Merge branch 'v0.9.2-dev-custom' into 'v0.9.2-dev'

解决custom allreduce在K100AI上新模型报错问题

See merge request dcutoolkit/deeplearing/vllm!461

csrc/custom_all_reduce.cuh

@@ -1062,7 +1062,7 @@ class CustomAllreduce {
 #define KL(ngpus, name)                                                       \
   {                                                                           \
     void* kernelArgs[] = {                                                    \
-      &ptrs, &sg_, &self_sg_, &output, &rank_, &size                          \
+      &ptrs, &sg_, &self_sg_, &output, &rank_, &size, &dev_curr_hdp_reg, &world_size_                         \
     };                                                                        \
     hipExtLaunchKernel(                                                       \
       (void*)name<T, ngpus>,                                                  \