support mqb and global

CMakeLists.txt

@@ -644,7 +644,18 @@ if (GGML_USE_CPU_HBM)
     add_definitions(-DGGML_USE_CPU_HBM)
     find_library(memkind memkind REQUIRED)
 endif()
+# 查找libibverbs库
+find_library(IBVERBS_LIBRARY NAMES ibverbs PATHS /usr/lib /usr/lib64 /usr/lib/x86_64-linux-gnu)
 
+# 确保找到了库
+if(NOT IBVERBS_LIBRARY)
+    message(FATAL_ERROR "libibverbs not found")
+endif()
+
+# 在需要链接libibverbs的目标上使用找到的库路径
+find_library(RDMACM_LIBRARY rdmacm HINTS /home/atr/local/lib)
+
+link_libraries(pthread ${IBVERBS_LIBRARY} ${RDMACM_LIBRARY})
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -664,6 +675,8 @@ add_library(ggml OBJECT
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+target_link_libraries(ggml PUBLIC ${IBVERBS_LIBRARY})
+
 if (GGML_USE_CPU_HBM)
     target_link_libraries(ggml PUBLIC memkind)
 endif()
@@ -697,7 +710,7 @@ target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
 target_link_libraries(llama PRIVATE
     ggml
-    rdma_common # for rdma_example 将rdma_example的头文件加入到llama中
+    # rdma_common # for rdma_example 将rdma_example的头文件加入到llama中
     ${LLAMA_EXTRA_LIBS}
     )
 
@@ -710,6 +723,10 @@ if (BUILD_SHARED_LIBS)
 endif()
 
 
+
+
+
+
 #
 # install
 #
@@ -802,10 +819,10 @@ endif ()
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
     add_subdirectory(pocs)
-    add_subdirectory(rdma-example)
+    # add_subdirectory(rdma-example)
 endif()
 
 
 
-target_include_directories(llama PUBLIC ${PROJECT_SOURCE_DIR}/rdma-example/src/)#将rdma_example的头文件加入到llama中
+# target_include_directories(llama PUBLIC ${PROJECT_SOURCE_DIR}/rdma-example/src/)#将rdma_example的头文件加入到llama中
 

README.md

@@ -184,12 +184,16 @@ For the same reason, we suggest keeping the same directory structure as PowerInf
 
 ### server
 ```
-./build/bin/main_server -m ./Falcon-40B/falcon-40b-relu.q4.powerinfer.gguf  -n 128 -t 8 -p "Once upon a time" 
+./build/bin/main_server -m ./Falcon-40B/falcon-40b-relu.q4.powerinfer.gguf  -n 128 -t 8 -p "Once upon a time" --no-mmap --seed 0 --temp 0 
+./build/bin/main_server -m ./ReluLLaMA-7B/llama-7b-relu.powerinfer.gguf -n 128 -t 8 -p "Once upon a time" --no-mmap
+
 ```
 
 ### client
 ```
-./build/bin/main_client -m ./Falcon-40B/falcon-40b-relu.q4.powerinfer.gguf  -n 128 -t 8 -p "Once upon a time"  --seed 0 --temp 0 
+./build/bin/main_client -m ./Falcon-40B/falcon-40b-relu.q4.powerinfer.gguf  -n 128 -t 8 -p "Once upon a time"  --seed 0 --temp 0 --no-mmap
+./build/bin/main_client -m ./ReluLLaMA-7B/llama-7b-relu.powerinfer.gguf -n 128 -t 8 -p "Once upon a time" --no-mmap --seed 0 --temp 0 
+
 ```
 
 For CPU-only and CPU-GPU hybrid inference with all available VRAM, you can use the following instructions to run PowerInfer:

examples/main/CMakeLists.txt

@@ -6,12 +6,14 @@ link_libraries(pthread ${IBVERBS_LIBRARY} ${RDMACM_LIBRARY})
 add_executable(${TARGET} main.cpp)
 if(LLAMA_BUILD_RDMA_server)
     set(TARGET_ main_server)
-    add_executable(${TARGET_} main_server.cpp  ${PROJECT_SOURCE_DIR}/rdma-example/src/rdma_common.h)
+    add_executable(${TARGET_} main_server.cpp )
+    # add_executable(${TARGET_} main_server.cpp  ${PROJECT_SOURCE_DIR}/rdma-example/src/rdma_common.h)
 elseif(LLAMA_BUILD_RDMA_client)
     set(TARGET_ main_client)
-    add_executable(${TARGET_} main_client.cpp  ${PROJECT_SOURCE_DIR}/rdma-example/src/rdma_common.h)
+    add_executable(${TARGET_} main_client.cpp )
+    # add_executable(${TARGET_} main_client.cpp  ${PROJECT_SOURCE_DIR}/rdma-example/src/rdma_common.h)
     endif()
-target_include_directories(${TARGET_} PRIVATE ${PROJECT_SOURCE_DIR}/rdma-example/src)
+# target_include_directories(${TARGET_} PRIVATE ${PROJECT_SOURCE_DIR}/rdma-example/src)
 
 install(TARGETS ${TARGET} RUNTIME)
 install(TARGETS ${TARGET_} RUNTIME)