[0.11.0]Chery pick pta upgrade change (#3940)

This PR cherry-pick two commit from main to upgrade torch-npu to 2.7.1
official release

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/ops/layernorm.py

@@ -22,8 +22,6 @@ from vllm.config import get_current_vllm_config
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm
 
-from vllm_ascend.utils import version_check
-
 
 def _addrmsnorm_forward_oot(
     self,
@@ -36,7 +34,6 @@ def _addrmsnorm_forward_oot(
 
     from vllm_ascend.utils import is_310p
 
-    torch_npu_check = version_check()
     if layer is not None and not is_310p():
         layer_cls_name = layer.__class__.__name__
         try:
@@ -53,23 +50,15 @@ def _addrmsnorm_forward_oot(
                 start_flag=x,
             )
         # add_rms_norm_quant
-        if torch_npu_check:
-            x, _, residual = torch_npu.npu_add_rms_norm_quant(
-                x,
-                residual,
-                self.weight,
-                layer.aclnn_input_scale,
-                layer.aclnn_input_offset,
-                beta=bias,
-                epsilon=self.variance_epsilon)
-        else:
-            x, _, residual = torch_npu.npu_add_rms_norm_quant(
-                x,
-                residual,
-                self.weight,
-                layer.aclnn_input_scale,
-                layer.aclnn_input_offset,
-                epsilon=self.variance_epsilon)
+        x, _, residual = torch_npu.npu_add_rms_norm_quant(
+            x,
+            residual,
+            self.weight,
+            layer.aclnn_input_scale,
+            layer.aclnn_input_offset,
+            beta=bias,
+            epsilon=self.variance_epsilon)
+
         # prefetch qkvo_proj.weight postprocess
         if weight_prefetch_method:
             weight_prefetch_method.maybe_prefetch_attn_weight_postprocess(
@@ -87,7 +76,7 @@ def _addrmsnorm_forward_oot(
         else:
             x, _, residual = torch_npu.npu_add_rms_norm(
                 x, residual, self.weight, self.variance_epsilon)
-        if torch_npu_check and bias is not None:
+        if bias is not None:
             x.add_(bias)
     torch.ops.vllm.maybe_wait_prefetch_done(x)
     return x, residual
@@ -106,9 +95,8 @@ class AscendRMSNorm(RMSNorm):
         super().__init__(hidden_size, eps, var_hidden_size, has_weight, 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)
@@ -128,7 +116,7 @@ class AscendRMSNorm(RMSNorm):
             return x, residual
         x, residual = torch_npu.npu_rms_norm(x, self.weight,
                                              self.variance_epsilon)
-        if self.torch_npu_check and self.bias is not None:
+        if self.bias is not None:
             x.add_(self.bias)
         return x
 

vllm_ascend/ops/weight_prefetch.py

@@ -7,7 +7,6 @@ from vllm.forward_context import get_forward_context
 from vllm_ascend.ascend_config import WeightPrefetchConfig
 from vllm_ascend.ops.linear import (AscendQKVParallelLinear,
                                     AscendRowParallelLinear)
-from vllm_ascend.utils import version_check
 
 SUPPORTED_MODULES = ["attn", "mlp", "moe"]
 MOE_PREFETCH_TOKEN_THRESHOLD = 96
@@ -83,8 +82,7 @@ class WeightPrefetchMethod:
         if not self.moe.is_active_this_forward:
             return
         forward_context = get_forward_context()
-        if not version_check():
-            forward_context.layer_idx += 1
+        # layer_idx is subtracted by 1 because layer_idx was incremented by 1 at layernorm.
         weight = forward_context.model_instance.model.layers[
             forward_context.layer_idx - 1].mlp.experts.w13_weight
         weight_size = weight.data.element_size() * weight.data.numel(