Question about running INT4 example

Thanks for your excellent work! However, I got some problems running this example
1) Do I have to install TensorRT in order to run this example? If so, where can I find a guide to install it (e.g. what version of TensorRT should I use)?
2) Can I running this INT4 example on my RTX3090 GPU? or it has to be A2?
3) When I running this example, I got an error of
File "/home/hyf/anaconda3/lib/python3.9/site-packages/megengine/utils/comp_graph_tools.py", line 499, in run
self._func.execute()
RuntimeError: var sanity check failed: var: {id:516, layout:{512(2048),32(64),1(64),1(64),64(1) @ LOWBITS{4,8} QuantizedS4}, QuantizedS4, owner:fuse():dimshuffle(reshape[3158])[3160]@3161{ImmutableTensor}, name:fuse():dimshuffle(reshape[3158])[3160]@3161, slot:0, CompNode("gpu0:0" from "xpux:0"), s, 2, 2} (checksum: expect={checksum:0x2d819168, last_int:1412490702, last_float:3.03908e+12} got={checksum:0x9b352857, last_int:1412490702, last_float:3.03908e+12}); receiver: conv_bias(conv_bias[4275],fuse[3160]@3161,const{1,8,1,1,64}[3182])[4282]{ConvBiasForward}(864); you can set MGB_DEBUG_VAR_SANITY_CHECK_LOG=516 to get more details; pass=0
How can I fix this?

Looking forward to your reply!!

The issue provides no repository file, test, or entry point to start from; begin by locating the INT4 example and reproducing the reported MegEngine quantized-tensor error. Check the required TensorRT and GPU prerequisites, then document a confirmed compatible setup and resolution for the sanity-check failure.

Written by the indexing model from the issue text.