Disable GGML_HIP_ROCWMMA_FATTN due to negative effect on PP numbers for bigger context sizes

Author: Slobodan Josic

.github/workflows/build-llamacpp-rocm.yml

@@ -312,7 +312,7 @@ jobs:
           -DGGML_HIP=ON ^
           -DGGML_OPENMP=OFF ^
           -DGGML_CUDA_FORCE_CUBLAS=OFF ^
-          -DGGML_HIP_ROCWMMA_FATTN=ON ^
+          -DGGML_HIP_ROCWMMA_FATTN=OFF ^
           -DLLAMA_CURL=OFF ^
           -DGGML_NATIVE=OFF ^
           -DGGML_STATIC=OFF ^
@@ -720,7 +720,7 @@ jobs:
           -DGGML_HIP=ON \
           -DGGML_OPENMP=OFF \
           -DGGML_CUDA_FORCE_CUBLAS=OFF \
-          -DGGML_HIP_ROCWMMA_FATTN=ON \
+          -DGGML_HIP_ROCWMMA_FATTN=OFF \
           -DLLAMA_CURL=OFF \
           -DGGML_NATIVE=OFF \
           -DGGML_STATIC=OFF \

docs/manual_instructions.md

@@ -50,7 +50,7 @@ set HIP_PLATFORM=amd
 cd "C:\<YOUR_LLAMACPP_PATH>\llama.cpp"
 mkdir build
 cd build
-cmake .. -G Ninja -DCMAKE_C_COMPILER="C:\opt\rocm\lib\llvm\bin\clang.exe" -DCMAKE_CXX_COMPILER="C:\opt\rocm\lib\llvm\bin\clang++.exe" -DCMAKE_CXX_FLAGS="-IC:\opt\rocm\include" -DCMAKE_CROSSCOMPILING=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS="gfx1151" -DBUILD_SHARED_LIBS=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_HIP=ON -DGGML_OPENMP=OFF -DGGML_CUDA_FORCE_CUBLAS=OFF -DGGML_HIP_ROCWMMA_FATTN=ON -DLLAMA_CURL=OFF -DGGML_NATIVE=OFF -DGGML_STATIC=OFF -DCMAKE_SYSTEM_NAME=Windows
+cmake .. -G Ninja -DCMAKE_C_COMPILER="C:\opt\rocm\lib\llvm\bin\clang.exe" -DCMAKE_CXX_COMPILER="C:\opt\rocm\lib\llvm\bin\clang++.exe" -DCMAKE_CXX_FLAGS="-IC:\opt\rocm\include" -DCMAKE_CROSSCOMPILING=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS="gfx1151" -DBUILD_SHARED_LIBS=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_HIP=ON -DGGML_OPENMP=OFF -DGGML_CUDA_FORCE_CUBLAS=OFF -DGGML_HIP_ROCWMMA_FATTN=OFF -DLLAMA_CURL=OFF -DGGML_NATIVE=OFF -DGGML_STATIC=OFF -DCMAKE_SYSTEM_NAME=Windows
 cmake --build . -j 24 2>&1 | findstr /i "error"
 ```
 
@@ -133,7 +133,7 @@ cmake .. -G Ninja \
   -DGGML_HIP=ON \
   -DGGML_OPENMP=OFF \
   -DGGML_CUDA_FORCE_CUBLAS=OFF \
-  -DGGML_HIP_ROCWMMA_FATTN=ON \
+  -DGGML_HIP_ROCWMMA_FATTN=OFF \
   -DLLAMA_CURL=OFF \
   -DGGML_NATIVE=OFF \
   -DGGML_STATIC=OFF \