ggml : automatic selection of best CPU backend

Author: slaren

.devops/llama-server.Dockerfile

@@ -3,22 +3,34 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-ENV LLAMA_CURL=1
 
-RUN make -j$(nproc) llama-server
+RUN \
+    # Build multiple versions of the CPU backend
+    scripts/build-cpu.sh avx         -DGGML_AVX=ON -DGGML_AVX2=OFF && \
+    scripts/build-cpu.sh avx2        -DGGML_AVX=ON -DGGML_AVX2=ON && \
+    scripts/build-cpu.sh avx512      -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON && \
+    scripts/build-cpu.sh amx         -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \
+    # Build llama-server
+    cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build --target llama-server -j $(nproc) && \
+    # Copy the built libraries to /app/lib
+    mkdir -p /app/lib && \
+    mv libggml-cpu* /app/lib/ && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
 
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/build/bin/llama-server /llama-server
+COPY --from=build /app/lib/ /
 
 ENV LC_ALL=C.utf8
 # Must be set to 0.0.0.0 so it can listen to requests from host machine

ggml/src/ggml-backend-impl.h

@@ -211,27 +211,45 @@ extern "C" {
     GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Add backend dynamic loading support to the backend
-    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
 
-    #ifdef GGML_BACKEND_DL
-        #ifdef __cplusplus
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
-                extern "C" {                                                     \
-                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                }                                                                \
-                ggml_backend_reg_t ggml_backend_init(void) {                     \
-                    return reg_fn();                                             \
-                }
-        #else
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                             \
-                GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                ggml_backend_reg_t ggml_backend_init(void) {                 \
-                    return reg_fn();                                         \
-                }
-        #endif
-    #else
-    #    define GGML_BACKEND_DL_IMPL(reg_fn)
-    #endif
+    // Initialize the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+    // Optional: obtain a score for the backend based on the system configuration
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
+    typedef int                (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+#    ifdef __cplusplus
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            extern "C" {                                                 \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+            }                                                            \
+            ggml_backend_reg_t ggml_backend_init(void) {                 \
+                return reg_fn();                                         \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
+            extern "C" {                                   \
+            GGML_BACKEND_API int ggml_backend_score(void); \
+            }                                              \
+            int ggml_backend_score(void) {                 \
+                return score_fn();                         \
+            }
+#    else
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
+            ggml_backend_reg_t                  ggml_backend_init(void) { \
+                return reg_fn();                                          \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
+            GGML_BACKEND_API int ggml_backend_score(void);  \
+            int                  ggml_backend_score(void) { \
+                return score_fn();                          \
+            }
+#    endif
+#else
+#    define GGML_BACKEND_DL_IMPL(reg_fn)
+#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
 
 #ifdef  __cplusplus
 }