Merge 904111af8c into 41f477879f

* Avoid using saved CUDA graph if scale changes and reset nodes/params on update

Fixes https://github.com/ggerganov/llama.cpp/issues/9451

* clear before resize

.github/workflows/build.yml

@@ -967,6 +967,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-latest-cmake-hip:
+    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
 
     steps:
@@ -994,8 +995,72 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
-          cmake --build build --config Release
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+
+  windows-latest-cmake-hip-release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        gpu_target: [gfx1100, gfx1101, gfx1030]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -1060,6 +1125,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
+      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

ggml/include/ggml.h

@@ -2503,6 +2503,9 @@ extern "C" {
     GGML_API int ggml_cpu_has_cann       (void);
     GGML_API int ggml_cpu_has_llamafile  (void);
 
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
+
     //
     // Internal types and functions exposed for tests and benchmarks
     //

ggml/src/ggml-aarch64.c

@@ -547,8 +547,8 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -659,8 +659,8 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -777,7 +777,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -843,12 +843,12 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }
@@ -998,8 +998,8 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1519,8 +1519,8 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1981,7 +1981,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (ggml_sve_cnt_b == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2392,12 +2392,12 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_cpu_get_sve_cnt() == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }

ggml/src/ggml-cuda.cu

@@ -2478,6 +2478,7 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
     }
+    memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
 }
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
@@ -2509,6 +2510,12 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
             return false;
         }
     }
+
+    if (node->op == GGML_OP_SCALE &&
+        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+
     return true;
 }
 
@@ -2720,7 +2727,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
             // First call with null argument gets number of nodes in graph
             CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
             // Subsequent call with non-null argument gets nodes
+            cuda_ctx->cuda_graph->nodes.clear();
             cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
+            cuda_ctx->cuda_graph->params.clear();
             cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
             if (cuda_ctx->cuda_graph->num_nodes > 0) {
                 CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));

ggml/src/ggml-cuda/common.cuh

@@ -569,6 +569,7 @@ struct ggml_graph_node_properties {
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
     void * src_address[GGML_MAX_SRC];
+    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 
 struct ggml_cuda_graph {

ggml/src/ggml-quants.c

@@ -4013,7 +4013,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     svfloat32_t sumv0 = svdup_n_f32(0.0f);
     svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    const int vector_length = ggml_sve_cnt_b*8;
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
 
     // VLA Implementation using switch case
     switch (vector_length) {
@@ -5597,7 +5597,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     svfloat32_t sumv0 = svdup_n_f32(0.0f);
     svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    const int vector_length = ggml_sve_cnt_b*8;
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
 
     //VLA Implemenation for SVE
     switch (vector_length) {

ggml/src/ggml-quants.h

@@ -142,10 +142,6 @@ void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 
-#if defined(__ARM_FEATURE_SVE)
-extern int ggml_sve_cnt_b;
-#endif
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml.c

@@ -39,9 +39,15 @@
 #include <unistd.h>
 #endif
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_ARCH)
-int ggml_sve_cnt_b = 0;
+struct ggml_arm_arch_features_type {
+    int has_neon;
+    int has_i8mm;
+    int has_sve;
+    int sve_cnt;
+} ggml_arm_arch_features = {-1, -1, -1, 0};
 #endif
+
 #if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
@@ -3643,6 +3649,66 @@ static inline int ggml_up(int n, int m) {
 
 ////////////////////////////////////////////////////////////////////////////////
 
+#if defined(__ARM_ARCH)
+
+#if defined(__linux__) && defined(__aarch64__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+static void ggml_init_arm_arch_features(void) {
+#if defined(__linux__) && defined(__aarch64__)
+    uint32_t hwcap = getauxval(AT_HWCAP);
+    uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+    ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#endif
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_neon = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_i8mm = oldp;
+
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#else
+// Run-time CPU feature detection not implemented for this platform, fallback to compile time
+#if defined(__ARM_NEON)
+    ggml_arm_arch_features.has_neon = 1;
+#else
+    ggml_arm_arch_features.has_neon = 0;
+#endif
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_arm_arch_features.has_i8mm = 1;
+#else
+    ggml_arm_arch_features.has_i8mm = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.has_sve = 1;
+    ggml_arm_arch_features.sve_cnt = 16;
+#else
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#endif
+#endif
+}
+#endif
+
 struct ggml_context * ggml_init(struct ggml_init_params params) {
     // make this function thread safe
     ggml_critical_section_start();
@@ -3693,6 +3759,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
+#if defined(__ARM_ARCH)
+        ggml_init_arm_arch_features();
+#endif
+
         is_first_call = false;
     }
 
@@ -3741,12 +3811,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_ASSERT_ALIGNED(ctx->mem_buffer);
 
-#if defined(__ARM_FEATURE_SVE)
-    if (!ggml_sve_cnt_b) {
-        ggml_sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-    }
-#endif
-
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
     ggml_critical_section_end();
@@ -23545,16 +23609,16 @@ int ggml_cpu_has_fma(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_NEON)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
 }
 
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_sve;
 #else
     return 0;
 #endif
@@ -23701,11 +23765,18 @@ int ggml_cpu_has_vsx(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
 #endif
 }
 
+int ggml_cpu_get_sve_cnt(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.sve_cnt;
+#else
+    return 0;
+#endif
+}
 ////////////////////////////////////////////////////////////////////////////////