Enabled Qwen2-MoE Tensor Parallelism (TP) inference (#6551)

Modified _replace_module in auto_tp.py :
The modification keeps the layers 'shared_expert_gate' and 'gate' in
qwen2-moe the original type torch.nn.Linear and not changes them into
LinearLayer. In this way, their weights will not be split into multiple
HPU/GPU cards. Then the qwen2-moe can run on multiple HPU/GPU cards.
Since the weights of 'gate' are not split into multiple HPU/GPU cards,
all gather operations are not needed, which may improve performance.

---------

Co-authored-by: Logan Adams <114770087+loadams@users.noreply.github.com>

deepspeed/module_inject/auto_tp.py

@@ -333,7 +333,8 @@ def _replace(self, child, name, conv_linear_layer):
         weight_shape = child.weight.shape
         mp_replace = ReplaceWithTensorSlicing(mp_group=self.mp_group)
         # For mixtral-7x8b, need to skip MoE gate linear replace.
-        if name == "block_sparse_moe.gate":
+        if name == "block_sparse_moe.gate" or (('mlp.shared_expert_gate' == name or 'mlp.gate' == name)
+                                               and 'qwen2_moe' in str(type(self.module))):
             return child
         # For Yuan model
         if 'Yuan' in str(self.module):

docs/_tutorials/automatic-tensor-parallelism.md

@@ -158,6 +158,7 @@ The following model families have been successfully tested with automatic tensor
 - plbart
 - qwen
 - qwen2
+- qwen2-moe
 - reformer
 - roberta
 - roformer