leezu commented on issue #18699:
URL:
https://github.com/apache/incubator-mxnet/issues/18699#issuecomment-664643062
@fhieber Sorry for being unclear. I'm referring to the 5 seconds overhead
that remains after reverting `parameter.py`. I think it may be introduced by
the extra C API call
leezu commented on issue #18699:
URL:
https://github.com/apache/incubator-mxnet/issues/18699#issuecomment-663711968
@mjdenkowski when Sockeye adopts MXNet 2, you will need to rename from
`hybrid_forward(F, ...)` to `forward(...)`. The overhead you currently
experience could be from the
leezu commented on issue #18699:
URL:
https://github.com/apache/incubator-mxnet/issues/18699#issuecomment-663621153
Thank you for splitting up the commit and comparing the frontend to backend
changes! To find out which part of the Python frontend changes is causing the
overhead, I'd
leezu commented on issue #18699:
URL:
https://github.com/apache/incubator-mxnet/issues/18699#issuecomment-661242552
@mjdenkowski
> This benchmark was run using a trained Sockeye model. This is a full
sequence-to-sequence model with HybridBlocks glued together by Python code.
leezu commented on issue #18699:
URL:
https://github.com/apache/incubator-mxnet/issues/18699#issuecomment-657744237
Can you share the benchmark script and point out where it uses the
HybridBlock? How many forward / backward operations are run by the script? We
need to identify if the