compile_model throws error about profiling

[dashesy]

When I compile a model that had nn.Conv2dBiasFewChannels I get this error:

<aitemplate.compiler.ops.conv.conv2d> Profile: conv2d_bias_few_channels_1: NI == 1 && HI == 384 && WI == 384 && CI == 4
<aitemplate.backend.profiler_runner> Using 1 GPU for profiling conv2d_bias_few_channels_1

RuntimeError: Profiler ./output/profiler/conv2d_bias_few_channels/cutlass_f16_s884fprop_fixed_channels_f16_256x128_32x3_nhwc_align_4_8 is not executable

This was confusing because I did not start any profiling, just did compile_model. Until I read that for some ops codegen require profiling. Is it to find the most optimized path? It is unusual for a compiler to run part of the code, conceptually.

1. My first question is this, which ops require profiling for codegen?

• Here are some that seem to create a profiler during compile: conv2d_bias_few_channels, conv2d_bias_add_identity, gemm_rcr_bias, bmm_crr, bmm_rcr, bmm_ccr_add, gemm_rcr_bias_fast_gelu, gemm_rcr_bias_add, bmm_rrr, conv2d_bias,

3. And my second question is why did this fail? Could it be that my V100 machine does not support that op? All other ops mentioned above have the executable built but this one. When I use nn.Conv2dBias instead, I get another issue "a/b is not aligned" which I am now looking into