nccl not support for float16?

[HuaYZhao]

Description

    inputs = flow.ones([4, 10],dtype=flow.float32, placement=placement, sbp=flow.sbp.broadcast)
    weight = flow.ones([10, 4],dtype=flow.float32, placement=placement, sbp=flow.sbp.split(1)) * 2
    y = flow.matmul(inputs, weight, )
    y.to_global(placement, flow.sbp.broadcast).to_local()

this is ok, but when i change the dtype to float16, just as fllow:

    inputs = flow.ones([4, 10],dtype=flow.float16, placement=placement, sbp=flow.sbp.broadcast)
    weight = flow.ones([10, 4],dtype=flow.float16, placement=placement, sbp=flow.sbp.split(1)) * 2
    y = flow.matmul(inputs, weight, )
    y.to_global(placement, flow.sbp.broadcast).to_local()

this is wrong, the error is :

RuntimeError: InferDataType Failed. Expected kFloat, but got kFloat16
  File "oneflow/core/functional/impl/nn_functor.cpp", line 359, in operator()
    OpInterpUtil::Dispatch<Tensor>(*matmul_op_, {cast_a, cast_b}, attrs)
  File "oneflow/core/framework/op_interpreter/op_interpreter_util.cpp", line 144, in Dispatch<oneflow::one::Tensor>
    Dispatch<TensorTuple>(op_expr, inputs, ctx)
  File "oneflow/core/framework/op_interpreter/op_interpreter_util.cpp", line 135, in Dispatch<oneflow::one::TensorTuple>
    Dispatch(op_expr, inputs, outputs.get(), ctx)
  File "oneflow/core/framework/op_interpreter/op_interpreter.cpp", line 103, in Apply
    internal_->Apply(op_expr, inputs, outputs, ctx)
  File "oneflow/core/framework/op_interpreter/eager_global_op_interpreter.cpp", line 219, in Interpret
    user_op_expr.mut_global_tensor_infer_cache()->GetOrInfer(*infer_args)
  File "oneflow/core/framework/global_tensor_infer_cache.cpp", line 367, in GetOrInfer
    Infer(*user_op_expr, infer_args)
  File "oneflow/core/framework/global_tensor_infer_cache.cpp", line 274, in Infer
    user_op_expr.InferLogicalTensorDesc( infer_args.attrs ... return output_mut_metas.at(i).mut_tensor_meta(); })
  File "oneflow/core/framework/op_expr.cpp", line 612, in InferLogicalTensorDesc
    dtype_infer_fn_(&infer_ctx)
  File "oneflow/user/ops/matmul_op.cpp", line 69, in InferDataType4Matmul
    CHECK_EQ_OR_RETURN(ctx->InputDType("b", 0), dtype)
Error Type: oneflow.ErrorProto.check_failed_error

please help me !

[clackhan]

I cannot reproduce your problem. Can you print dtype of input and weight before doing matmul to make sure they are the same? If the dtype of input and weight are the same and the problem still occurs, please provide the version of oneflow, thanks.