You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I evaluate Profile-Guided Optimization (PGO) applicability to different kinds of software - all my results are available in my repo. From my experience, PGO helps with achieving better performance in many scenarios.
Recently I performed PGO tests for HuggingFace Tokenizer project - the results are located here. Since the results are quite promising (up to 20% performance improvements in some scenarios), I think it could be interesting to perform the same PGO benchmarks for SentencePiece. As far as I understand from the technical highlights - performance is one of the goals of the project.
Did anyone try before to optimize SentencePiece performance with PGO? If yes, could you please share the benchmark results? If no - is there an established methodology/make command/anything else to perform such benchmarks?
Thanks in advance.
The text was updated successfully, but these errors were encountered:
Hi!
I evaluate Profile-Guided Optimization (PGO) applicability to different kinds of software - all my results are available in my repo. From my experience, PGO helps with achieving better performance in many scenarios.
Recently I performed PGO tests for HuggingFace Tokenizer project - the results are located here. Since the results are quite promising (up to 20% performance improvements in some scenarios), I think it could be interesting to perform the same PGO benchmarks for SentencePiece. As far as I understand from the technical highlights - performance is one of the goals of the project.
Did anyone try before to optimize SentencePiece performance with PGO? If yes, could you please share the benchmark results? If no - is there an established methodology/make command/anything else to perform such benchmarks?
Thanks in advance.
The text was updated successfully, but these errors were encountered: