Add pipeline parallel support to TransformersModel (#12832)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <2037008807@qq.com>

vllm/model_executor/models/utils.py

@@ -472,6 +472,16 @@ class PPMissingLayer(torch.nn.Identity):
 
     def __init__(self, *args, **kwargs):
         super().__init__()
+        self.return_tuple = kwargs.get("return_tuple", False)
+
+    def forward(self, *args, **kwargs):
+        """
+        Return the first arg from args or the first value from kwargs.
+
+        Wraps the input in a tuple if `self.return_tuple` is True.
+        """
+        input = args[0] if args else next(iter(kwargs.values()))
+        return (input, ) if self.return_tuple else input
 
 
 _CPU_OFFLOAD_BYTES = 0
@@ -650,4 +660,4 @@ def cast_overflow_tensors(
     if tensors.isinf().any() or tensors.isnan().any():
         clamp_value = torch.finfo(tensors.dtype).max - offset
         tensors = torch.clamp(tensors, min=-clamp_value, max=clamp_value)
-    return tensors
+    return tensors