Yavor Ivanov

I think [conv_2d](https://github.com/ggerganov/ggml/blob/08c57df1b98ff94d065c0ee2f42294a5bde6bb7b/src/ggml.c#L6936) has been finished and its interface [generalized](https://github.com/ggerganov/ggml/pull/313). Further optimizations of the operations will be done in a separate issue.

I am having this exact same issue on Ubuntu 22.04 with GCC 10.3. I ran AddressSanitizer and this is what I got: ``` ==2807805==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x6070000016b8 at...

Here is what I did to be as explicit as possible and the issue was still reproduced (the address sanitizer complained). If I start it in Debug without sanitizer it...

I tested using GCC 11.3.0 & GCC 12.1.0 and I am unable to reproduce the issue with both of the compilers. However, I do reproduce it with GCC 10.3 as...

I was trying to find a way to circumvent the issue and in isolated example and tried just doing random stuff as adding a key to a std::unordered_map similar to...

I am unable to determine anything wrong currently. I tried purging GCC 10.3 and building with GCC 10.4 and I do not reproduce the issue. May have been somehow related...

> After implementing the kernel with simdgroup matrix ops, it is now universally better than the `master` version. Tested with head size 128 - might needs some extra work for...

In order to copy and persist the current key and values (Kcur and Vcur) to the kv cache.

This function seems to assume that the src0 is contigous, so you can do ggml_cont(..) on the first argument of the add you do.

The MLA approach can probably be combined with the Pyramid KV cache - https://arxiv.org/abs/2405.12532