Vulkan Optimizations and Fixes (#8959)

* Optimize Vulkan REPEAT performance

* Use Vulkan GLSL fused multiply-add instruction where possible

* Add GGML_VULKAN_PERF option to output performance data per operator

* Rework and fix Vulkan descriptor set and descriptor pool handling

* Fix float32 concat f16 shader validation error

* Add Vulkan GROUP_NORM eps parameter

* Fix validation error with transfer queue memory barrier flags

* Remove trailing whitespaces

.devops/llama-server.Dockerfile

@@ -3,7 +3,7 @@ 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 curl
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -16,7 +16,7 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 COPY --from=build /app/llama-server /llama-server
 

.github/workflows/bench.yml

@@ -129,6 +129,8 @@ jobs:
 
       - name: Server bench
         id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -137,7 +139,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -47,7 +47,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -105,7 +105,7 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -222,7 +222,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -696,22 +696,20 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'

.github/workflows/python-check-requirements.yml

@@ -6,15 +6,13 @@ on:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

.gitignore

@@ -79,7 +79,6 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
-
 zig-out/
 zig-cache/
 

CONTRIBUTING.md

@@ -5,6 +5,7 @@
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
 - Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
   - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
+- Consider allowing write access to your branch for faster review
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)

Makefile

@@ -19,6 +19,7 @@ BUILD_TARGETS = \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
+	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -762,6 +763,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -888,15 +893,16 @@ ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
 	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -1205,6 +1211,7 @@ clean:
 	rm -rvf ggml/*.dll
 	rm -rvf ggml/*.so
 	rm -vrf ggml/src/*.o
+	rm -rvf ggml/src/llamafile/*.o
 	rm -rvf common/build-info.cpp
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
@@ -1451,15 +1458,20 @@ libllava.a: examples/llava/llava.cpp \
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/clip.h \
-	examples/llava/clip.cpp \
-	examples/llava/llava.h \
 	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
+llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift

README.md

@@ -95,8 +95,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -145,6 +153,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -177,10 +186,12 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 
 **Infrastructure:**
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
+- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.

common/common.cpp

@@ -684,14 +684,24 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--lora") {
         CHECK_ARG
-        params.lora_adapter.emplace_back(argv[i], 1.0f);
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
-        const char* lora_adapter = argv[i];
+        std::string lora_adapter = argv[i];
         CHECK_ARG
-        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
+        return true;
+    }
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
         return true;
     }
     if (arg == "--control-vector") {
@@ -1654,6 +1664,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
     options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
                                                                         "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 });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
 
 #ifndef LOG_DISABLE_LOGS
     options.push_back({ "logging" });
@@ -1766,6 +1777,17 @@ std::string string_get_sortable_timestamp() {
     return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
 }
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
+    }
+}
+
 void string_process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -2039,8 +2061,8 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
@@ -2055,7 +2077,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -2064,7 +2086,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
@@ -2075,7 +2097,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -2087,21 +2109,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
     }
 
-    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-        float lora_scale = std::get<1>(params.lora_adapter[i]);
-        auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
-        if (adapter == nullptr) {
-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
     if (params.ignore_eos) {
@@ -2129,13 +2156,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
             tmp.clear();
             tmp.push_back(decoder_start_token_id);
         }
-        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        if (llama_model_has_decoder(model)) {
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
-    return std::make_tuple(model, lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -3160,19 +3200,18 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     }
 
     fprintf(stream, "lora:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) != 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
         }
-        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) == 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
         }
-        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
     }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);

common/common.h

@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -126,8 +135,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -277,6 +286,8 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -308,8 +319,13 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
+
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
@@ -317,6 +333,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);

common/grammar-parser.cpp

@@ -369,6 +369,9 @@ namespace grammar_parser {
             }
             // Validate the state to ensure that all rules are defined
             for (const auto & rule : state.rules) {
+                if (rule.empty()) {
+                    throw std::runtime_error("Undefined rule");
+                }
                 for (const auto & elem : rule) {
                     if (elem.type == LLAMA_GRETYPE_RULE_REF) {
                         // Ensure that the rule at that location exists

convert_hf_to_gguf.py

@@ -251,12 +251,7 @@ class Model:
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del name, new_name, bid, n_dims  # unused
-
-        return False
-
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid, n_dims  # unused
 
         return False
@@ -285,55 +280,47 @@ class Model:
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
                 data: np.ndarray  # type hint
                 n_dims = len(data.shape)
-                data_dtype = data.dtype
-                data_qtype: gguf.GGMLQuantizationType | None = None
-
-                # when both are True, f32 should win
-                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
-                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
+                data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                extra_f32 = any(cond for cond in (
-                    extra_f32,
-                    n_dims == 1,
-                    new_name.endswith("_norm.weight"),
-                ))
-
-                # Some tensor types are always in float32
-                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
-                    gguf.MODEL_TENSOR.FFN_GATE_INP,
-                    gguf.MODEL_TENSOR.POS_EMBD,
-                    gguf.MODEL_TENSOR.TOKEN_TYPES,
-                ))
-
-                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = any(cond for cond in (
-                    extra_f16,
-                    (name.endswith(".weight") and n_dims >= 2),
-                ))
-
-                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.uint16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
-                        data = gguf.quantize_q8_0(data)
-                        assert data.dtype == np.uint8
-                        data_qtype = gguf.GGMLQuantizationType.Q8_0
-
-                    else:  # default to float16 for quantized tensors
-                        if data_dtype != np.float16:
-                            data = data.astype(np.float16)
-                        data_qtype = gguf.GGMLQuantizationType.F16
-
-                if data_qtype is None:  # by default, convert to float32
-                    if data_dtype != np.float32:
-                        data = data.astype(np.float32)
+                if n_dims <= 1 or new_name.endswith("_norm.weight"):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                # Some tensor types are always in float32
+                if data_qtype is False and (
+                    any(
+                        self.match_model_tensor_name(new_name, key, bid)
+                        for key in (
+                            gguf.MODEL_TENSOR.FFN_GATE_INP,
+                            gguf.MODEL_TENSOR.POS_EMBD,
+                            gguf.MODEL_TENSOR.TOKEN_TYPES,
+                        )
+                    )
+                    or not name.endswith(".weight")
+                ):
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                # No override (data_qtype is False), or wants to be quantized (data_qtype is True)
+                if isinstance(data_qtype, bool):
+                    if self.ftype == gguf.LlamaFileType.ALL_F32:
+                        data_qtype = gguf.GGMLQuantizationType.F32
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+                    else:
+                        raise ValueError(f"Unknown file type: {self.ftype.name}")
+
+                try:
+                    data = gguf.quants.quantize(data, data_qtype)
+                except gguf.QuantError as e:
+                    logger.warning("%s, %s", e, "falling back to F16")
+                    data_qtype = gguf.GGMLQuantizationType.F16
+                    data = gguf.quants.quantize(data, data_qtype)
+
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
                 # reverse shape to make it similar to the internal ggml dimension order
@@ -1765,7 +1752,7 @@ class DbrxModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid  # unused
 
         return n_dims > 1
@@ -2506,6 +2493,112 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
+@Model.register("XLMRobertaModel")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA
@@ -2680,18 +2773,22 @@ class MambaModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del n_dims  # unused
-
-        return bid is not None and new_name in (
-            self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+        if bid is not None and new_name in (
+            self.format_tensor_name(
+                n, bid, ".weight" if name.endswith(".weight") else ""
+            )
+            for n in [
                 gguf.MODEL_TENSOR.SSM_CONV1D,
                 gguf.MODEL_TENSOR.SSM_X,
                 gguf.MODEL_TENSOR.SSM_DT,
                 gguf.MODEL_TENSOR.SSM_A,
                 gguf.MODEL_TENSOR.SSM_D,
             ]
-        )
+        ):
+            return gguf.GGMLQuantizationType.F32
+
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
 @Model.register("CohereForCausalLM")
@@ -3227,6 +3324,145 @@ class T5Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("T5EncoderModel")
+class T5EncoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.T5ENCODER
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("JAISLMHeadModel")
 class JaisModel(Model):
     model_arch = gguf.MODEL_ARCH.JAIS

docs/backend/SYCL.md

@@ -80,7 +80,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +95,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -237,6 +244,13 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic
 ### II. Build llama.cpp
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
@@ -276,23 +290,26 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
+
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
@@ -304,12 +321,37 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 ```
 
+#### Choose level-zero devices
 
-4. Launch inference
+|Chosen Device ID|Setting|
+|-|-|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
+
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
+- Single device: Use one device assigned by user. Default device id is 0.
 - Multiple devices: Automatically choose the devices with the same backend.
 
 In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
@@ -326,11 +368,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -338,12 +375,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -390,7 +421,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -411,6 +442,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -425,12 +468,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -442,7 +481,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -450,23 +491,25 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
@@ -479,9 +522,27 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+#### Choose level-zero devices
 
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
-4. Launch inference
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
+
+2. Command line
+
+Launch inference
 
 There are two device selection modes:
 
@@ -508,11 +569,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -526,17 +583,18 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
 | Name               | Value                             | Function                                    |
 |--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -572,9 +630,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  It's same for other projects including llama.cpp SYCL backend.
+
+
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA

examples/baby-llama/baby-llama.cpp

@@ -1,7 +1,6 @@
 #include "ggml.h"
 #include "train.h"
 
-#include <vector>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>

examples/batched-bench/batched-bench.cpp

@@ -69,7 +69,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 

examples/cvector-generator/cvector-generator.cpp

@@ -414,9 +414,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);

examples/embedding/README.md

@@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
@@ -50,11 +50,11 @@ The above command will output space-separated float values.
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```

examples/embedding/embedding.cpp

@@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to decode\n", __func__);
+        }
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        const float * embd = nullptr;
+        int embd_pos = 0;
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
+            embd = llama_get_embeddings_ith(ctx, i);
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            embd_pos = batch.seq_id[i][0];
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        }
+
+        float * out = output + embd_pos * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -79,11 +101,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -93,8 +115,9 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+
+    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
         return 1;
     }
 
@@ -153,13 +176,23 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // count number of embeddings
+    int n_embd_count = 0;
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        for (int k = 0; k < n_prompts; k++) {
+            n_embd_count += inputs[k].size();
+        }
+    } else {
+        n_embd_count = n_prompts;
+    }
+
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
+    int e = 0; // number of embeddings already stored
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -169,11 +202,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
+            float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            llama_batch_clear(batch);
-            p += s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
+            llama_batch_clear(batch);
         }
 
         // add to batch
@@ -182,40 +215,63 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + p * n_embd;
+    float * out = emb + e * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                }
-            }
-            fprintf(stdout, "\n");
-        }
 
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
-            }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            for (int j = 0; j < n_embd_count; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < std::min(3, n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, " ... ");
+                for (int i = n_embd - 3; i < n_embd; i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
                 }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
                 fprintf(stdout, "\n");
             }
+        } else {
+            // print the first part of the embeddings or for a single prompt, the full embedding
+            for (int j = 0; j < n_prompts; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, "\n");
+            }
+
+            // print cosine similarity matrix
+            if (n_prompts > 1) {
+                fprintf(stdout, "\n");
+                printf("cosine similarity matrix:\n\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    fprintf(stdout, "%6.6s ", prompts[i].c_str());
+                }
+                fprintf(stdout, "\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    for (int j = 0; j < n_prompts; j++) {
+                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        fprintf(stdout, "%6.2f ", sim);
+                    }
+                    fprintf(stdout, "%1.10s", prompts[i].c_str());
+                    fprintf(stdout, "\n");
+                }
+            }
         }
     }
 
@@ -233,23 +289,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
         fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
             fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
+                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
                     float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                    if (j < n_embd_count) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }

examples/eval-callback/eval-callback.cpp

@@ -163,9 +163,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/export-lora/README.md

@@ -17,9 +17,9 @@ For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2-q8_0.gguf \
-    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```
 
 Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:

examples/export-lora/export-lora.cpp

@@ -10,6 +10,12 @@
 
 static bool g_verbose = false;
 
+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
 static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
     int id = gguf_find_key(ctx_gguf, key.c_str());
     return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@@ -50,20 +56,6 @@ static struct gguf_context * load_gguf(std::string & fname, struct ggml_context
     return ctx_gguf;
 }
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
-    }
-    s = std::move(result);
-}
-
 struct file_input {
     struct ggml_context * ctx_meta = nullptr;
     struct gguf_context * ctx_gguf = nullptr;
@@ -135,7 +127,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<std::tuple<std::string, float>> & lora_files,
+            std::vector<llama_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -144,9 +136,9 @@ struct lora_merge_ctx {
             throw std::runtime_error("split model is not yet supported");
         }
 
-        for (auto lora_inp : lora_files) {
-            auto fname = std::get<0>(lora_inp);
-            auto scale = std::get<1>(lora_inp);
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
             std::unique_ptr<file_input> adapter(new file_input(fname, scale));
             check_metadata_lora(adapter.get());
             adapters.push_back(std::move(adapter));
@@ -212,8 +204,7 @@ struct lora_merge_ctx {
         }
 
         // mapping base tensor to out tensor (same shape with base, but different type)
-        // if out_tensor == nullptr, we only copy it
-        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
+        std::vector<tensor_transformation> trans;
         for (auto & it : base_model.tensors) {
             bool t_a = true;
             bool t_b = true;
@@ -226,14 +217,22 @@ struct lora_merge_ctx {
                 // only copy
                 struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
                 ggml_set_name(cpy_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
                 gguf_add_tensor(ctx_out, cpy_tensor);
             } else if (t_a && t_b) {
                 // need merging
                 struct ggml_tensor * out_tensor = ggml_new_tensor(
                     ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
                 ggml_set_name(out_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
                 gguf_add_tensor(ctx_out, out_tensor);
             } else {
                 throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -248,12 +247,12 @@ struct lora_merge_ctx {
 
         // process base model tensors
         size_t n_merged = 0;
-        for (auto & it : base_to_out_tensors) {
-            if (it.second != nullptr) {
-                merge_tensor(it.first, it.second);
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
                 n_merged++;
             } else {
-                copy_tensor(it.first);
+                copy_tensor(it.in);
             }
         }
 
@@ -266,7 +265,7 @@ struct lora_merge_ctx {
         }
 
         printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -299,6 +298,10 @@ struct lora_merge_ctx {
         for (size_t i = 0; i < adapters.size(); ++i) {
             auto t_a = adapters[i]->get_tensor(name_lora_a);
             auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
             inp_a[i] = ggml_dup_tensor(ctx, t_a);
             inp_b[i] = ggml_dup_tensor(ctx, t_b);
         }
@@ -407,7 +410,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/imatrix/imatrix.cpp

@@ -611,10 +611,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/infill/infill.cpp

@@ -179,7 +179,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);

examples/llama-bench/llama-bench.cpp

@@ -27,6 +27,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -96,6 +104,27 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+    }
+    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;

examples/llava/CMakeLists.txt

@@ -36,3 +36,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET llama-minicpmv-cli)
+add_executable(${TARGET} minicpmv-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/llava/README-minicpmv2.5.md

@@ -0,0 +1,99 @@
+## MiniCPM-Llama3-V 2.5
+
+### Prepare models and code
+
+Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder.
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+### Usage
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
+
+```bash
+python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+python ./convert-hf-to-gguf.py ../MiniCPM-Llama3-V-2_5/model
+
+# quantize int4 version
+./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+Build for Linux or Mac
+
+```bash
+make
+make llama-minicpmv-cli
+```
+
+Inference on Linux or Mac
+```
+# run f16 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run quantized int4 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+
+# or run in interactive mode
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+```
+
+### Android
+
+#### Build on Android device using Termux
+We found that build on Android device would bring better runtime performance, so we recommend to build on device.
+
+[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
+
+Install tools in Termux:
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+
+```bash
+mkdir build-android
+cd build-android
+export NDK=/your_ndk_path
+cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+```

examples/llava/clip.cpp

@@ -74,26 +74,27 @@ static std::string format(const char * fmt, ...) {
 // key constants
 //
 
-#define KEY_FTYPE          "general.file_type"
-#define KEY_NAME           "general.name"
-#define KEY_DESCRIPTION    "general.description"
-#define KEY_HAS_TEXT_ENC   "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC    "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
-#define KEY_USE_GELU       "clip.use_gelu"
-#define KEY_N_EMBD         "clip.%s.embedding_length"
-#define KEY_N_FF           "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK        "clip.%s.block_count"
-#define KEY_N_HEAD         "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM       "clip.%s.projection_dim"
-#define KEY_TOKENS         "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS    "clip.text.context_length"
-#define KEY_IMAGE_SIZE     "clip.vision.image_size"
-#define KEY_PATCH_SIZE     "clip.vision.patch_size"
-#define KEY_IMAGE_MEAN     "clip.vision.image_mean"
-#define KEY_IMAGE_STD      "clip.vision.image_std"
-#define KEY_PROJ_TYPE      "clip.projector_type"
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -127,12 +128,20 @@ static std::string format(const char * fmt, ...) {
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY "resampler.query"
+#define TN_MINICPMV_PROJ "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
+#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN "resampler.ln_%s.%s"
+
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -140,6 +149,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
 };
 
 
@@ -200,17 +210,14 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 }
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
     }
-    s = std::move(result);
 }
 
 static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@@ -492,12 +499,33 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_mlp_2_b;
     struct ggml_tensor * mm_model_peg_0_w;
     struct ggml_tensor * mm_model_peg_0_b;
+
+    // MINICPMV projection
+    struct ggml_tensor * mm_model_pos_embed_k;
+    struct ggml_tensor * mm_model_query;
+    struct ggml_tensor * mm_model_proj;
+    struct ggml_tensor * mm_model_kv_proj;
+    struct ggml_tensor * mm_model_attn_q_w;
+    struct ggml_tensor * mm_model_attn_q_b;
+    struct ggml_tensor * mm_model_attn_k_w;
+    struct ggml_tensor * mm_model_attn_k_b;
+    struct ggml_tensor * mm_model_attn_v_w;
+    struct ggml_tensor * mm_model_attn_v_b;
+    struct ggml_tensor * mm_model_attn_o_w;
+    struct ggml_tensor * mm_model_attn_o_b;
+    struct ggml_tensor * mm_model_ln_q_w;
+    struct ggml_tensor * mm_model_ln_q_b;
+    struct ggml_tensor * mm_model_ln_kv_w;
+    struct ggml_tensor * mm_model_ln_kv_b;
+    struct ggml_tensor * mm_model_ln_post_w;
+    struct ggml_tensor * mm_model_ln_post_b;
 };
 
 struct clip_ctx {
     bool has_text_encoder    = false;
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
+    bool has_minicpmv_projector = false;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -522,9 +550,11 @@ struct clip_ctx {
 
     ggml_backend_t backend       = NULL;
     ggml_gallocr_t compute_alloc = NULL;
+
+    struct clip_image_size * load_image_size;
 };
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -533,20 +563,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size           = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        if (load_image_size == nullptr) {
+            load_image_size = clip_image_size_init();
+        }
+        LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        image_size_width  = load_image_size->width;
+        image_size_height = load_image_size->height;
+        if (is_inf) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
     const int patch_size           = hparams.patch_size;
-    const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
-    const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
-    const int n_layer              = hparams.n_layer;
+    int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
 
     const int batch_size = imgs->size;
 
-    if (ctx->has_llava_projector) {
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
@@ -559,7 +602,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
-    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
@@ -572,19 +615,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
         inp = ggml_add(ctx0, inp, model.patch_bias);
     }
-
-    // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = inp;
-    if (ctx->has_class_embedding) {
-        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-        ggml_set_name(embeddings, "embeddings");
-        ggml_set_input(embeddings);
-        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-        embeddings = ggml_acc(ctx0, embeddings, inp,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
-    }
+    struct ggml_tensor * pos_embed = nullptr;
 
+    if (ctx->has_llava_projector) {
+        // concat class_embeddings and patch_embeddings
+        if (ctx->has_class_embedding) {
+            embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+            ggml_set_name(embeddings, "embeddings");
+            ggml_set_input(embeddings);
+            embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+            embeddings = ggml_acc(ctx0, embeddings, inp,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+        }
+    }
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
@@ -593,6 +638,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     embeddings =
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
+    if (ctx->has_minicpmv_projector) {
+        int pos_w = image_size_width/patch_size;
+        int pos_h = image_size_height/patch_size;
+        pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        ggml_set_name(pos_embed, "pos_embed");
+        ggml_set_input(pos_embed);
+    }
+
     // pre-layernorm
     if (ctx->has_pre_norm) {
         embeddings = ggml_norm(ctx0, embeddings, eps);
@@ -602,6 +655,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // loop over layers
+    if (ctx->has_minicpmv_projector) {
+        n_layer += 1;
+    }
     for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
 
@@ -691,7 +747,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // llava projector
-    {
+    if (ctx->has_llava_projector) {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
@@ -872,6 +928,65 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ABORT("fatal error");
         }
     }
+    // minicpmv projector
+    else if (ctx->has_minicpmv_projector)
+    {
+        if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            struct ggml_tensor * q = model.mm_model_query;
+            { // layernorm
+                q = ggml_norm(ctx0, q, eps);
+                q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            }
+            struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+            { // layernorm
+                v = ggml_norm(ctx0, v, eps);
+                v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
+            }
+            struct ggml_tensor * k;
+            { // position
+                // q = ggml_add(ctx0, q, model.mm_model_pos_embed);
+                k = ggml_add(ctx0, v, pos_embed);
+            }
+
+            { // attention
+                const int hidden_size = 4096;
+                const int d_head = 128;
+                const int n_head = hidden_size/d_head;
+                const int num_query = 96;
+
+                struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
+                Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+                struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
+                struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
+                // permute
+                Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
+                Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+                Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
+                K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+                K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+                K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+                V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+                V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+                V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                KQ = ggml_soft_max_inplace(ctx0, KQ);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+                KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
+                KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
+
+                embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
+            }
+            { // layernorm
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
+            }
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+        }
+        else {
+            GGML_ASSERT(false);
+        }
+    }
 
     // build the graph
     ggml_build_forward_expand(gf, embeddings);
@@ -1029,7 +1144,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
         }
 
-        GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+        idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ);
+        if (idx != -1) {
+            new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+
         GGML_ASSERT(new_clip->has_vision_encoder);
         GGML_ASSERT(!new_clip->has_text_encoder);
 
@@ -1040,6 +1161,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
             LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
@@ -1281,6 +1403,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
             vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
+            vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K);
+            vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY);
+            vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ);
+            vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ);
+            vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight"));
+            vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight"));
+            vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight"));
+            vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias"));
+            vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias"));
+            vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias"));
+            vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight"));
+            vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias"));
+            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight"));
+            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias"));
+            vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight"));
+            vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias"));
+            vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
+            vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1319,7 +1462,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
-        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
         LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
@@ -1328,6 +1471,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     return new_clip;
 }
 
+void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
+    ctx_clip->load_image_size = load_image_size;
+}
+
+struct clip_image_size * clip_image_size_init() {
+    struct clip_image_size * load_image_size = new struct clip_image_size();
+    load_image_size->width = 448;
+    load_image_size->height = 448;
+    return load_image_size;
+}
+
 struct clip_image_u8 * clip_image_u8_init() {
     return new clip_image_u8();
 }
@@ -1598,9 +1752,184 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
     return patches;
 }
 
+static int ensure_divide(int length, int patch_size) {
+    return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
+}
+
+static std::pair<int, int> uhd_find_best_resize(std::pair<int, int> original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width = original_size.first;
+    int height = original_size.second;
+    if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
+        float r = static_cast<float>(width) / height;
+        height = static_cast<int>(scale_resolution / std::sqrt(r));
+        width = static_cast<int>(height * r);
+    }
+    int best_width = ensure_divide(width, patch_size);
+    int best_height = ensure_divide(height, patch_size);
+    return std::make_pair(best_width, best_height);
+}
+
+static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size, std::pair<int, int> grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width, height;
+    std::tie(width, height) = original_size;
+    int grid_x, grid_y;
+    std::tie(grid_x, grid_y) = grid;
+
+    int refine_width = ensure_divide(width, grid_x);
+    int refine_height = ensure_divide(height, grid_y);
+
+    int grid_width = refine_width / grid_x;
+    int grid_height = refine_height / grid_y;
+
+   // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line)
+    auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair
+    int best_grid_width, best_grid_height;
+    std::tie(best_grid_width, best_grid_height) = best_grid_size;
+
+  //  std::pair<int, int> refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line)
+    std::pair<int, int> refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line)
+    return refine_size;
+}
+
+inline int clip(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
+    std::vector<int> candidate_split_grids_nums;
+    for (int i : {multiple - 1, multiple, multiple + 1}) {
+        if (i == 1 || i > max_slice_nums) {
+            continue;
+        }
+        candidate_split_grids_nums.push_back(i);
+    }
+
+    std::vector<std::pair<int, int>> candidate_grids;
+    for (int split_grids_nums : candidate_split_grids_nums) {
+        int m = 1;
+        while (m <= split_grids_nums) {
+            if (split_grids_nums % m == 0) {
+                candidate_grids.emplace_back(m, split_grids_nums / m);
+            }
+            ++m;
+        }
+    }
+
+    std::pair<int, int> best_grid{1, 1};
+    float min_error = std::numeric_limits<float>::infinity();
+    for (const auto& grid : candidate_grids) {
+        float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
+        if (error < min_error) {
+            best_grid = grid;
+            min_error = error;
+        }
+    }
+    return best_grid;
+}
+
+// inspired from LLaVA-UHD:
+//    -> https://arxiv.org/pdf/2403.11703
+//    -> https://github.com/thunlp/LLaVA-UHD
+//    -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
+static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
+    const std::pair<int, int> original_size={img->nx,img->ny};
+    const int original_width = img->nx;
+    const int original_height = img->ny;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+
+    std::vector<std::vector<clip_image_u8 *>> images;
+    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    images.push_back(std::vector<clip_image_u8 *>());
+
+    if (multiple <= 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.resize(best_size, Image.Resampling.BICUBIC)
+        images[images.size()-1].push_back(source_image);
+    }
+    else if (multiple > 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
+        LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
+        images[images.size()-1].push_back(source_image);
+
+        std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+
+        auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
+        clip_image_u8 * refine_image = clip_image_u8_init();
+        bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
+
+        LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
+
+        // split_to_patches
+        int width = refine_image->nx;
+        int height = refine_image->ny;
+        int grid_x = int(width / best_grid.first);
+        int grid_y = int(height / best_grid.second);
+        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
+            images.push_back(std::vector<clip_image_u8 *>());
+            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
+                clip_image_u8 * patch = clip_image_u8_init();
+                patch->nx = grid_x;
+                patch->ny = grid_y;
+                patch->buf.resize(3 * patch->nx * patch->ny);
+                for (int y = patches_i; y < patches_i + grid_y; ++y) {
+                    for (int x = patches_j; x < patches_j + grid_x; ++x) {
+                        const int i = 3 * (y * refine_image->nx + x);
+                        const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
+                        patch->buf[j]   = refine_image->buf[i];
+                        patch->buf[j+1] = refine_image->buf[i+1];
+                        patch->buf[j+2] = refine_image->buf[i+2];
+                    }
+                }
+                images[images.size()-1].push_back(patch);
+            }
+        }
+    }
+    return images;
+}
+
+int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
+    const int max_slice_nums=9;
+    const int scale_resolution=448;
+    const int original_width = ctx_clip->load_image_size->width;
+    const int original_height = ctx_clip->load_image_size->height;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+    std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+    return best_grid.first;
+}
+
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
+    if (clip_is_minicpmv(ctx)) {
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img);
+        res_imgs->size = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            res_imgs->size += imgs[i].size();
+        }
+        res_imgs->data = new clip_image_f32[res_imgs->size];
+        int idx = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            for (size_t j = 0; j < imgs[i].size(); ++j) {
+                LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+                clip_image_f32 * res = clip_image_f32_init();
+                normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
+                res_imgs->data[idx++] = *res;
+                clip_image_f32_free(res);
+            }
+        }
+        return true;
+    }
+
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1816,11 +2145,99 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
     if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
+    } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        n_patches = 96;
     }
 
     return n_patches;
 }
 
+static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
+    assert(embed_dim % 2 == 0);
+    int H = pos.size();
+    int W = pos[0].size();
+
+    std::vector<float> omega(embed_dim / 2);
+    for (int i = 0; i < embed_dim / 2; ++i) {
+        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
+    }
+
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                float out_value = pos[h][w] * omega[d];
+                emb[h][w][d] = sin(out_value);
+                emb[h][w][d + embed_dim / 2] = cos(out_value);
+            }
+        }
+    }
+
+    return emb;
+}
+
+static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
+    assert(embed_dim % 2 == 0);
+    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
+    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
+
+    int H = emb_h.size();
+    int W = emb_h[0].size();
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                emb[h][w][d] = emb_h[h][w][d];
+                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
+            }
+        }
+    }
+    return emb;
+}
+
+static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
+    int grid_h_size = image_size.first;
+    int grid_w_size = image_size.second;
+
+    std::vector<float> grid_h(grid_h_size);
+    std::vector<float> grid_w(grid_w_size);
+
+    for (int i = 0; i < grid_h_size; ++i) {
+        grid_h[i] = static_cast<float>(i);
+    }
+    for (int i = 0; i < grid_w_size; ++i) {
+        grid_w[i] = static_cast<float>(i);
+    }
+
+    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid[h][w] = grid_w[w];
+        }
+    }
+    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid_2d[0][h][w] = grid_h[h];
+            grid_2d[1][h][w] = grid_w[w];
+        }
+    }
+
+    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
+
+    int H = image_size.first;
+    int W = image_size.second;
+    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
+        }
+    }
+
+    return pos_embed_2d;
+}
+
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1843,18 +2260,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1); // TODO: support multiple images
     }
+    if (ctx->has_minicpmv_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
 
     // build the inference graph
-    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
     ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
 
     // set inputs
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size    = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        image_size_width  = imgs->data[0].nx;
+        image_size_height = imgs->data[0].ny;
+    }
     const int patch_size    = hparams.patch_size;
-    const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
 
     {
@@ -1864,7 +2290,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
             const int ny = imgs->data[i].ny;
-            GGML_ASSERT(nx == image_size && ny == image_size);
+            if (!ctx->has_minicpmv_projector) {
+                GGML_ASSERT(nx == image_size && ny == image_size);
+            }
 
             const int n = nx * ny;
 
@@ -1881,37 +2309,75 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
         free(data);
     }
-
-    {
-        if (ctx->has_class_embedding) {
-            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
-
-            void* zero_mem = malloc(ggml_nbytes(embeddings));
-            memset(zero_mem, 0, ggml_nbytes(embeddings));
-            ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
-            free(zero_mem);
+    if (ctx->has_minicpmv_projector) {
+        {
+            // inspired from siglip:
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = std::floor(70.0*i/num_positions);
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
         }
-    }
 
-    {
-        struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+        {
+            // inspired from resampler of Qwen-VL:
+            //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
+            //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
+            struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
+            if(ctx->load_image_size==nullptr){
+                ctx->load_image_size= clip_image_size_init();
+            }
+            int pos_w = ctx->load_image_size->width/patch_size;
+            int pos_h = ctx->load_image_size->height/patch_size;
+            int embed_dim = 4096;
+            auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
-        int* positions_data = (int*)malloc(ggml_nbytes(positions));
-        for (int i = 0; i < num_positions; i++) {
-            positions_data[i] = i;
+            float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
+            for(int i=0;i<pos_w * pos_h;++i){
+                for(int j=0;j<embed_dim;++j){
+                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+                }
+            }
+
+            ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
+            free(pos_embed_data);
         }
-        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
-        free(positions_data);
-    }
+    } else {
+        {
+            if (ctx->has_class_embedding) {
+                struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
-    {
-        struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-        int* patches_data = (int*)malloc(ggml_nbytes(patches));
-        for (int i = 0; i < num_patches; i++) {
-            patches_data[i] = i + 1;
+                void* zero_mem = malloc(ggml_nbytes(embeddings));
+                memset(zero_mem, 0, ggml_nbytes(embeddings));
+                ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+                free(zero_mem);
+            }
+        }
+
+        {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+
+        {
+            struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
+            }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
         }
-        ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-        free(patches_data);
     }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
@@ -2081,7 +2547,14 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
         return ctx->vision_model.mm_3_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        return 4096;
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
 }
+
+bool clip_is_minicpmv(const struct clip_ctx * ctx) {
+    return ctx->has_minicpmv_projector;
+}

examples/llava/clip.h

@@ -18,14 +18,17 @@
 #    define CLIP_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 struct clip_ctx;
 
+struct clip_image_size {
+    int width;
+    int height;
+};
+
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
     size_t size;
@@ -55,6 +58,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
 CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
+CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
+CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+
+CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
@@ -78,6 +85,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
+CLIP_API bool clip_is_minicpmv(const struct clip_ctx * ctx);
+
 #ifdef __cplusplus
 }
 #endif

examples/llava/llava.cpp

@@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     return true;
 }
 
+static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
+    int width = image->nx;
+    int height = image->ny;
+    int num_patches = (height / patch_size) * (width / patch_size);
+    clip_image_f32 * patch = clip_image_f32_init();
+    patch->nx = patch_size * num_patches;
+    patch->ny = patch_size;
+    patch->buf.resize(3 * patch->nx * patch->ny);
+
+    int patch_index = 0;
+
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            for (int pi = 0; pi < patch_size; ++pi) {
+                for (int pj = 0; pj < patch_size; ++pj) {
+                    int input_index = ((i + pi) * width + (j + pj)) * 3;
+                    int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
+                    patch->buf[output_index] = image->buf[input_index];
+                    patch->buf[output_index+1] = image->buf[input_index+1];
+                    patch->buf[output_index+2] = image->buf[input_index+2];
+                }
+            }
+            patch_index++;
+        }
+    }
+    return patch;
+}
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
@@ -218,7 +245,44 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+    if (clip_is_minicpmv(ctx_clip)) {
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            const int64_t t_img_enc_step_start_us = ggml_time_us();
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
+            int patch_size=14;
+            load_image_size->width = img_res_v.data[i].nx;
+            load_image_size->height = img_res_v.data[i].ny;
+            clip_add_load_image_size(ctx_clip, load_image_size);
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            if (!encoded) {
+                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+            const int64_t t_img_enc_steop_batch_us = ggml_time_us();
+            LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        int n_img_pos_out = 0;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
+            n_img_pos_out += clip_n_patches(ctx_clip);
+        }
+        *n_img_pos = n_img_pos_out;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+        load_image_size->width = img->nx;
+        load_image_size->height = img->ny;
+        clip_add_load_image_size(ctx_clip, load_image_size);
+        LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+    }
+    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
@@ -228,7 +292,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
             return false;
         }
-    } else {
+    }
+    else {
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
@@ -297,7 +362,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
+    int num_max_patches = 6;
+    if (clip_is_minicpmv(ctx_clip)) {
+        num_max_patches = 10;
+    }
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
     if (!image_embd) {
         LOG_TEE("Unable to allocate memory for image embeddings\n");
         return false;

examples/llava/llava.h

@@ -17,12 +17,11 @@
 #    define LLAVA_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+struct clip_ctx;
 struct llava_image_embed {
     float * embed;
     int n_image_pos;
@@ -37,8 +36,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip,
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** free an embedding made with llava_image_embed_make_* */
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);

examples/llava/minicpmv-cli.cpp

@@ -0,0 +1,309 @@
+#include "ggml.h"
+#include "log.h"
+#include "common.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
+static struct llama_model * llava_init(gpt_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    if (params->n_ctx < 2048) {
+        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
+        LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
+        ctx_params.n_ctx = 2048;
+    } else {
+        ctx_params.n_ctx = params->n_ctx;
+    }
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+static struct clip_ctx * clip_init_context(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+    return ctx_clip;
+}
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+}
+
+static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
+    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
+    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
+
+    auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    slice_embed->embed = image_embed;
+    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
+    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
+    llava_image_embed_free(slice_embed);
+}
+
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
+    std::string system_prompt;
+    int idx = 0;
+    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
+    system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
+    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+    if (num_image_embeds > 1) {
+        size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+        eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+        for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+            for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+                process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                if (j == num_image_embeds_col - 1) {
+                    eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                }
+            }
+        }
+        eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+    }
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+}
+
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
+    auto ctx_clip = clip_init_context(params);
+    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
+    if (!embeds) {
+        std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
+        return NULL;
+    }
+
+    // process the prompt
+    if (params->prompt.empty() && params->interactive == false) {
+        LOG_TEE("prompt should be given or interactive mode should be on");
+        return NULL;
+    }
+
+    auto model = llava_init(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
+        return NULL;
+    }
+    const int64_t t_llava_init_start_us = ggml_time_us();
+    auto ctx_llava = llava_init_context(params, model);
+    ctx_llava->ctx_clip = ctx_clip;
+    const int64_t t_llava_init_end_us = ggml_time_us();
+    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
+    LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
+
+    const int64_t t_process_image_start_us = ggml_time_us();
+    process_image(ctx_llava, embeds, params, n_past);
+    const int64_t t_process_image_end_us = ggml_time_us();
+    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
+    LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
+
+    llava_image_embed_free(embeds);
+    return ctx_llava;
+}
+
+static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
+    std::string user_prompt = prompt;
+    if (!is_first) user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+    eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    // generate the response
+
+    LOG_TEE("\n");
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    return ctx_sampling;
+}
+
+static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
+
+    const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+    return tmp;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("llava", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
+    if (params.mmproj.empty() || (params.image.empty())) {
+        gpt_params_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    for (auto & image : params.image) {
+        int n_past = 0;
+        auto ctx_llava = minicpmv_init(&params, image, n_past);
+
+        if (!params.prompt.empty()) {
+            LOG_TEE("<user>%s\n", params.prompt.c_str());
+            LOG_TEE("<assistant>");
+            auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
+            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+            std::string response = "";
+            bool have_tmp = false;
+            for (int i = 0; i < max_tgt_len; i++) {
+                auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                response += tmp;
+                if (strcmp(tmp, "</s>") == 0){
+                    if(!have_tmp)continue;
+                    else break;
+                }
+                if (strstr(tmp, "###")) break; // Yi-VL behavior
+                have_tmp = true;
+                printf("%s", tmp);
+                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+
+                fflush(stdout);
+            }
+            llama_sampling_free(ctx_sampling);
+        }else {
+            while (true) {
+                LOG_TEE("<user>");
+                std::string prompt;
+                std::getline(std::cin, prompt);
+                LOG_TEE("<assistant>");
+                auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
+                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+                std::string response = "";
+                for (int i = 0; i < max_tgt_len; i++) {
+                    auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                    response += tmp;
+                    if (strcmp(tmp, "</s>") == 0) break;
+                    if (strstr(tmp, "###")) break; // Yi-VL behavior
+                    printf("%s", tmp);// mistral llava-1.6
+                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+                    fflush(stdout);
+                }
+                llama_sampling_free(ctx_sampling);
+            }
+        }
+        printf("\n");
+        llama_print_timings(ctx_llava->ctx_llama);
+
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+    }
+
+    return 0;
+}

examples/llava/minicpmv-convert-image-encoder-to-gguf.py

@@ -0,0 +1,382 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def add_key_str(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_minicpmv and name in ["visual_projection.weight"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+# if args.clip_model_is_vision or args.clip_model_is_openclip:
+#     model = CLIPVisionModel.from_pretrained(dir_model)
+#     processor = None
+# else:
+#     model = CLIPModel.from_pretrained(dir_model)
+#     processor = CLIPProcessor.from_pretrained(dir_model)
+
+default_vision_config = {
+        "hidden_size": 1152,
+        "image_size": 980,
+        "intermediate_size": 4304,
+        "model_type": "idefics2",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 27,
+        "patch_size": 14,
+    }
+vision_config = Idefics2VisionConfig(**default_vision_config)
+model = Idefics2VisionTransformer(vision_config)
+
+processor = None
+# if model.attn_pool is not None:
+#     model.attn_pool = torch.nn.Identity()
+
+# model.blocks = model.blocks[:-1]
+model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip")))
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_minicpmv_projector = False
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.minicpmv_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_minicpmv_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector)
+fout.add_file_type(ftype)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_minicpmv_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_minicpmv_projector:
+    fout.add_description("image encoder for MiniCPM-V")
+    # add projector type
+    fout.add_string("clip.projector_type", "resampler")
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", 448)
+    fout.add_uint32("clip.vision.patch_size", 14)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    block_count = 26
+    fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = True
+fout.add_bool("clip.use_gelu", use_gelu)
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+def _replace_name_resampler(s, v):
+    if re.match("resampler.pos_embed", s):
+        return {
+            s: v,
+            re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
+        }
+    if re.match("resampler.proj", s):
+        return {
+            re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
+            re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
+        }
+    if re.match("resampler.attn.in_proj_.*", s):
+        return {
+            re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
+            re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
+            re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
+        }
+    return {s: v}
+
+if has_minicpmv_projector:
+    projector = torch.load(args.minicpmv_projector)
+    new_state_dict = {}
+    for k, v in projector.items():
+        kvs = _replace_name_resampler(k, v)
+        for nk, nv in kvs.items():
+            new_state_dict[nk] = nv
+    projector = new_state_dict
+    ftype_cur = 0
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        data = data.squeeze().numpy()
+
+        n_dims = len(data.shape)
+        if ftype == 1:
+            if name[-7:] == ".weight" and n_dims == 2:
+                print("  Converting to float16")
+                data = data.astype(np.float16)
+                ftype_cur = 1
+            else:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+        else:
+            if data.dtype != np.float32:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+
+        fout.add_tensor(name, data)
+        print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+
+    print("Projector tensors added\n")
+
+def _replace_name(s, v):
+    s = "vision_model." + s
+    if re.match("vision_model.embeddings.position_embedding", s):
+        v = v.unsqueeze(0)
+        return {s: v}
+
+    return {s: v}
+
+state_dict = model.state_dict()
+new_state_dict = {}
+for k, v in state_dict.items():
+    kvs = _replace_name(k, v)
+    for nk, nv in kvs.items():
+        new_state_dict[nk] = nv
+state_dict = new_state_dict
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

examples/llava/minicpmv-surgery.py

@@ -0,0 +1,47 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to MiniCPM-V-2.5 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/minicpmv.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/minicpmv.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config._name_or_path = "openbmb/MiniCPM-Llama3-V-2.5"
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/model")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/model")
+# os.system(f"cp {args.model}/modeling_minicpm.py {args.model}/MiniCPM_l3/modeling_minicpm.py")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")

examples/llava/requirements.txt

@@ -2,3 +2,4 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=10.2.0
 torch~=2.2.1
+torchvision~=0.17.1

examples/lookahead/lookahead.cpp

@@ -58,11 +58,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup-create.cpp

@@ -22,11 +22,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt

examples/lookup/lookup-stats.cpp

@@ -26,11 +26,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup.cpp

@@ -34,11 +34,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/main/main.cpp

@@ -207,7 +207,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
     if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);

examples/parallel/parallel.cpp

@@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {

examples/perplexity/perplexity.cpp

@@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/quantize/quantize.cpp

@@ -91,7 +91,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {

examples/retrieval/retrieval.cpp

@@ -148,11 +148,12 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/rpc/README.md

@@ -1,5 +1,9 @@
 ## Overview
 
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
 The `rpc-server` allows  running `ggml` backend on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:

examples/rpc/rpc-server.cpp

@@ -16,7 +16,7 @@
 #include <stdio.h>
 
 struct rpc_server_params {
-    std::string host        = "0.0.0.0";
+    std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
 };
@@ -114,6 +114,17 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "Invalid parameters\n");
         return 1;
     }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
     ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");

examples/save-load-state/save-load-state.cpp

@@ -28,10 +28,11 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/server/README.md

@@ -207,41 +207,6 @@ model:
   -hff,  --hf-file FILE           Hugging Face model file (default: unused)
   -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
 
-retrieval:
-
-         --context-file FNAME     file to load context from (repeat to specify multiple files)
-         --chunk-size N           minimum length of embedded text chunks (default: 64)
-         --chunk-separator STRING
-                                  separator between chunks (default: '
-                                  ')
-
-passkey:
-
-         --junk N                 number of times to repeat the junk text (default: 250)
-         --pos N                  position of the passkey in the junk text (default: -1)
-
-imatrix:
-
-  -o,    --output FNAME           output file (default: 'imatrix.dat')
-         --output-frequency N     output the imatrix every N iterations (default: 10)
-         --save-frequency N       save an imatrix copy every N iterations (default: 0)
-         --process-output         collect data for the output tensor (default: false)
-         --no-ppl                 do not compute perplexity (default: true)
-         --chunk N                start processing the input from chunk N (default: 0)
-
-bench:
-
-  -pps                            is the prompt shared across parallel sequences (default: false)
-  -npp n0,n1,...                  number of prompt tokens
-  -ntg n0,n1,...                  number of text generation tokens
-  -npl n0,n1,...                  number of parallel prompts
-
-embedding:
-
-         --embd-normalize         normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-         --embd-output-format     empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-         --embd-separator         separator of embendings (default \n) for example "<#sep#>"
-
 server:
 
          --host HOST              ip address to listen (default: 127.0.0.1)
@@ -267,7 +232,8 @@ server:
                                   https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
   -sps,  --slot-prompt-similarity SIMILARITY
                                   how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-
+         --lora-init-without-apply
+                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
 
 logging:
 
@@ -279,15 +245,6 @@ logging:
          --log-file FNAME         Specify a log filename (without extension)
          --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
          --log-append             Don't truncate the old log file.
-
-cvector:
-
-  -o,    --output FNAME           output file (default: 'control_vector.gguf')
-         --positive-file FNAME    positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
-         --negative-file FNAME    negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
-         --pca-batch N            batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
-         --pca-iter N             number of iterations used for PCA (default: 1000)
-         --method {pca,mean}      dimensionality reduction method to be used (default: pca)
 ```
 
 
@@ -411,7 +368,8 @@ node index.js
 
 ## API Endpoints
 
-- **GET** `/health`: Returns the current state of the server:
+### GET `/health`: Returns the current state of the server
+
   - 503 -> `{"status": "loading model"}` if the model is still being loaded.
   - 500 -> `{"status": "error"}` if the model failed to load.
   - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
@@ -420,7 +378,7 @@ node index.js
 
   If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
-- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -498,7 +456,7 @@ node index.js
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
-### Result JSON
+**Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -537,7 +495,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
-- **POST** `/tokenize`: Tokenize a given text.
+### POST `/tokenize`: Tokenize a given text
 
     *Options:*
 
@@ -545,13 +503,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
-- **POST** `/detokenize`: Convert tokens to text.
+### POST `/detokenize`: Convert tokens to text
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
-- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+### POST `/embedding`: Generate embedding of a given text
+
+The same as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -559,7 +519,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+### POST `/infill`: For code infilling.
+
+Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -571,7 +533,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - **GET** `/props`: Return current server settings.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -589,7 +551,9 @@ Notice that each `probs` is an array of length `n_probs`.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
     *Options:*
 
@@ -641,7 +605,7 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
     *Options:*
 
@@ -675,9 +639,9 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
 
-### Result JSON
+**Response format**
 
 ```json
 [
@@ -738,7 +702,7 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -750,13 +714,13 @@ Available metrics:
 - `llamacpp:requests_processing`: Number of requests processing.
 - `llamacpp:requests_deferred`: Number of requests deferred.
 
-- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
+### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 
     *Options:*
 
     `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -770,13 +734,13 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
+### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
 
     *Options:*
 
     `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -790,9 +754,9 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
+### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -801,6 +765,42 @@ Available metrics:
 }
 ```
 
+### GET `/lora-adapters`: Get list of all LoRA adapters
+
+If an adapter is disabled, the scale will be set to 0.
+
+**Response format**
+
+```json
+[
+    {
+        "id": 0,
+        "path": "my_adapter_1.gguf",
+        "scale": 0.0
+    },
+    {
+        "id": 1,
+        "path": "my_adapter_2.gguf",
+        "scale": 0.0
+    }
+]
+```
+
+### POST `/lora-adapters`: Set list of LoRA adapters
+
+To disable an adapter, either remove it from the list below, or set scale to 0.
+
+**Request format**
+
+To know the `id` of the adapter, use GET `/lora-adapters`
+
+```json
+[
+  {"id": 0, "scale": 0.2},
+  {"id": 1, "scale": 0.8}
+]
+```
+
 ## More examples
 
 ### Change system prompt on runtime

examples/server/server.cpp

@@ -78,6 +78,7 @@ enum server_task_type {
     SERVER_TASK_TYPE_SLOT_SAVE,
     SERVER_TASK_TYPE_SLOT_RESTORE,
     SERVER_TASK_TYPE_SLOT_ERASE,
+    SERVER_TASK_TYPE_SET_LORA,
 };
 
 struct server_task {
@@ -622,6 +623,7 @@ struct server_response {
 struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
 
     gpt_params params;
 
@@ -629,6 +631,7 @@ struct server_context {
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
+    bool has_eos_token  = false;
 
     int32_t n_ctx; // total context for all clients / slots
 
@@ -677,7 +680,11 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+        model = llama_init.model;
+        ctx = llama_init.context;
+        lora_adapters = llama_init.lora_adapters;
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -687,7 +694,7 @@ struct server_context {
         n_ctx = llama_n_ctx(ctx);
 
         add_bos_token = llama_should_add_bos_token(model);
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
+        has_eos_token = llama_add_eos_token(model) != 1;
 
         return true;
     }
@@ -747,13 +754,13 @@ struct server_context {
         default_generation_settings_for_props = get_formated_generation(slots.front());
         default_generation_settings_for_props["seed"] = -1;
 
-        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
         // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
         {
             const int32_t n_batch = llama_n_batch(ctx);
 
             // only a single seq_id per token is needed
-            batch = llama_batch_init(n_batch, 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
         }
 
         metrics.init();
@@ -900,7 +907,7 @@ struct server_context {
 
         slot.params.stream             = json_value(data, "stream",            false);
         slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
         slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -969,6 +976,8 @@ struct server_context {
                 (prompt->is_array() &&  prompt->size() == 1 && prompt->at(0).is_string()) ||
                 (prompt->is_array() && !prompt->empty()     && prompt->at(0).is_number_integer())) {
                 slot.prompt = *prompt;
+            } else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
+                slot.prompt = prompt->at(0);
             } else {
                 send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
                 return false;
@@ -1023,7 +1032,7 @@ struct server_context {
         {
             slot.sparams.logit_bias.clear();
 
-            if (json_value(data, "ignore_eos", false)) {
+            if (json_value(data, "ignore_eos", false) && has_eos_token) {
                 slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
             }
 
@@ -1128,28 +1137,19 @@ struct server_context {
         if (!system_prompt.empty()) {
             system_tokens = ::llama_tokenize(ctx, system_prompt, true);
 
-            llama_batch_clear(batch);
-
-            for (int i = 0; i < (int)system_tokens.size(); ++i) {
-                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-            }
-
             const int32_t n_batch = llama_n_batch(ctx);
+            const int32_t n_tokens_prompt = system_tokens.size();
 
-            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                    0, 0, 0, // unused
-                };
+            for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
+                const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
 
-                if (llama_decode(ctx, batch_view) != 0) {
+                llama_batch_clear(batch);
+
+                for (int32_t j = 0; j < n_tokens; ++j) {
+                    llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
+                }
+
+                if (llama_decode(ctx, batch) != 0) {
                     LOG_ERROR("llama_decode() failed", {});
                     return;
                 }
@@ -1847,6 +1847,14 @@ struct server_context {
                     };
                     queue_results.send(result);
                 } break;
+            case SERVER_TASK_TYPE_SET_LORA:
+                {
+                    llama_lora_adapters_apply(ctx, lora_adapters);
+                    server_task_result result;
+                    result.id = task.id;
+                    result.data = json{{ "success", true }};
+                    queue_results.send(result);
+                } break;
         }
     }
 
@@ -3325,6 +3333,55 @@ int main(int argc, char ** argv) {
         return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
+    const auto handle_lora_adapters_list = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        json result = json::array();
+        for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
+            auto & la = ctx_server.lora_adapters[i];
+            result.push_back({
+                {"id", i},
+                {"path", la.path},
+                {"scale", la.scale},
+            });
+        }
+        res.set_content(result.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
+    const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+
+        const std::vector<json> body = json::parse(req.body);
+        int max_idx = ctx_server.lora_adapters.size();
+
+        // clear existing value
+        for (auto & la : ctx_server.lora_adapters) {
+            la.scale = 0.0f;
+        }
+
+        // set value
+        for (auto entry : body) {
+            int id      = entry.at("id");
+            float scale = entry.at("scale");
+            if (0 <= id && id < max_idx) {
+                ctx_server.lora_adapters[id].scale = scale;
+            } else {
+                throw std::runtime_error("invalid adapter id");
+            }
+        }
+
+        server_task task;
+        task.type = SERVER_TASK_TYPE_SET_LORA;
+        const int id_task = ctx_server.queue_tasks.post(task);
+        ctx_server.queue_results.add_waiting_task_id(id_task);
+
+        server_task_result result = ctx_server.queue_results.recv(id_task);
+        ctx_server.queue_results.remove_waiting_task_id(id_task);
+
+        res.set_content(result.data.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
         return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
             res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
@@ -3363,7 +3420,6 @@ int main(int argc, char ** argv) {
 
     // register API routes
     svr->Get ("/health",              handle_health);
-    svr->Get ("/slots",               handle_slots);
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Get ("/v1/models",           handle_models);
@@ -3378,6 +3434,11 @@ int main(int argc, char ** argv) {
     svr->Post("/v1/embeddings",       handle_embeddings);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
+    // LoRA adapters hotswap
+    svr->Get ("/lora-adapters",       handle_lora_adapters_list);
+    svr->Post("/lora-adapters",       handle_lora_adapters_apply);
+    // Save & load slots
+    svr->Get ("/slots",               handle_slots);
     if (!params.slot_save_path.empty()) {
         // only enable slot endpoints if slot_save_path is set
         svr->Post("/slots/:id_slot",  handle_slots_action);

examples/server/tests/features/lora.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@lora
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
+    And   a model file stories15M_MOE-F16.gguf
+    And   a model alias stories15M_MOE
+    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Completion LoRA disabled
+    Given switch off lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching little|girl|three|years|old
+
+  Scenario: Completion LoRA enabled
+    Given switch on lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching eye|love|glass|sun

examples/server/tests/features/steps/steps.py

@@ -7,6 +7,7 @@ import subprocess
 import sys
 import threading
 import time
+import requests
 from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
@@ -70,6 +71,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.user_api_key = None
     context.response_format = None
     context.temperature = None
+    context.lora_file = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -82,6 +84,12 @@ def step_download_hf_model(context, hf_file: str, hf_repo: str):
     context.model_hf_file = hf_file
     context.model_file = os.path.basename(hf_file)
 
+@step('a lora adapter file from {lora_file_url}')
+def step_download_lora_file(context, lora_file_url: str):
+    file_name = lora_file_url.split('/').pop()
+    context.lora_file = f'../../../{file_name}'
+    with open(context.lora_file, 'wb') as f:
+        f.write(requests.get(lora_file_url).content)
 
 @step('a model file {model_file}')
 def step_model_file(context, model_file: str):
@@ -849,6 +857,17 @@ async def step_erase_slot(context, slot_id):
             context.response = response
 
 
+@step('switch {on_or_off} lora adapter {lora_id:d}')
+@async_run_until_complete
+async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
+    async with aiohttp.ClientSession() as session:
+        async with session.post(f'{context.base_url}/lora-adapters',
+                                json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
+                                headers={"Content-Type": "application/json"}) as response:
+            context.response = response
+            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
+
+
 @step('the server responds with status code {status_code:d}')
 def step_server_responds_with_status_code(context, status_code):
     assert context.response.status == status_code
@@ -1326,6 +1345,8 @@ def start_server_background(context):
         server_args.extend(['--grp-attn-w', context.n_ga_w])
     if context.debug:
         server_args.append('--verbose')
+    if context.lora_file:
+        server_args.extend(['--lora', context.lora_file])
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
 

examples/server/tests/requirements.txt

@@ -4,3 +4,4 @@ huggingface_hub~=0.20.3
 numpy~=1.26.4
 openai~=1.30.3
 prometheus-client~=0.20.0
+requests~=2.32.3

examples/server/utils.hpp

@@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(
 
     llama_params["__oaicompat"] = true;
 
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
     // Apply chat template to the list of messages
     llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
 

examples/simple/README.md

@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 

examples/speculative/speculative.cpp

@@ -66,7 +66,9 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
     params.model = params.model_draft;
@@ -75,7 +77,9 @@ int main(int argc, char ** argv) {
         params.n_threads = params.n_threads_draft;
     }
     params.n_threads_batch = params.n_threads_batch_draft;
-    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
+    model_dft = llama_init_dft.model;
+    ctx_dft = llama_init_dft.context;
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG("vocab_type tgt: %d\n", vocab_type_tgt);

examples/sycl/README.md

@@ -12,9 +12,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 List all SYCL devices with ID, compute capability, max work group size, ect.
 
-1. Build the llama.cpp for SYCL for all targets.
+1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 
-2. Enable oneAPI running environment
+2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
 
 ```
 source /opt/intel/oneapi/setvars.sh
@@ -29,19 +29,13 @@ source /opt/intel/oneapi/setvars.sh
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+found 2 SYCL devices:
+|  |                   |                                       |       |Max    |        |Max  |Global |                     |
+|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
+|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
+|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
+| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29138|
+| 1| [level_zero:gpu:1]|                 Intel UHD Graphics 750|    1.3|     32|     512|   32| 62631M|            1.3.29138|
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|

examples/sycl/win-run-llama2.bat

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1719994518,
-        "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=",
+        "lastModified": 1722555600,
+        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7",
+        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1722062969,
-        "narHash": "sha256-QOS0ykELUmPbrrUGmegAUlpmUFznDQeR4q7rFhl8eQg=",
+        "lastModified": 1723175592,
+        "narHash": "sha256-M0xJ3FbDUc4fRZ84dPGx5VvgFsOzds77KiBMW/mMTnI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "b73c2221a46c13557b1b3be9c2070cc42cf01eb3",
+        "rev": "5e0ca22929f3342b19569b21b2f3462f053e497b",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1719876945,
-        "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=",
+        "lastModified": 1722555339,
+        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       }
     },
     "root": {

ggml/CMakeLists.txt

@@ -129,13 +129,13 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
 
-option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)

ggml/include/ggml-metal.h

@@ -50,6 +50,8 @@ GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
 
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

ggml/include/ggml.h

@@ -244,6 +244,8 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 #define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
@@ -1140,16 +1142,17 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     GGML_API struct ggml_tensor * ggml_group_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1452,11 +1455,10 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
-    // if mode & 2 == 1, GPT-NeoX style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
-    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1473,6 +1475,7 @@ extern "C" {
             int                   mode);
 
     // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/CMakeLists.txt

@@ -602,6 +602,10 @@ if (GGML_VULKAN)
             add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
         endif()
 
+        if (GGML_VULKAN_PERF)
+            add_compile_definitions(GGML_VULKAN_PERF)
+        endif()
+
         if (GGML_VULKAN_VALIDATE)
             add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()

ggml/src/ggml-aarch64.c

@@ -16,6 +16,8 @@
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED GGML_UNUSED

ggml/src/ggml-backend.c

@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
@@ -1782,7 +1777,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 

ggml/src/ggml-cann.cpp

@@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
 GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
                                                        const void* src,
                                                        void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
  */
 GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
     const ggml_tensor* tensor, void* src, void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -898,11 +896,10 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
  * @param size Size of the data to be copied, in bytes.
  */
 GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor* tensor, const void* data,
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
     size_t offset, size_t size) {
-    // GGML_ASSERT(size == ggml_nbytes(tensor));
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
 
     ggml_cann_set_device(ctx->device);
     // TODO: refer to cann(#6017), it use thread's default stream.
@@ -910,22 +907,21 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, (const char*)data + offset,
-                              size, ACL_MEMCPY_HOST_TO_DEVICE));
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size) ==
-                    0);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size,
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         free(transform_buffer);
     }
@@ -947,21 +943,20 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
 GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
     size_t offset, size_t size) {
-    GGML_ASSERT(size == ggml_nbytes(tensor));
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
 
     ggml_cann_set_device(ctx->device);
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char*)data + offset, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
     } else {
         void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1450,42 +1445,41 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
  * @param size Size of the data to copy in bytes.
  */
 GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                                         ggml_tensor* tensor,
-                                                         const void* data,
+                                                         ggml_tensor *tensor,
+                                                         const void *data,
                                                          size_t offset,
                                                          size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(
-            tensor->data, size, (const char*)data + offset, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size));
+        GGML_ASSERT(memcmp(data, check_buffer, size));
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpyAsync(tensor->data, size, transform_buffer, size,
-                                   ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
         free(transform_buffer);
     }
 }
 
 GGML_CALL static void ggml_backend_cann_get_tensor_async(
-    ggml_backend_t backend, const ggml_tensor* tensor, void* data,
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
     ggml_backend_buffer_t buf =
         tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -1493,17 +1487,16 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
                 "unsupported buffer type");
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char*)data + offset, size, tensor->data,
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
                                    size, ACL_MEMCPY_DEVICE_TO_HOST,
                                    cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(transform_buffer, size, tensor->data, size,
-                                   ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1666,10 +1659,13 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                // case GGML_TYPE_Q4_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1694,6 +1690,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;

ggml/src/ggml-cann/acl_tensor.cpp

@@ -37,6 +37,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT16;
         case GGML_TYPE_I32:
             return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
         default:
             return ACL_DT_UNDEFINED;
     }
@@ -89,33 +93,6 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   size_t type_size, int64_t* ne, size_t* nb,
-                                   int64_t dims, aclFormat format,
-                                   size_t offset) {
-    int64_t tmp_ne[GGML_MAX_DIMS * 2];
-    int64_t tmp_stride[GGML_MAX_DIMS * 2];
-
-    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
-    for (int i = 0; i < dims; i++) {
-        tmp_stride[i] = nb[i] / type_size;
-    }
-
-    std::reverse(tmp_ne, tmp_ne + dims);
-    std::reverse(tmp_stride, tmp_stride + dims);
-
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
-
-    return acl_tensor;
-}
-
 int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
                                   const ggml_tensor* src1,
                                   int64_t* bcast_src0_ne,

ggml/src/ggml-cann/acl_tensor.h

@@ -23,6 +23,9 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
+#include <algorithm>
+#include <cstring>
+
 #include <aclnn/aclnn_base.h>
 #include "common.h"
 
@@ -65,7 +68,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                              size_t offset = 0);
 
 /**
- * @brief   Creates an ACL tensor from provided parameters.
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
  *
  * @details This function creates an ACL tensor using the provided data pointer,
  *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -83,10 +87,34 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
+template<typename TYPE>
 aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                             size_t type_size, int64_t* ne, size_t* nb,
-                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}
 
 /**
  * @brief   Checks if tensors require broadcasting based on their shapes.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -464,9 +464,11 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    const float eps = 1e-6f;  // TODO: make this a parameter
     int n_groups = dst->op_params[0];
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -910,6 +912,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F16) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -971,6 +980,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -2463,21 +2479,33 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst) {
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
     // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
     // is regarded as batch. weight need transpose.
     int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
-    size_t weight_elem_size = sizeof(uint8_t);
-    size_t weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
     // size of one matrix is element_size * height * width.
     size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
     int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
@@ -2541,8 +2569,9 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
-                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
-                weight_elem_size, weight_ne, weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
                 scale_elem_size, scale_ne, scale_nb, 2);
@@ -2596,11 +2625,9 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16:
             ggml_cann_mat_mul_fp(ctx, dst);
             break;
-        // case GGML_TYPE_Q4_0:
-        //     ggml_cann_mul_mat_q4_0(ctx, dst);
-        //     break;
+        case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_q8_0(ctx, dst);
+            ggml_cann_mul_mat_quant(ctx, dst, type);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -2854,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
                              beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     // init cos/sin cache
     ggml_cann_pool_alloc sin_allocator(

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -9,6 +9,7 @@ file(GLOB SRC_FILES
     get_row_q8_0.cpp
     quantize_f32_q8_0.cpp
     quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
     dup.cpp
 )
 
@@ -29,4 +30,4 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/ascendc_kernels.h

@@ -8,6 +8,8 @@
 
 #include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
 #include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
+#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
 
 #include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
 #include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"

ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -0,0 +1,278 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define Group_Size 32
+
+template <typename SRC_T>
+class QUANTIZE_FLOAT_TO_Q4_0 {
+   public:
+    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
+        //                         permute=[0,0,0,0]):
+        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // input stride of data elements
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        // output stride of data elements
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / Group_Size;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t) / 2;
+
+        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
+        pipe.InitBuffer(output_queue, BUFFER_NUM,
+                            Group_Size * sizeof(int8_t) / 2);
+        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
+        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
+        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
+        DataCopy(input_local, input_gm[offset], Group_Size);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
+        // and using DataCopyPad to avoid 32 bits align.
+        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
+        LocalTensor<int8_t> output_int8_local =
+                                    output_local.ReinterpretCast<int8_t>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
+        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
+
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<float> input_local) {
+        DataCopy(cast_local, input_local, Group_Size);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<half> input_local) {
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + Group_Size * group;
+
+        // output_offset is stride for output_gm which datatype is int8_t and
+        // divided by 2 is needed for int4b_t.
+        const int64_t output_offset = (i1 * output_stride[1] +
+                                       i2 * output_stride[2] +
+                                       i3 * output_stride[3] +
+                                       Group_Size * group) / 2;
+        copy_in(input_offset);
+
+        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
+        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
+        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
+        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
+
+        input_to_cast(cast_local, input_local);
+
+        ReduceMax(max_local, cast_local, work_local, Group_Size);
+        ReduceMin(min_local, cast_local, work_local, Group_Size);
+        const float max_value = max_local.GetValue(0);
+        const float min_value = min_local.GetValue(0);
+        float d = max_value;
+        if (min_value < 0 && (-1 * min_value) > max_value) {
+            d = min_value;
+        }
+
+        d = d / (-8);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
+        }
+
+        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
+        float scalar = 8.5f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
+        scalar = 15.0f;
+        Mins(cast_local, cast_local, scalar, Group_Size);
+        scalar = -8.0f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+
+        // float->half->int4b
+        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
+        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
+
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        max_queue.FreeTensor(max_local);
+        min_queue.FreeTensor(min_local);
+        int8_queue.FreeTensor(int8_local);
+        half_queue.FreeTensor(half_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                // Copy Group_Size/2 length data each time.
+                if (scale_local_offset == Group_Size / 2) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local,
+                                      Group_Size / 2);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += Group_Size / 2;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+        scale_queue.FreeTensor(scale_local);
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<half> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<float> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}

ggml/src/ggml-cuda.cu

@@ -1501,7 +1501,7 @@ static void ggml_cuda_op_mul_mat(
         }
 
         // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
@@ -2358,33 +2358,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
 }
 
 GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
-
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
-    // device -> device
+    // device -> device copy
     ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
     ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
 
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+#ifndef NDEBUG
+        GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
+#endif
+        return false;
+    }
+
     if (backend_src != backend_dst) {
-        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
-
-        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
-        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
-
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
-            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
         } else {
 #ifdef GGML_CUDA_NO_PEER_COPY
             return false;
@@ -2393,7 +2395,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
 #endif
         }
 
-        // record event on src stream
+        // record event on src stream after the copy
         if (!cuda_ctx_src->copy_event) {
             ggml_cuda_set_device(cuda_ctx_src->device);
             CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2405,7 +2407,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
         CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
     } else {
         // src and dst are on the same backend
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
     }
     return true;
 }
@@ -2742,11 +2744,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_MUL_MAT_ID:
             {
                 struct ggml_tensor * a = op->src[0];
-                if (op->op == GGML_OP_MUL_MAT) {
-                    struct ggml_tensor * b = op->src[1];
-                    if (a->ne[3] != b->ne[3]) {
-                        return false;
-                    }
+                struct ggml_tensor * b = op->src[1];
+                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
+                    return false;
                 }
                 switch (a->type) {
                     case GGML_TYPE_F32:
@@ -2877,7 +2880,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
 #else
             if (op->src[0]->ne[0] == 128) {
                 return true;

ggml/src/ggml-cuda/norm.cu

@@ -142,8 +142,7 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     }
 }
 
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
-    static const float eps = 1e-6f;
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
     if (group_size < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@@ -196,8 +195,12 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
 }
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/rope.cu

@@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const int32_t * pos = (const int32_t *) src1_d;
 

ggml/src/ggml-metal.m

@@ -210,7 +210,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
-struct ggml_metal_context {
+struct ggml_backend_metal_context {
     int n_cb;
 
     id<MTLDevice>       device;
@@ -224,6 +224,10 @@ struct ggml_metal_context {
     bool support_simdgroup_mm;
 
     bool should_capture_next_compute;
+
+    // abort ggml_metal_graph_compute if callback returns true
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
 };
 
 // MSL code
@@ -289,7 +293,7 @@ static void * ggml_metal_host_malloc(size_t n) {
     return data;
 }
 
-static struct ggml_metal_context * ggml_metal_init(int n_cb) {
+static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
 #if TARGET_OS_OSX && !GGML_METAL_NDEBUG
@@ -306,7 +310,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -668,7 +672,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     return ctx;
 }
 
-static void ggml_metal_free(struct ggml_metal_context * ctx) {
+static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
     for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
@@ -734,7 +738,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
     return nil;
 }
 
-static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
+static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
     for (size_t i = 0, n = 3; i < n; ++i) {
         if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
             return false;
@@ -845,7 +849,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
 }
 
 static enum ggml_status ggml_metal_graph_compute(
-        struct ggml_metal_context * ctx,
+        struct ggml_backend_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
     @autoreleasepool {
@@ -878,8 +882,11 @@ static enum ggml_status ggml_metal_graph_compute(
         id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBufferWithUnretainedReferences];
         command_buffer_builder[cb_idx] = command_buffer;
 
-        // enqueue the command buffers in order to specify their execution order
-        [command_buffer enqueue];
+        // always enqueue the first two command buffers
+        // enqueue all of the command buffers if we don't need to abort
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer enqueue];
+        }
     }
 
     const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
@@ -2229,10 +2236,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(ne00 % 4 == 0);
                         GGML_ASSERT(ggml_is_contiguous(src0));
 
-                        //float eps;
-                        //memcpy(&eps, dst->op_params, sizeof(float));
-
-                        const float eps = 1e-6f; // TODO: temporarily hardcoded
+                        float eps;
+                        memcpy(&eps, dst->op_params + 1, sizeof(float));
 
                         const int32_t n_groups = ((int32_t *) dst->op_params)[0];
 
@@ -2308,7 +2313,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-                        const bool is_neox = mode & 2;
+                        const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
                         id<MTLComputePipelineState> pipeline = nil;
 
@@ -2829,7 +2834,9 @@ static enum ggml_status ggml_metal_graph_compute(
 
         [encoder endEncoding];
 
-        [command_buffer commit];
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer commit];
+        }
     });
 
     // Wait for completion and check status of each command buffer
@@ -2849,6 +2856,23 @@ static enum ggml_status ggml_metal_graph_compute(
 
             return GGML_STATUS_FAILED;
         }
+
+        id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil);
+        if (!next_buffer) {
+            continue;
+        }
+
+        bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
+        if (next_queued) {
+            continue;
+        }
+
+        if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
+            GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i);
+            return GGML_STATUS_ABORTED;
+        }
+
+        [next_buffer commit];
     }
 
     if (should_capture) {
@@ -3152,7 +3176,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ggml_metal_free(ctx);
     free(backend);
 }
@@ -3164,13 +3188,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
 }
 
 GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
 GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_supports_op(metal_ctx, op);
 }
@@ -3215,9 +3239,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) {
 }
 
 ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
-
+    struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
     if (ctx == NULL) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return NULL;
     }
 
@@ -3239,15 +3263,24 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
 }
 
+void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = user_data;
+}
+
 bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
@@ -3255,7 +3288,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ctx->should_capture_next_compute = true;
 }
 

ggml/src/ggml-rpc.cpp

@@ -197,6 +197,10 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
         fprintf(stderr, "Failed to set SO_REUSEADDR\n");
         return nullptr;
     }
+    if (inet_addr(host) == INADDR_NONE) {
+        fprintf(stderr, "Invalid host address: %s\n", host);
+        return nullptr;
+    }
     struct sockaddr_in serv_addr;
     serv_addr.sin_family = AF_INET;
     serv_addr.sin_addr.s_addr = inet_addr(host);
@@ -879,6 +883,14 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
     if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
         return nullptr;
     }
+
+    // require that the tensor data does not go beyond the buffer end
+    uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
+    uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+    uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+    GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
+    GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+
     result->op = (ggml_op) tensor->op;
     for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
         result->op_params[i] = tensor->op_params[i];
@@ -898,7 +910,7 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
     const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
     uint64_t offset;
     memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
+    const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
 
     struct ggml_init_params params {
         /*.mem_size   =*/ ggml_tensor_overhead(),
@@ -913,6 +925,17 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
     ggml_backend_tensor_set(tensor, data, offset, size);
     ggml_free(ctx);
@@ -943,6 +966,17 @@ bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     // output serialization format: | data (size bytes) |
     output.resize(size, 0);
     ggml_backend_tensor_get(tensor, output.data(), offset, size);

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -874,7 +874,7 @@ namespace dpct
         inline std::string get_preferred_gpu_platform_name() {
             std::string result;
 
-            std::string filter = "level-zero";
+            std::string filter = "";
             char* env = getenv("ONEAPI_DEVICE_SELECTOR");
             if (env) {
                 if (std::strstr(env, "level_zero")) {
@@ -892,11 +892,24 @@ namespace dpct
                 else {
                     throw std::runtime_error("invalid device filter: " + std::string(env));
                 }
+            } else {
+                auto default_device = sycl::device(sycl::default_selector_v);
+                auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
+
+                if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "HIP")) {
+                    filter = "hip";
+                }
             }
 
-            auto plaform_list = sycl::platform::get_platforms();
+            auto platform_list = sycl::platform::get_platforms();
 
-            for (const auto& platform : plaform_list) {
+            for (const auto& platform : platform_list) {
                 auto devices = platform.get_devices();
                 auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
                     return d.is_gpu();

ggml/src/ggml-sycl/norm.cpp

@@ -225,9 +225,8 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
 }
 
 static void group_norm_f32_sycl(const float* x, float* dst,
-    const int num_groups, const int group_size,
+    const int num_groups, const float eps, const int group_size,
     const int ne_elements, queue_ptr stream, int device) {
-    static const float eps = 1e-6f;
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
@@ -343,8 +342,12 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor*
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
+    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
 
     (void)src1;
     (void)dst;

ggml/src/ggml-sycl/rope.cpp

@@ -226,7 +226,7 @@ void ggml_sycl_op_rope(
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const int32_t * pos = (const int32_t *) src1_dd;
 

ggml/src/ggml.c

@@ -56,6 +56,9 @@ int ggml_sve_cnt_b = 0;
 // disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
+
+// unreachable code because of multiple instances of code after GGML_ABORT
+#pragma warning(disable: 4702)
 #endif
 
 #if defined(_WIN32)
@@ -2312,7 +2315,7 @@ inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) {
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
-inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
+inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
@@ -3721,7 +3724,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_tensor  * view_src,
         size_t                view_offs) {
 
-    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
+    GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
     // find the base tensor and absolute offset
     if (view_src != NULL && view_src->view_src != NULL) {
@@ -5374,6 +5378,7 @@ static struct ggml_tensor * ggml_group_norm_impl(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
     int n_groups,
+    float eps,
     bool inplace) {
 
     bool is_node = false;
@@ -5384,7 +5389,8 @@ static struct ggml_tensor * ggml_group_norm_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op_params[0] = n_groups;
+    ggml_set_op_params_i32(result, 0, n_groups);
+    ggml_set_op_params_f32(result, 1, eps);
 
     result->op = GGML_OP_GROUP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5396,15 +5402,17 @@ static struct ggml_tensor * ggml_group_norm_impl(
 struct ggml_tensor * ggml_group_norm(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, false);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, false);
 }
 
 struct ggml_tensor * ggml_group_norm_inplace(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, true);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
 }
 
 // ggml_mul_mat
@@ -12095,10 +12103,11 @@ static void ggml_compute_forward_group_norm_f32(
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    const float eps = 1e-6f; // TODO: make this a parameter
-
     // TODO: optimize
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int n_channels = src0->ne[2];
     int n_groups = dst->op_params[0];
     int n_channels_per_group = (n_channels + n_groups - 1) / n_groups;
@@ -14085,7 +14094,7 @@ static void ggml_compute_forward_rope_f32(
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -14210,7 +14219,7 @@ static void ggml_compute_forward_rope_f16(
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -21120,7 +21129,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[2] *
                 (int64_t) info->ne[3];
 
-            if (ne % ggml_blck_size(info->type) != 0) {
+            if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
                 fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                         __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);

ggml/src/kompute-shaders/op_rope_f16.comp

@@ -11,7 +11,7 @@ void main() {
     const uint i2 = gl_WorkGroupID.y;
     const uint i1 = gl_WorkGroupID.x;
 
-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
 
     float corr_dims[2];
     rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

ggml/src/kompute-shaders/op_rope_f32.comp

@@ -11,7 +11,7 @@ void main() {
     const uint i2 = gl_WorkGroupID.y;
     const uint i1 = gl_WorkGroupID.x;
 
-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
 
     float corr_dims[2];
     rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

ggml/src/kompute-shaders/rope_common.comp

@@ -1,5 +1,7 @@
 #include "common.comp"
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 // TODO: use a local size of 32 or more (Metal uses 1024)
 layout(local_size_x = 1) in;
 

ggml/src/vulkan-shaders/CMakeLists.txt

@@ -1,5 +1,7 @@
+find_package (Threads REQUIRED)
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)

ggml/src/vulkan-shaders/add.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/clamp.comp

@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }

ggml/src/vulkan-shaders/concat.comp

@@ -0,0 +1,39 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    const int dim = p.param3;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
+    const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
+    const uint i2_offset = i2*p.ne21*p.ne20;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
+
+    uint o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
+    const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+#else
+    if (is_src0) {
+        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+    } else {
+        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+    }
+#endif
+}

ggml/src/vulkan-shaders/copy.comp

@@ -4,13 +4,15 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
+    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
 #endif
 }

ggml/src/vulkan-shaders/div.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) / FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/gelu.comp

@@ -13,7 +13,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 void main() {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;

ggml/src/vulkan-shaders/gelu_quick.comp

@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_QUICK_COEF = -1.702f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * (1.0f / (1.0f + exp(GELU_QUICK_COEF * x))));
+}

ggml/src/vulkan-shaders/generic_binary_head.comp

@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint d_offset;
-    float param1; float param2;
+    float param1; float param2; int param3;
 } p;
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
@@ -16,6 +16,10 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;

ggml/src/vulkan-shaders/generic_unary_head.comp

@@ -14,6 +14,10 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;

ggml/src/vulkan-shaders/group_norm.comp

@@ -0,0 +1,66 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared float tmp[BLOCK_SIZE];
+
+void main() {
+    const uint group_size = p.KX;
+    const float eps = p.param1;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint start = gl_WorkGroupID.x * group_size + tid;
+    const uint end = start + group_size;
+
+    tmp[tid] = 0.0f;
+
+    // Calculate mean
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        tmp[tid] += float(data_a[col]);
+    }
+
+    // tmp up partial tmps and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = tmp[0] / group_size;
+    barrier();
+    tmp[tid] = 0.0f;
+
+    // Calculate variance
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        const float xi = float(data_a[col]) - mean;
+        data_d[col] = D_TYPE(xi);
+        tmp[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float variance = tmp[0] / group_size;
+    const float scale = inversesqrt(variance + eps);
+
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        data_d[col] *= D_TYPE(scale);
+    }
+}

ggml/src/vulkan-shaders/im2col.comp

@@ -0,0 +1,57 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint batch_offset; uint offset_delta;
+    uint IC;
+    uint IW; uint IH;
+    uint OW; uint OH;
+    uint KW; uint KH;
+    uint pelements;
+    uint CHW;
+    int s0; int s1;
+    int p0; int p1;
+    int d0; int d1;
+} p;
+
+#include "types.comp"
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+    if (i >= p.pelements) {
+        return;
+    }
+
+    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint kx = i / ksize;
+    const uint kd = kx * ksize;
+    const uint ky = (i - kd) / p.OW;
+    const uint ix = i % p.OW;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+    const uint offset_dst =
+        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+        (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
+        data_d[offset_dst] = D_TYPE(0.0f);
+    } else {
+        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    }
+}

ggml/src/vulkan-shaders/leaky_relu.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float val = float(data_a[i]);
+    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
+}

ggml/src/vulkan-shaders/mul.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/mul_mat_vec.comp

@@ -16,6 +16,13 @@ void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
     const uint tid = gl_LocalInvocationID.x;
 
+    // There are not enough cols to use all threads
+    if (tid >= p.ncols) {
+        return;
+    }
+
+    const uint block_size = min(p.ncols, BLOCK_SIZE);
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -23,8 +30,8 @@ void main() {
 
     tmp[tid] = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
-        const uint col = i*BLOCK_SIZE + 2*tid;
+    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
+        const uint col = i*block_size + 2*tid;
         const uint ib = (row*p.ncols + col)/QUANT_K; // block index
         const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
         const uint iybs = col - col%QUANT_K; // y block start index
@@ -32,13 +39,12 @@ void main() {
         vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
 
         // matrix multiplication
-        tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) +
-                    FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
     }
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/vulkan-shaders/mul_mat_vec_nc.comp

@@ -53,7 +53,7 @@ void main() {
 
         const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 
-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_p021.comp

@@ -52,7 +52,7 @@ void main() {
         // y is not transposed but permuted
         const uint iy = channel*nrows_y + row_y;
 
-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
     }
 
     // dst is not transposed and not permuted