From d0086d432ac0f40ce6ea37e802c2fb364cc43481 Mon Sep 17 00:00:00 2001
From: Wang Yixuan <88923622+hust17yixuan@users.noreply.github.com>
Date: Thu, 23 Oct 2025 22:53:13 +0800
Subject: [PATCH] fix deepseek torchair recompile (#3679)

### What this PR does / why we need it?
The #3624 PR fix the precision of deepseek torchair, but don't consider
the limitation of torch compile which results in the recompile, This PR
fixs this problem. PR to main #3678


### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->

Signed-off-by: hust17yixuan <303660421@qq.com>
---
 vllm_ascend/torchair/ops/torchair_layernorm.py | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

diff --git a/vllm_ascend/torchair/ops/torchair_layernorm.py b/vllm_ascend/torchair/ops/torchair_layernorm.py
index 2f52dab..583a376 100644
--- a/vllm_ascend/torchair/ops/torchair_layernorm.py
+++ b/vllm_ascend/torchair/ops/torchair_layernorm.py
@@ -21,8 +21,6 @@ import torch
 from vllm.config import get_current_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 
-from vllm_ascend.utils import version_check
-
 _original_re_init = RMSNorm.__init__
 
 
@@ -38,9 +36,8 @@ def torchair_rmsnorm_init_(
                       dtype)
     vllm_config = get_current_vllm_config()
     self.bias = None
-    self.torch_npu_check = version_check()
     # quantization with anti_method m4 will generate none-zero norm bias
-    if self.torch_npu_check and vllm_config.quant_config is not None and \
+    if vllm_config.quant_config is not None and \
             any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
         self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
                                        requires_grad=False)
@@ -59,7 +56,6 @@ def torchair_rmsnorm_forward_oot(
     """
 
     import torch_npu
-    torch_npu_check = version_check()
 
     from vllm_ascend.utils import is_310p
     if residual is not None:
@@ -72,11 +68,11 @@ def torchair_rmsnorm_forward_oot(
         else:
             x, _, residual = torch_npu.npu_add_rms_norm(
                 x, residual, self.weight, self.variance_epsilon)
-        if torch_npu_check and self.bias is not None:
+        if self.bias is not None:
             x.add_(self.bias)
         return x, residual
 
     x, residual = torch_npu.npu_rms_norm(x, self.weight, self.variance_epsilon)
-    if torch_npu_check and self.bias is not None:
+    if self.bias is not None:
         x.add_(self.bias)
     return x