Xiaohan Ding

icon indicating an email [email protected]

icon location Shenzhen, China icon location ML, CV, Multimodal LLM, PhD from Tsinghua University

Results 1 issues of Xiaohan Ding

Update depthwise_conv2d_implicit_gemm.py

1 comment

Sometimes when the tensor format changes after this conv (e.g., NCHW -> NHWC for layer normalization), calling backward will raise an "input must be contiguous" error. Making the grad contiguous...