[Meta_schedule] InternalError: Check failed: original_producers.size() == 1u (0 vs. 1):

Troubleshooting
#1

Hi there,

I am trying to perform metascheduling on a resnet50 workload, and the focus is to run it on cuda tensorcores. The workload specification code is given below:

   mod, params = testing.resnet.get_workload(
   batch_size = batch_size, image_shape=input_image_shape,layout="NHCW",dtype="float16")

The parameter specifications are :

batch_size = 1
num_class = 1000
input_image_shape = (224,224,3)
data_shape = (batch_size,) + input_image_shape
output_shape = (batch_size, num_class)
dtype = "float16"

However, when I begin the metascheduling, I encounter this error:

image
image1682×510 40 KB

I checked this issue but it has been stagnant for a while now. No contribution has been made to it either. This error however, doesn’t appear when I use the workload with NCHW layout and float32 datatype. The tuning happens as usual.

What is the reason behind this error for resnet with this specific NHWC format? Please help me solve this error. I have put up the full code here : resnet_ms

@zxybazh @junrushao Kindly share your thoughts on what might be going wrong here.

#2

Hello,

I did some further probing into this issue and found that this error occurs only when the data type is specified as float16. I tried meta-scheduling with float 32 for different models (with both NCHW and NHWC layouts) and the tuning seemed to carry on as usual.

However, when the datatype is specified as float16 or when a ToMixedPrecision Pass is used to convert the relay module to float16 from float32, this error occurs. Hope this helps.

@comaniac @zxybazh @junrushao