Eugene Chereshnev
Eugene Chereshnev
> That's why I would prefer a flexible API for frameworks to pass all the known information, e.g., able to mark individual dim as dynamic and provide size hint for...
> @karturov, > > > Graph mode: guarantee that you set a hint only for really changing dimensions, so, for example, only relevant Hifagan layers are affected, but SD/ResNet-50 layers...
@petercad Do we need a similar PR for gemmstone repo?
> The result has an ~3% bias of rounding up, likely related to the distribution of bits generated by the Philox engine, or the seeds generated by the Mersenne Twister...
> Because benchdnn also uses the stream_profiler functionality in perf mode, the verbose profiler is disabled when benchdnn is run in perf mode to record aggregate timing info. There is...
> I have adjusted the design to removing the resetting action which allows it to be used with the benchdnn perf mode and the experimental profiler. @avmanerikar Thanks for the...
> This would be a concern if the profiler is reset by the experimental API while the verbose profiler is yet to print the timing data. But since the verbose...
@ZackyLake Thanks for the PR! Do you observe the related overhead in real workloads? I agree with your comment - we can fail much earlier in case of insufficient SLM....