masahi edited a comment on pull request #5755: URL: https://github.com/apache/incubator-tvm/pull/5755#issuecomment-642983173
ok I now understand why it makes importing some model slower. Previously there are fewer but heavy `infer_value` calls, now there are many but they are all single node compile, so LLVM overhead dominates. I wish there were a "fast-path" compile option for llvm for use case like this. Turning on all optimization to compile a single node and get a compile time constant is clearly overkill. At least we can lower the opt level (always 3 in our backend) for const eval mode @tqchen Maybe we can add some option here to tell llvm that we want a fast path compile https://github.com/apache/incubator-tvm/blob/master/python/tvm/relay/frontend/common.py#L509 ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org
