
 
@@ -37,3 +38,14 @@ class AscendGemmaRMSNorm310(AscendGemmaRMSNorm):
 
         x, _ = torch_npu.npu_rms_norm(x, 1.0 + self.weight, self.variance_epsilon)
         return x
+
+
+class AscendRMSNormGated310(RMSNormGated):
+    def forward_oot(
+        self,
+        x: torch.Tensor,
+        z: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        # 310P should not depend on the Triton-gated layernorm path.
+        # Reuse the upstream native implementation directly.
+        return super().forward_native(x, z)

vllm_ascend/utils.py

@@ -661,7 +661,11 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
     if is_310p():
         from vllm_ascend._310p.fused_moe.fused_moe import AscendFusedMoE310, AscendSharedFusedMoE310
         from vllm_ascend._310p.ops.activation import AscendSiluAndMul310
-        from vllm_ascend._310p.ops.layernorm import AscendGemmaRMSNorm310, AscendRMSNorm310
+        from vllm_ascend._310p.ops.layernorm import (
+            AscendGemmaRMSNorm310,
+            AscendRMSNorm310,
+            AscendRMSNormGated310,
+        )
         from vllm_ascend._310p.ops.mm_encoder_attention import AscendMMEncoderAttention310
         from vllm_ascend._310p.ops.rotary_embedding import AscendRotaryEmbedding310
         from vllm_ascend._310p.ops.vocab_parallel_embedding import (
@@ -675,6 +679,7 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
                 "RotaryEmbedding": AscendRotaryEmbedding310,
                 "RMSNorm": AscendRMSNorm310,
                 "GemmaRMSNorm": AscendGemmaRMSNorm310,
+                "RMSNormGated": AscendRMSNormGated310,
                 "FusedMoE": AscendFusedMoE310,
                 "SharedFusedMoE": AscendSharedFusedMoE310,
                 "ParallelLMHead": AscendParallelLMHead310,