Xiaodong (Vincent) Huang
Xiaodong (Vincent) Huang
This dynamic weights feature will be supported in next release. thanks!
Closing since no response for more than 3 weeks, please reopen if you still have question, thanks!
@matudouchen , The could not find implementation issue might because some old version trt missing conv + swish kernel implementation. And fixed in newer version TRT. The error in the...
Closing since no response for more than 3 weeks, please reopen if you still have question, thanks!
@pangr , what's the op after the `add`, also have you tried insert Q/DQ after the `add`? thanks
Hello @zhaohb , this `copyPackedKernel` come from the `reformat` between layers that has different data layout, we have some dedicated kernels for `reformat`, and will fallback to this `genericReformat::copyPackedKernel` if...
@zhaohb , do you also have the verbose build log by enable `TRT_LOGGER = trt.Logger(trt.Logger.VERBOSE)`? thanks
Hello @zhaohb , I need the log to figure out why reformat needed, but I cannot access the log, and have request the access to it.
Hello @zhaohb , just checked those are not really reformat, they are `slice` implementation. And slice is memory bound, to optimize this we need consider other approaches. I have two...
Hello @zhaohb , Enable verbose log by `TRT_LOGGER = trt.Logger(trt.Logger.VERBOSE).` can tell us where the `CopyPackedKernel` come from. I have a question, How did you implement `slice` without copy? for...