[SOLVED] TVMError: Check failed: !use_count_.count(v)

Troubleshooting
#1

Hi everyone,

I’m currently getting the following error when trying to compile my model with target tvm.target.cuda():

TVMError: Check failed: !use_count_.count(v): variable dh has been used before definition!
Error during compile function

It’s due a conv2d_transpose layer:

nn.conv2d_transpose(%p0, %p1, channels=256, kernel_size=[4, 4], strides=[2, 2], padding=[1, 1])

I found this reporting the same error but the case seems different.

Any idea?

#2

Anyone can give some insight? Is it a bug?

@tqchen @yongwww @kevinthesun @yzhliu @ziheng @umangyadav @optima2005 @huyuwei

#3

Not sure if it is the same issue. To do a quick check, could you add back the workaround in you local code of this file “topi/python/topi/cuda/conv2d_transpose_nchw.py”. To see if it is still needed in your case.

def schedule_conv2d_transpose_nchw_cuda(cfg, outs):
    """TOPI Schedule callback for conv2d transpose operator.
   ... ...
    def _callback(op):
        if op.tag == 'conv2d_transpose_nchw':
        ... ...


        if cfg.is_fallback:
            ko = int(kernel.shape[1])
            kh = int(kernel.shape[2])
            kw = int(kernel.shape[3])
            stride_h, stride_w = cfg.stride
            # Workaround to make CUDA compilation work. Issue #4470
            # TODO make _fallback_schedule work for all kernel/strides combinations
            #  after issue #4470 is resolved
            do_fallback = True
            if ko == 1:
                do_fallback = False
            elif (kh, kw) == (1, 1):
                do_fallback = True
            elif (stride_h, stride_w) == (2, 2):
                do_fallback = False
            elif (kh, kw) == (stride_h, stride_w):
                do_fallback = False

            if do_fallback:
                N, F, Y, X = get_const_tuple(conv.shape)
                _fallback_schedule(N, F, Y, X)
1 Like
#4

Thanks @optima2005! The workaround fixed the issue. We need to implement a stable fix for this issue though.