HIP: force max threads per block to be 1024

Some old compilers still use 256. Explicitly set it to 1024 to get correct
result from ops like ARGMAX and GROUP_NORM.

Related: #10610, #11619
Signed-off-by: fxzjshm <[email protected]>

Author: fxzjshm

ggml/src/ggml-hip/CMakeLists.txt

@@ -46,6 +46,9 @@ endif()
 
 message(STATUS "HIP and hipBLAS found")
 
+# Workaround old compilers
+set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} --gpu-max-threads-per-block=1024")
+
 file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
 list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")