Thomas Bocek

I just tried this PR (ac6ae5daca029e554af08281a3fd839169725c8c), but could not compile. I installed rocwmma from source and I'm using ROCm 6.0.2. I tried to complie with: `make -j32 LLAMA_HIPBLAS=1 AMDGPU_TARGETS=gfx1100` ```...

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 issues now: ``` ggml-cuda/fattn.cu:425:34: error: use of...

nvm, there are still rocwmma compilation issue with c++17: ``` /opt/rocm/include/rocwmma/rocwmma_impl.hpp:189:9: note: candidate template ignored: deduced conflicting types for parameter 'DataT' ('__half' vs. 'float') fill_fragment(fragment& frag, ^ ``` Something is...

@jdecourval just tried your suggestions (I also know nothing about HIP/CUDA), but ChatGPT has a similar proposal for hmax2. And for me it compiles and works. I tried: ``` ./main...

@jdecourval I applied your patches. Here are my results with -p 4096: ``` ./llama-bench -m ../models/codellama-7b.Q4_K_M.gguf -m ../models/openhermes-2.5-mistral-7b-16k.Q5_K_M.gguf -ngl 99 --flash-attn 1 --flash-attn 0 -r 10 -p 4096 ggml_cuda_init: GGML_CUDA_FORCE_MMQ:...

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).

I'm also running into this issue. It says: local-sui-1 | Fullnode RPC URL: http://127.0.0.1:9000 local-sui-1 | Faucet URL: http://0.0.0.0:9123 And the commit https://github.com/MystenLabs/sui/commit/fd09307bb3ba48950f93c414b9c73f1090b27e7f makes the Faucet work, but not yet...

I believe the gnome mapping happens after I captures the keycodes. In order to make this work, a 5th modifier (CAPS_LOCK) needs to be added here: https://github.com/tbocek/dvorak/blob/c72932ee331f33b6e818788e2cd28a3db5a88ec7/dvorak.c#L90 and here https://github.com/tbocek/dvorak/blob/c72932ee331f33b6e818788e2cd28a3db5a88ec7/dvorak.c#L439,...

I am also interested in this feature, to run FIM with a local model. Qwen2.5-Coder does also a good job at inline completion.

I have the same problem with the most recent version 2.0.5. However, only on certain devices. I am also using arch. Not sure what I should be looking for to...