From 71e9b379c84d4d68764fcda4ef833e21884f8341 Mon Sep 17 00:00:00 2001
From: zhangxinyuehfad <59153331+zhangxinyuehfad@users.noreply.github.com>
Date: Tue, 2 Dec 2025 09:20:05 +0800
Subject: [PATCH] [Bugfix] Fix Qwen2.5-Omni-7B accuarcy test (#4556)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

### What this PR does / why we need it?
Fix Qwen2.5-Omni-7B accuarcy test
issue：https://github.com/vllm-project/vllm-ascend/issues/4480
Depends on : https://github.com/vllm-project/vllm-ascend/pull/4534

- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

Signed-off-by: hfadzxy <starmoon_zhang@163.com>
---
 vllm_ascend/ops/layernorm.py           | 6 +++---
 vllm_ascend/ops/register_custom_ops.py | 8 ++++++--
 2 files changed, 9 insertions(+), 5 deletions(-)

diff --git a/vllm_ascend/ops/layernorm.py b/vllm_ascend/ops/layernorm.py
index da5051c0..8dad11c2 100644
--- a/vllm_ascend/ops/layernorm.py
+++ b/vllm_ascend/ops/layernorm.py
@@ -108,13 +108,13 @@ class AscendRMSNorm(RMSNorm):
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
         import torch_npu
-
         if residual is not None:
             residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
             assert x.size(0) == residual.size(0)
+            next_need_quant_fusion_linear = getattr(
+                self, 'next_need_quant_fusion_linear', None)
             x, residual = _addrmsnorm_forward_oot(
-                self, x, residual, self.next_need_quant_fusion_linear,
-                self.bias)
+                self, x, residual, next_need_quant_fusion_linear, self.bias)
             return x, residual
         x, residual = torch_npu.npu_rms_norm(x, self.weight,
                                              self.variance_epsilon)
diff --git a/vllm_ascend/ops/register_custom_ops.py b/vllm_ascend/ops/register_custom_ops.py
index 03bea554..7c7fd6f0 100644
--- a/vllm_ascend/ops/register_custom_ops.py
+++ b/vllm_ascend/ops/register_custom_ops.py
@@ -173,7 +173,9 @@ def _maybe_prefetch_mlp_down_proj_impl(x_dependency: torch.Tensor) -> None:
     except AssertionError:
         return
 
-    if not forward_context.prefetch_mlp_enabled:
+    prefetch_mlp_enabled = getattr(forward_context, 'prefetch_mlp_enabled',
+                                   False)
+    if not prefetch_mlp_enabled:
         return
     forward_context.prefetch_mlp_down_proj = True
     model_instance = forward_context.model_instance
@@ -202,7 +204,9 @@ def _maybe_wait_prefetch_done_impl(x: torch.Tensor) -> None:
     except AssertionError:
         return
 
-    if not forward_context.prefetch_mlp_enabled:
+    prefetch_mlp_enabled = getattr(forward_context, 'prefetch_mlp_enabled',
+                                   False)
+    if not prefetch_mlp_enabled:
         return
     if forward_context.prefetch_mlp_gate_up_proj or \
         forward_context.prefetch_mlp_down_proj: