Jan Wassenberg comments

Results 633 comments of


                                            Jan Wassenberg

Added support for dynamic dispatch for macOS/iOS/iPadOS on AArch64

Thanks for rebasing! Unfortunately internal CI is failing with a compiler crash for convert_test, but possibly also for all other targets. ``` clang: error: clang frontend command failed due to...

Added support for dynamic dispatch for macOS/iOS/iPadOS on AArch64

Indeed still breaking. I am patching the PR to disable runtime dispatch for Clang < 17, and fix the detection - Apple Clang 15.3 was mistakenly found to be 17...

Added support for dynamic dispatch for macOS/iOS/iPadOS on AArch64

I have patched this PR as mentioned :)

Different test results using Clang when enabling Debug or not on targer RVV

Thanks for reporting. We have also seen issues with rounding mode on QEMU - is that how you are running the tests, or is it on real HW? > Debug,...

Different test results using Clang when enabling Debug or not on targer RVV

I think the issue is solved, thanks @johnplatts :)

HWY_EXPORT_AND_DYNAMIC_DISPATCH_T for multiple template parameters function with Clang/GCC

Hi, we do document that this is intended for a single template argument. At a minimum, highway.h would probably have to wrap all FUNC_NAME in something like SINGLE_ARG - want...

HWY_EXPORT_AND_DYNAMIC_DISPATCH_T for multiple template parameters function with Clang/GCC

Makes sense. Thanks for confirming! Supporting a comma in macro args does seem to be a higher-risk (of breaking across compiler versions) thing that I'd prefer to avoid if we...

HWY_EXPORT_AND_DYNAMIC_DISPATCH_T for multiple template parameters function with Clang/GCC

Yes, it's mainly MSVC that differs in its macro/comma parsing. I understand this would be useful given you have many functions. If you can verify your patch works in MSVC...

Gemma.cpp hangs on a Gemma 7B model that was finetuned using huggingface peft(QLoRA)

Hi, thanks for reaching out :) There are two likely causes. One is that Gemma is trained for a format like "user\n", but because the command line specifies 7b-pt instead...

Gemma.cpp hangs on a Gemma 7B model that was finetuned using huggingface peft(QLoRA)

Thank you for sharing. So we indeed want PT and your model is trained for the prompt format used. Also sounds like there is no range overrun for the compressed...