Merge branch 'remoteManagement' into remoteman_stripped

Author: Nexesenex

.github/workflows/kcpp-build-release-arm64.yaml

@@ -78,6 +78,8 @@ jobs:
               --add-data './kcpp_adapters:./kcpp_adapters' \
               --add-data './koboldcpp.py:.' \
               --add-data './json_to_gbnf.py:.' \
+              --add-data "./LICENSE.md:."  \
+              --add-data "./MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md:." \
               --add-data './klite.embd:.' \
               --add-data './kcpp_docs.embd:.' \
               --add-data './kcpp_sdui.embd:.' \

.github/workflows/kcpp-build-release-osx.yaml

@@ -36,7 +36,7 @@ jobs:
           make LLAMA_METAL=1 LLAMA_PORTABLE=1
           chmod +x './create_ver_file.sh'
           . create_ver_file.sh
-          pyinstaller --noconfirm --onefile --collect-all customtkinter --collect-all psutil --collect-all pdfplumber --add-data './koboldcpp_default.so:.' --add-data './ggml-metal-merged.metal:.' --add-data './kcpp_adapters:./kcpp_adapters' --add-data './koboldcpp.py:.' --add-data './json_to_gbnf.py:.' --add-data './klite.embd:.' --add-data './kcpp_docs.embd:.' --add-data './kcpp_sdui.embd:.' --add-data './taesd.embd:.' --add-data './taesd_xl.embd:.' --add-data './taesd_f.embd:.' --add-data './taesd_3.embd:.' --add-data './rwkv_vocab.embd:.' --add-data './rwkv_world_vocab.embd:.' --version-file './version.txt' --clean --console koboldcpp.py -n "koboldcpp-mac-arm64"
+          pyinstaller --noconfirm --onefile --collect-all customtkinter --collect-all psutil --collect-all pdfplumber --add-data './koboldcpp_default.so:.' --add-data './ggml-metal-merged.metal:.' --add-data './kcpp_adapters:./kcpp_adapters' --add-data './koboldcpp.py:.' --add-data './json_to_gbnf.py:.' --add-data './LICENSE.md:.' --add-data './MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md:.' --add-data './klite.embd:.' --add-data './kcpp_docs.embd:.' --add-data './kcpp_sdui.embd:.' --add-data './taesd.embd:.' --add-data './taesd_xl.embd:.' --add-data './taesd_f.embd:.' --add-data './taesd_3.embd:.' --add-data './rwkv_vocab.embd:.' --add-data './rwkv_world_vocab.embd:.' --version-file './version.txt' --clean --console koboldcpp.py -n "koboldcpp-mac-arm64"
 
       - name: Test
         id: test

.github/workflows/release.yml

@@ -238,14 +238,19 @@ jobs:
       matrix:
         include:
           - build: 'cpu-x64'
+            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
           #- build: 'openblas-x64'
+          #  arch: 'x64'
           #  defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
+            arch: 'x64'
             defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'cpu-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
           - build: 'opencl-adreno-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
     steps:
@@ -312,6 +317,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -339,7 +346,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
+          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts

MIT_LICENSE_GGML_LLAMACPP_ONLY renamed to MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md

@@ -22,5 +22,13 @@ SOFTWARE.
 
 ===================================
 
-Note that the above license applies ONLY to the GGML library and llama.cpp by ggerganov which are licensed under the MIT License
-KoboldAI Lite by Concedo and the provided python ctypes bindings in koboldcpp dlls are licensed under the AGPL v3.0 License
+Note that the above MIT license applies ONLY to the GGML library, the UNMODIFIED stable-diffusion.cpp portions and llama.cpp portions which are licensed under the MIT License
+
+KoboldAI Lite and the rest of KoboldCpp python and C++ code are licensed under the AGPL v3.0 License
+
+- Llama.cpp source repo is at https://github.com/ggml-org/llama.cpp (MIT)
+- Stable-diffusion.cpp source repo is at https://github.com/leejet/stable-diffusion.cpp (MIT)
+- KoboldCpp source repo is at https://github.com/LostRuins/koboldcpp (AGPL)
+- KoboldAI Lite source repo is at https://github.com/LostRuins/lite.koboldai.net (AGPL)
+
+For any further enquiries, contact @concedo on discord, or LostRuins on github.

README.md

@@ -293,9 +293,14 @@ and it will install everything required. Alternatively, you can download the abo
 - Since v1.75, openblas has been deprecated and removed in favor of the native CPU implementation.
 
 ## License
-- The original GGML library and llama.cpp by ggerganov are licensed under the MIT License
+- The original GGML library, stable-diffusion.cpp and llama.cpp by ggerganov are licensed under the MIT License
 - However, KoboldAI Lite is licensed under the AGPL v3.0 License
 - KoboldCpp code and other files are also under the AGPL v3.0 License unless otherwise stated
+- Llama.cpp source repo is at https://github.com/ggml-org/llama.cpp (MIT)
+- Stable-diffusion.cpp source repo is at https://github.com/leejet/stable-diffusion.cpp (MIT)
+- KoboldCpp source repo is at https://github.com/LostRuins/koboldcpp (AGPL)
+- KoboldAI Lite source repo is at https://github.com/LostRuins/lite.koboldai.net (AGPL)
+- For any further enquiries, contact @concedo on discord, or LostRuins on github.
 
 ## Notes
 - If you wish, after building the koboldcpp libraries with `make`, you can rebuild the exe yourself with pyinstaller by using `make_pyinstaller.bat`

common/arg.cpp

@@ -2586,7 +2586,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.n_junk = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_PASSKEY}));
+    ).set_examples({LLAMA_EXAMPLE_PASSKEY, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--pos"}, "N",
         string_format("position of the passkey in the junk text (default: %d)", params.i_pos),
@@ -2649,7 +2649,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.is_pp_shared = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_BENCH}));
+    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"-npp"}, "n0,n1,...",
         "number of prompt tokens",
@@ -2881,6 +2881,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
+    add_opt(common_arg(
+        {"--no-prefill-assistant"},
+        string_format(
+            "whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)\n"
+            "when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled\n"
+        ),
+        [](common_params & params) {
+            params.prefill_assistant = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_PREFILL_ASSISTANT"));
     add_opt(common_arg(
         {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
         string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),

common/common.h

@@ -364,6 +364,7 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
 

ggml/src/ggml-metal/ggml-metal.m

@@ -415,6 +415,13 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
@@ -1362,6 +1369,13 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,      flash_attn_ext_vec_f16_h64,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,     flash_attn_ext_vec_bf16_h64,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,     flash_attn_ext_vec_q4_0_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64,     flash_attn_ext_vec_q4_1_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64,     flash_attn_ext_vec_q5_0_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64,     flash_attn_ext_vec_q5_1_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64,     flash_attn_ext_vec_q8_0_h64,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,      flash_attn_ext_vec_f16_h96,      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,     flash_attn_ext_vec_bf16_h96,     has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,     flash_attn_ext_vec_q4_0_h96,     has_simdgroup_reduction);
@@ -4358,7 +4372,7 @@ static bool ggml_metal_encode_node(
                 // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
                 //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 //       these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
-                if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
+                if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
@@ -4539,6 +4553,24 @@ static bool ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 64:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 96:
                             {
                                 switch (src1->type) {

ggml/src/ggml-metal/ggml-metal.metal

@@ -4124,6 +4124,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  64, 64, 8>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 64, 64, 8>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 64, 64, 8>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -5896,10 +5896,17 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     vk_pipeline *pipelines;
     bool small_rows = N <= get_fa_num_small_rows(path);
 
+    // coopmat1 does not actually support "small rows" (it needs 16 rows).
+    // So use scalar instead.
     if (small_rows && path == FA_COOPMAT1) {
         path = FA_SCALAR;
     }
 
+    // scalar is faster than coopmat2 when N==1
+    if (N == 1 && path == FA_COOPMAT2) {
+        path = FA_SCALAR;
+    }
+
     bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
 
     switch (path) {