goerch

I'd tend to agree with @Dampfinchen : being interested in the Intel platform I don't believe we'll be able to outperform the MKL and oneAPI engineers.

Is `ggml_graph_compute_thread` covered by the tests? How should I check the merged version otherwise?

@GoWind: I have been thinking about this one, too. Maybe @ggerganov means something along the lines of ``` struct test { int ne[4]; }; #define GGML_LOCALS_1(type, prefix, pointer, array) \...

@ggerganov : possible simplification: ``` struct test { int ne[4]; }; #define GGML_LOCALS_1(prefix, pointer, array) \ const auto prefix##0 = (pointer)->array[0]; #define GGML_LOCALS_2(prefix, pointer, array) \ GGML_LOCALS_1(prefix, pointer, array) \...

Now that #309 is merged: does it make sense if I adapt other accelerators without being able to test them (maybe I get OpenCL working)? Does CI cover enough of...

I'd be interested in helping with the 'add more tests' part of this because of some unanswered [question](https://github.com/ggerganov/llama.cpp/pull/1237#issuecomment-1610181085). But I believe it would be reasonable to have some directions here....

@ggerganov : My tries to get `lcov` working on Windows failed miserably, but I got `clang/llvm` [coverage analysis](https://clang.llvm.org/docs/SourceBasedCodeCoverage.html) working. Here is a first summary: ``` Filename Regions Missed Regions Cover...

@Igoorx : @klosax just made me aware that we are working on weaknesses of the tokenizer(s) at the same time. I'd greatly appreciate any cooperation on this. If I'd go...

> @goerch Here is a simple test: [6f7daba](https://github.com/ggerganov/llama.cpp/commit/6f7dabab441566078446ef868e573cd309fe62be) Nice, thank you! Then my plan would be to try to first integrate your changes into #2315 and afterwards migrate the PR...

> @goerch Yeah, that's fine. You just have to be aware about: ... IANAL, but I'm equally concerned about compatibility with the `sentencepiece` license (Apache-2.0). They use a trie too,...