Jérôme de Courval

To replicate, just try to show two stacked bar datasets with one having negative values. Here is an example that shows the issue on version 1.6.1: ```html Title import {Chart}...

> > With your changes, it now works with ROCm HIP (with patch below), but it is slower, making it likely not worth enabling it on that platform. I'm using...

I managed to compile the branch. > I just tried to compile with C++17, `MK_CXXFLAGS="-std=c++17 -fPIC"`, now at least I don't see any rocm compilation issues. However I have these...

Thanks @tbocek, it was indeed something on my side, probably a leftover from another test. Having the code compile is a nice improvement, but it still doesn't work here if...

Here are some results. Token generation is about 2-3% faster. ``` ❯ ./bin/llama-bench -m ggml-c4ai-command-r-35b-v01-iq4_xs.gguf -m openhermes-2.5-mistral-7b-16k.Q5_K_M.gguf -m Nous-Hermes-2-Mixtral-8x7B-DPO.i1-IQ3_XXS.gguf -ngl 99 --flash-attn 1 --flash-attn 0 -r 10 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no...

> If it's only token generation that is faster then this PR is pretty much pointless because the FlashAttention kernel for batch size 1 does not use tensor cores at...

> My current stance is that I don't think that a speedup of 2% is large enough to justify adding a dependency, especially when there is no dev with the...

> On Arch Linux at least, I had to manually install the rocWMMA dependency from github, as there is no current library (also not in AUR). There's this one: https://aur.archlinux.org/packages/rocwmma...

> is this still a thing? Yes. `dpms` is one way to trigger the bug, but I can also just shutdown my monitor. This bug forced me back to sway,...

> > is this still a thing? > > Yes. `dpms` is one way to trigger the bug, but I can also just shutdown my monitor. This bug forced me...