server : fix incorrect usage of llama_get_embeddings() (#14225)

* server : fix incorrect usage of llama_get_embeddings()

ggml-ci

* cont : fix the fix

ggml-ci

.github/workflows/build.yml

@@ -693,7 +693,7 @@ jobs:
           - build: 'openblas-x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=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'
           - build: 'llvm-arm64-opencl-adreno'
@@ -778,6 +778,7 @@ jobs:
           cmake -S . -B build ${{ matrix.defines }} `
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll

ci/run.sh

@@ -39,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"

common/arg.cpp

@@ -988,10 +988,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
-    if (params.reranking && params.embedding) {
-        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
-    }
-
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -2747,9 +2743,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
         [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(

common/common.cpp

@@ -767,6 +767,9 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
+#include <iostream>
+
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -784,9 +787,16 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // process path from front to back, procedurally creating directories
     while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
         const std::wstring subpath = wpath.substr(0, pos_slash);
-        const wchar_t * test = subpath.c_str();
 
-        const bool success = CreateDirectoryW(test, NULL);
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -800,8 +810,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
-
-        pos_slash += 1;
     }
 
     return true;
@@ -897,34 +905,6 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    if (params.reranking) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -966,6 +946,35 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
+    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1143,11 +1152,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
 
-    if (params.reranking) {
-        cparams.embeddings    = true;
-        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
-    }
-
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
 

common/common.h

@@ -355,7 +355,6 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port

convert_hf_to_gguf.py

@@ -519,7 +519,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -2020,6 +2020,20 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("ArceeForCausalLM")
+class ArceeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.ARCEE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self._try_set_pooling_type()
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -4062,6 +4076,34 @@ class NomicBertModel(BertModel):
         raise ValueError(f"unknown tokenizer: {toktyp}")
 
 
+@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
+class NeoBert(BertModel):
+    model_arch = gguf.MODEL_ARCH.NEO_BERT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # NeoBERT uses 2/3 of the intermediate size as feed forward length
+        self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
+        self.gguf_writer.add_rope_freq_base(10000.0)  # default value for NeoBERT
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+        f_rms_eps = self.hparams.get("norm_eps", 1e-6)  # default value for NeoBERT
+        self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
+
+        self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name.startswith("decoder."):
+            return []
+
+        if name.startswith("model."):
+            name = name[6:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT

examples/gritlm/gritlm.cpp

@@ -41,12 +41,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_memory_clear(llama_get_memory(ctx), true);
-        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -103,7 +102,6 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     llama_token eos_token = llama_vocab_eos(vocab);
 
     llama_memory_clear(llama_get_memory(ctx), true);
-    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -166,6 +164,8 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
+    cparams.embeddings = true;
+
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,6 +213,8 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
+    llama_set_embeddings(ctx, false);
+
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {

ggml/CMakeLists.txt

@@ -172,6 +172,7 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
+option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    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)

ggml/cmake/common.cmake

@@ -36,8 +36,7 @@ function(ggml_get_system_arch)
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
         set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
-            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)

ggml/src/CMakeLists.txt

@@ -311,18 +311,28 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
-    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
-        # Many of these features are optional so we build versions with popular
-        # combinations and name the backends based on the version they were
-        # first released with
-        ggml_add_cpu_backend_variant(armv8.0_1)
-        ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
-        ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-        ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
-        ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
-        ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
-        ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
-        ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            # Many of these features are optional so we build versions with popular
+            # combinations and name the backends based on the version they were
+            # first released with
+            ggml_add_cpu_backend_variant(armv8.0_1)
+            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
+            # Android-specific backends with SoC-compatible feature sets
+            ggml_add_cpu_backend_variant(android_armv8.0_1)
+            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
+        else()
+            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()

ggml/src/ggml-cpu/CMakeLists.txt

@@ -158,48 +158,45 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
                 elseif(GGML_CPU_ALL_VARIANTS)
-                    if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-                        # Begin with the lowest baseline
-                        set(ARM_MCPU "armv8-a")
-                        set(ARCH_TAGS "")
-                        set(ARCH_DEFINITIONS "")
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+                    set(ARCH_DEFINITIONS "")
 
-                        # When a feature is selected, bump the MCPU to the first
-                        # version that supported it
-                        if (GGML_INTERNAL_DOTPROD)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
-                        endif()
-                        if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+fp16")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
-                        endif()
-                        if (GGML_INTERNAL_SVE)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sve")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
-                        endif()
-                        if (GGML_INTERNAL_MATMUL_INT8)
-                            set(ARM_MCPU "armv8.6-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
-                        endif()
-                        if (GGML_INTERNAL_SVE2)
-                            set(ARM_MCPU "armv8.6-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sve2")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
-                        endif()
-                        if (GGML_INTERNAL_SME)
-                            set(ARM_MCPU "armv9.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sme")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
-                        endif()
-
-                        list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
-                        ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
                     endif()
+                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                    endif()
+                    if (GGML_INTERNAL_SVE)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                    endif()
+                    if (GGML_INTERNAL_MATMUL_INT8)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                    endif()
+                    if (GGML_INTERNAL_SVE2)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                    endif()
+                    if (GGML_INTERNAL_SME)
+                        set(ARM_MCPU "armv9.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                    endif()
+                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                 endif()
             endif()
 

ggml/src/ggml-cpu/apple-fallback.h

@@ -0,0 +1,88 @@
+#pragma once
+
+// Solve alias issue for Apple targets (currently PowerPC, x86, and ARM64).
+// Mach-O has a weak alias equivalent but no practical compiler support can
+// be found, so we need to do it manually.
+// ref: https://stackoverflow.com/questions/42757744
+//
+// This file is a complement to native implementations in the `arch` folder.
+// A kernel in quants.c or repack.cpp is either:
+// - implemented in the `arch` folder, or
+// - defined in this file to remove the `_generic` suffix
+
+#if defined(GGML_CPU_GENERIC)
+// quants.c
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__aarch64__) || defined(__arm__)
+// repack.cpp
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#elif defined(__x86_64__) || defined(__i386__)
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__POWERPC__)
+// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -503,13 +503,16 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
+int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
+
 #ifdef __cplusplus
 }
 #endif
 
 #define GGML_DO_PRAGMA_(x) _Pragma (#x)
 #define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
-#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__) || defined(__APPLE__)
 // Note for Apple targets:
 // - clang: aliases are not supported on darwin
 // - all native kernels need to be implemented in both x86 and arm files

ggml/src/ggml-cpu/ggml-cpu.c

@@ -559,6 +559,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
+    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
+    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -53,7 +53,6 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
-#include <atomic>
 #include <array>
 #include <type_traits>
 
@@ -394,8 +393,6 @@ class tinyBLAS {
 
     template <int RM, int RN, int BM>
     NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
-        static std::atomic<int64_t> current_chunk;
-
         GGML_ASSERT(m % (RM * BM) == 0);
         const int64_t ytiles = m / (RM * BM);
         const int64_t xtiles = (n + RN -1) / RN;
@@ -410,7 +407,7 @@ class tinyBLAS {
         if (params->ith == 0) {
             GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
             // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
         }
 
         ggml_barrier(params->threadpool);
@@ -439,8 +436,7 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
 
         ggml_barrier(params->threadpool);

ggml/src/ggml-cpu/quants.c

@@ -5,6 +5,10 @@
 #include "ggml-quants.h"
 #include "quants.h"
 
+#if defined(__APPLE__)
+#include "apple-fallback.h"
+#endif
+
 #include <string.h>
 #include <assert.h>
 #include <float.h>

ggml/src/ggml-cpu/quants.h

@@ -84,33 +84,6 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
-#if defined(GGML_CPU_GENERIC)
-#define quantize_row_q8_0_generic quantize_row_q8_0
-#define quantize_row_q8_1_generic quantize_row_q8_1
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
-#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
-#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
-#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-#endif
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/repack.cpp

@@ -8,6 +8,10 @@
 #include "ggml-cpu-impl.h"
 #include "traits.h"
 
+#if defined(__APPLE__)
+#include "apple-fallback.h"
+#endif
+
 #include <cmath>
 #include <cstring>
 #include <cassert>

ggml/src/ggml-cpu/repack.h

@@ -67,7 +67,7 @@ extern "C" {
 // Workaround for clang:
 // clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
 // repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
-#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || defined(__APPLE__) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -98,22 +98,6 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
-#if defined(GGML_CPU_GENERIC)
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#endif
-
 #if defined(__cplusplus)
 } // extern "C"
 #endif

ggml/src/ggml-cuda/common.cuh

@@ -207,9 +207,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE

ggml/src/ggml-cuda/ssm-scan.cu

@@ -10,6 +10,8 @@ __global__ void __launch_bounds__(splitD, 2)
                  float * __restrict__ dst, const int64_t L) {
     GGML_UNUSED(src1_nb0);
     GGML_UNUSED(src2_nb0);
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     const int bidx = blockIdx.x;  // split along B
     const int bidy = blockIdx.y;  // split along D
     const int tid  = threadIdx.x;
@@ -44,16 +46,16 @@ __global__ void __launch_bounds__(splitD, 2)
     if (N == 16) {
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
+            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
             // todo: bank conflict
             // I am always confused with how to use the swizzling method to solve
             // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
-            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
+            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
+    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -49,15 +49,7 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    set(VULKAN_SHADER_GEN_CMAKE_ARGS
-        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
-        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
-    )
-
-    set(VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS "")
-    if (CMAKE_BUILD_TYPE AND CMAKE_BUILD_TYPE MATCHES "Debug|Release|MinSizeRel|RelWithDebInfo")
-        list(APPEND VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS --config=${CMAKE_BUILD_TYPE})
-    endif()
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
 
     # Test all shader extensions
     test_shader_extension_support(
@@ -136,42 +128,39 @@ if (Vulkan_FOUND)
         set(HOST_CMAKE_TOOLCHAIN_FILE "")
     endif()
 
-    # Always use ExternalProject_Add approach
     include(ExternalProject)
 
-    # Add toolchain file if cross-compiling
     if (CMAKE_CROSSCOMPILING)
         list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE})
         message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
     endif()
 
-    # Native build through ExternalProject_Add
     ExternalProject_Add(
         vulkan-shaders-gen
         SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
-        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
-        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
-        INSTALL_DIR ${CMAKE_BINARY_DIR}
+        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/$<CONFIG>
+                   -DCMAKE_INSTALL_BINDIR=.
+                   -DCMAKE_BUILD_TYPE=$<CONFIG>
+                   ${VULKAN_SHADER_GEN_CMAKE_ARGS}
+
+        BUILD_COMMAND   ${CMAKE_COMMAND} --build   . --config $<CONFIG>
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install . --config $<CONFIG>
     )
     ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
 
     set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>)
-    set (_ggml_vk_genshaders_cmd ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/vulkan-shaders-gen${_ggml_vk_host_suffix})
-    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
-    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
-    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
-    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+    set (_ggml_vk_genshaders_dir "${CMAKE_BINARY_DIR}/$<CONFIG>")
+    set (_ggml_vk_genshaders_cmd "${_ggml_vk_genshaders_dir}/vulkan-shaders-gen${_ggml_vk_host_suffix}")
+    set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
+    set (_ggml_vk_source     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp")
+    set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
+    set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
 
-    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
-    set (_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen)
-
-    # Add build and install dependencies for all builds
-    set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
+    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
-                ${_ggml_vk_source}
+               ${_ggml_vk_source}
 
         COMMAND ${_ggml_vk_genshaders_cmd}
             --glslc      ${Vulkan_GLSLC_EXECUTABLE}
@@ -181,7 +170,11 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps}
+        DEPENDS ${_ggml_vk_shader_files}
+                vulkan-shaders-gen
+                vulkan-shaders-gen-build
+                vulkan-shaders-gen-install
+
         COMMENT "Generate vulkan shaders"
     )
 

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

@@ -168,6 +168,11 @@ struct vk_command_pool {
     vk_queue *q;
 };
 
+// Prevent simultaneous submissions to the same queue.
+// This could be per vk_queue if we stopped having two vk_queue structures
+// sharing the same vk::Queue.
+static std::mutex queue_mutex;
+
 struct vk_queue {
     uint32_t queue_family_index;
     vk::Queue queue;
@@ -1266,6 +1271,7 @@ static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_command
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         if (fence) {
+            std::lock_guard<std::mutex> guard(queue_mutex);
             ctx->p->q->queue.submit({}, fence);
         }
         return;
@@ -1335,6 +1341,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
         }
     }
 
+    std::lock_guard<std::mutex> guard(queue_mutex);
     ctx->p->q->queue.submit(submit_infos, fence);
 
     ctx->seqs.clear();

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -25,15 +25,3 @@ add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
-
-# Configure output directories for MSVC builds
-if(MSVC)
-    # Get the main project's runtime output directory if possible
-    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
-        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
-            string(TOUPPER ${CONFIG} CONFIG)
-            set_target_properties(${TARGET} PROPERTIES
-                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-        endforeach()
-    endif()
-endif()

gguf-py/gguf/constants.py

@@ -291,6 +291,7 @@ class MODEL_ARCH(IntEnum):
     BERT             = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
+    NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -344,6 +345,7 @@ class MODEL_ARCH(IntEnum):
     PLM              = auto()
     BAILINGMOE       = auto()
     DOTS1            = auto()
+    ARCEE            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -572,6 +574,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
     MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
+    MODEL_ARCH.NEO_BERT:         "neo-bert",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -624,7 +627,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
-    MODEL_ARCH.DOTS1:            "dots1"
+    MODEL_ARCH.DOTS1:            "dots1",
+    MODEL_ARCH.ARCEE:            "arcee",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1079,6 +1083,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NEO_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2070,6 +2086,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.ARCEE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -271,7 +271,7 @@ class GGUFWriter:
 
     def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
         if any(key in kv_data for kv_data in self.kv_data):
-            raise ValueError(f'Duplicated key name {key!r}')
+            logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
 
         self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
 

gguf-py/gguf/tensor_mapping.py

@@ -31,6 +31,7 @@ class TensorNameMap:
             "model.embeddings",                          # rwkv7
             "model.word_embeddings",                     # bailingmoe
             "language_model.model.embed_tokens",         # llama4
+            "encoder",                                   # neobert
         ),
 
         # Token type embeddings
@@ -134,6 +135,7 @@ class TensorNameMap:
             "rwkv.blocks.{bid}.ln1",                                # rwkv6
             "model.layers.{bid}.ln1",                               # rwkv7
             "model.layers.{bid}.input_layernorm",                   # llama4
+            "transformer_encoder.{bid}.attention_norm",             # neobert
         ),
 
         # Attention norm 2
@@ -161,6 +163,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
+            "transformer_encoder.{bid}.qkv",                                       # neobert
         ),
 
         # Attention query
@@ -236,6 +239,7 @@ class TensorNameMap:
             "transformer.layers.{bid}.attn.out_proj",                       # openelm
             "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
             "model.layers.{bid}.self_attn.o_proj",                          # llama4
+            "transformer_encoder.{bid}.wo",                                 # neobert
         ),
 
         # Attention output norm
@@ -276,6 +280,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
+            "transformer_encoder.{bid}.ffn_norm",                            # neobert
         ),
 
         # Post feed-forward norm
@@ -340,6 +345,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
             "model.layers.{bid}.feed_forward.up_proj",                # llama4
+            "transformer_encoder.{bid}.ffn.w12",                      # neobert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -422,6 +428,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
             "model.layers.{bid}.feed_forward.down_proj",              # llama4
+            "transformer_encoder.{bid}.ffn.w3",                       # neobert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -832,12 +839,14 @@ class TensorNameMap:
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
+            "layer_norm",               # neobert
         ),
 
         MODEL_TENSOR.CLS: (
             "classifier",       # jina
             "classifier.dense", # roberta
             "pre_classifier",   # distillbert
+            "dense",            # neobert
         ),
 
         MODEL_TENSOR.CLS_OUT: (

include/llama.h

@@ -254,7 +254,10 @@ extern "C" {
     // - seq_id : the sequence to which the respective token belongs
     //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
-    //            (if set to NULL, only the logits for last token will be returned)
+    //            (if set to NULL:
+    //               - if embeddings: all tokens are output
+    //               - if not:        only the last token is output
+    //            )
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -262,8 +265,8 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
-        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
         int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
@@ -961,8 +964,8 @@ extern "C" {
     // Get the number of threads used for prompt and batch processing (multiple token).
     LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
-    // Set whether the model is in embeddings mode or not
-    // If true, embeddings will be returned but logits will not
+    // Set whether the context outputs embeddings or not
+    // TODO: rename to avoid confusion with llama_get_embeddings()
     LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
 
     // Set whether to use causal attention or not

src/llama-arch.cpp

@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
+    { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -73,6 +74,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
     { LLM_ARCH_DOTS1,            "dots1"            },
+    { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -244,6 +246,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_ARCEE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -495,6 +515,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_NEO_BERT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
+            { LLM_TENSOR_CLS,             "cls" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {

src/llama-arch.h

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
+    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -77,6 +78,7 @@ enum llm_arch {
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
     LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-batch.cpp

@@ -299,7 +299,8 @@ llama_batch_allocr::llama_batch_allocr() {
 bool llama_batch_allocr::init(
         const llama_batch & batch_inp,
         const llama_vocab & vocab,
-        const llama_memory_i * memory) {
+        const llama_memory_i * memory,
+        bool embd_all) {
     clear();
 
     batch = batch_inp;
@@ -378,10 +379,31 @@ bool llama_batch_allocr::init(
     }
 
     if (!batch.logits) {
-        // by default return the output only for the last token
-        output.resize(batch.n_tokens);
-        output[output.size() - 1] = true;
+        if (embd_all) {
+            // return the output for all tokens
+            output.resize(batch.n_tokens, true);
+        } else {
+            // return the output only for the last token
+            output.resize(batch.n_tokens, false);
+            output[output.size() - 1] = true;
+        }
+
         batch.logits = output.data();
+    } else if (embd_all) {
+        bool warn = false;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.logits[i] == 0) {
+                warn = true;
+            }
+        }
+
+        if (warn) {
+            LLAMA_LOG_WARN("%s: embeddings required but some input tokens were not marked as outputs -> overriding\n", __func__);
+
+            output.resize(batch.n_tokens, true);
+            batch.logits = output.data();
+        }
     }
 
     //

src/llama-batch.h

@@ -88,7 +88,8 @@ public:
     bool init(
             const llama_batch & batch_inp,
             const llama_vocab & vocab,
-            const llama_memory_i * memory);
+            const llama_memory_i * memory,
+            bool embd_all);
 
     const llama_batch & get_batch() const;
 

src/llama-context.cpp

@@ -728,7 +728,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     }
 
     // note: during encode, we always pass the full sequence starting from pos = 0
-    if (!batch_allocr->init(batch_inp, model.vocab, nullptr)) {
+    if (!batch_allocr->init(batch_inp, model.vocab, nullptr, true)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -894,7 +894,10 @@ int llama_context::decode(const llama_batch & batch_inp) {
         return -1;
     }
 
-    if (!batch_allocr->init(batch_inp, model.vocab, memory.get())) {
+    // when computing embeddings, all tokens are output
+    const bool embd_all = cparams.embeddings;
+
+    if (!batch_allocr->init(batch_inp, model.vocab, memory.get(), embd_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -911,12 +914,9 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // this indicates we are doing pooled embedding
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
     const uint32_t n_outputs_all = batch_allocr->get_n_outputs();
 
-    if (embd_pooled) {
+    if (embd_all) {
         // require that all tokens are output
         if (n_outputs_all != n_tokens_all) {
             LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n",
@@ -945,7 +945,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     llama_memory_state_ptr mstate;
 
     while (true) {
-        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled);
+        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_all);
         if (!mstate) {
             return -2;
         }
@@ -1058,7 +1058,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
-        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
+        auto * t_logits = res->get_logits();
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
@@ -1222,9 +1222,8 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_vocab = vocab.n_tokens();
     const auto n_embd  = hparams.n_embd;
 
-    // TODO: use a per-batch flag for logits presence instead
-    bool has_logits = !cparams.embeddings;
-    bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
+    bool has_logits = true;
+    bool has_embd   = cparams.embeddings;
 
     // TODO: hacky enc-dec support
     if (model.arch == LLM_ARCH_T5) {
@@ -2044,14 +2043,11 @@ void llama_context::opt_epoch_iter(
 
         n_queued_tokens += n_tokens_all;
 
-        // this indicates we are doing pooled embedding
-        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
         embd_seq.clear();
 
         uint32_t n_outputs_all = n_tokens_all;
 
-        auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled);
+        auto mstate = memory->init_batch(batch, cparams.n_ubatch, true);
         if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;

src/llama-kv-cache-recurrent.cpp

@@ -359,9 +359,7 @@ llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled) {
-    GGML_UNUSED(embd_pooled);
-
+llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
     auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
 
     std::vector<llama_ubatch> ubatches;
@@ -369,8 +367,8 @@ llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch &
     while (sbatch.n_tokens > 0) {
         llama_ubatch ubatch;
 
-        if (embd_pooled) {
-            // Pooled embeddings cannot be split across ubatches (yet)
+        if (embd_all) {
+            // if all tokens are output, split by sequence
             ubatch = sbatch.split_seq(n_ubatch);
         } else {
             ubatch = sbatch.split_equal(n_ubatch);

src/llama-kv-cache-recurrent.h

@@ -32,7 +32,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 

src/llama-kv-cache-unified-iswa.cpp

@@ -95,8 +95,8 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled) {
-    GGML_UNUSED(embd_pooled);
+llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
 
     // first try simple split
     do {

src/llama-kv-cache-unified-iswa.h

@@ -34,7 +34,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 

src/llama-kv-cache-unified.cpp

@@ -310,8 +310,8 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
 llama_memory_state_ptr llama_kv_cache_unified::init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled) {
-    GGML_UNUSED(embd_pooled);
+            bool embd_all) {
+    GGML_UNUSED(embd_all);
 
     do {
         auto sbatch = llama_sbatch(batch, hparams.n_embd, true);

src/llama-kv-cache-unified.h

@@ -59,7 +59,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 

src/llama-memory.h

@@ -73,7 +73,7 @@ struct llama_memory_i {
     virtual llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled) = 0;
+            bool embd_all) = 0;
 
     // simulate full cache, used for allocating worst-case compute buffers
     virtual llama_memory_state_ptr init_full() = 0;

src/llama-model.cpp

@@ -599,6 +599,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.use_kq_norm = false;
                 }
             } break;
+        case LLM_ARCH_ARCEE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                // Arcee uses the same structure as Llama
+                switch (hparams.n_layer) {
+                    case 36: type = LLM_TYPE_4B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -739,6 +749,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     }
                 }
             } break;
+        case LLM_ARCH_NEO_BERT:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_CAUSAL,            hparams.causal_attn);
+                ml.get_key(LLM_KV_POOLING_TYPE,                hparams.pooling_type);
+
+                if (hparams.n_layer == 28) {
+                    type = LLM_TYPE_250M;
+                }
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2202,6 +2222,32 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_NEO_BERT:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+
+                    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+                    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+
+                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+
+                    output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff*2}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_JINA_BERT_V2:
                 {
                     tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
@@ -4190,6 +4236,37 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_ARCEE:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6141,6 +6218,117 @@ struct llm_build_bert : public llm_graph_context {
     }
 };
 
+struct llm_build_neo_bert : public llm_graph_context {
+    llm_build_neo_bert(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        // construct input embeddings (token, type, position)
+        inpL = build_inp_embd(model.tok_embd);
+        cb(inpL, "inp_embd", -1);
+
+        auto * inp_attn = build_attn_inp_no_cache();
+
+        // iterate layers
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * cur = inpL;
+
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // pre-norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+
+            // self-attention
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+            Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+            Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            // RoPE
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn, gf,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+
+            if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // re-add the layer input
+            cur = ggml_add(ctx0, cur, inpL);
+
+            ggml_tensor * ffn_inp = cur;
+            cb(ffn_inp, "ffn_inp", il);
+
+            // pre-norm
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,
+                    NULL, NULL, NULL, NULL, NULL,
+                    model.layers[il].ffn_down,
+                    NULL, NULL, NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+
+            // attentions bypass the intermediate layer
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm_enc, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_embd", -1);
+        res->t_embd = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_bloom : public llm_graph_context {
     llm_build_bloom(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -13411,6 +13599,141 @@ struct llm_build_dots1 : public llm_graph_context {
     }
 };
 
+struct llm_build_arcee : public llm_graph_context {
+    llm_build_arcee(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            // ARCEE uses relu^2 instead of silu
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
@@ -13419,6 +13742,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
+        case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 res = nullptr;
@@ -13527,6 +13851,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_bert>(*this, params, gf);
             } break;
+        case LLM_ARCH_NEO_BERT:
+            {
+                llm = std::make_unique<llm_build_neo_bert>(*this, params, gf);
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 llm = std::make_unique<llm_build_bloom>(*this, params, gf);
@@ -13753,6 +14081,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_dots1>(*this, params, gf);
             } break;
+        case LLM_ARCH_ARCEE:
+            {
+                llm = std::make_unique<llm_build_arcee>(*this, params, gf);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -13902,6 +14234,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
         case LLM_ARCH_BAILINGMOE:
+        case LLM_ARCH_NEO_BERT:
+        case LLM_ARCH_ARCEE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

src/llama-quant.cpp

@@ -585,7 +585,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
-                gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64);
+                // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {

src/llama-vocab.cpp

@@ -1987,6 +1987,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|eom_id|>"
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"
+                    || t.first == "<|end_of_text|>"
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

src/llama.cpp

@@ -198,14 +198,18 @@ static struct llama_model * llama_model_load_from_file_impl(
 
     // if using single GPU mode, remove all except the main GPU
     if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
-        if (params.main_gpu < 0 || params.main_gpu >= (int)model->devices.size()) {
-            LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %d)\n", __func__, params.main_gpu, (int)model->devices.size());
-            llama_model_free(model);
-            return nullptr;
+        if (params.main_gpu < 0) {
+            model->devices.clear();
+        } else {
+            if (params.main_gpu >= (int)model->devices.size()) {
+                LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %zu)\n", __func__, params.main_gpu, model->devices.size());
+                llama_model_free(model);
+                return nullptr;
+            }
+            ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
+            model->devices.clear();
+            model->devices.push_back(main_gpu);
         }
-        ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
-        model->devices.clear();
-        model->devices.push_back(main_gpu);
     }
 
     for (auto * dev : model->devices) {

tests/CMakeLists.txt

@@ -185,6 +185,8 @@ llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(test-regex-partial.cpp)
 
+llama_build_and_test(test-thread-safety.cpp ARGS -hf ggml-org/models -hff tinyllamas/stories15M-q4_0.gguf -ngl 99 -p "The meaning of life is" -n 128 -c 256 -ub 32 -np 4)
+
 # this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
 if (NOT WIN32)
     llama_build_and_test(test-arg-parser.cpp)

tests/test-thread-safety.cpp

@@ -0,0 +1,152 @@
+// thread safety test
+// - Loads a copy of the same model on each GPU, plus a copy on the CPU
+// - Creates n_parallel (--parallel) contexts per model
+// - Runs inference in parallel on each context
+
+#include <thread>
+#include <vector>
+#include <atomic>
+#include "llama.h"
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "sampling.h"
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+        return 1;
+    }
+
+    common_init();
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+
+    //llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
+    //    if (level == GGML_LOG_LEVEL_ERROR) {
+    //        common_log_add(common_log_main(), level, "%s", text);
+    //    }
+    //}, NULL);
+
+    auto cparams = common_context_params_to_llama(params);
+
+    int dev_count = ggml_backend_dev_count();
+    int gpu_dev_count = 0;
+    for (int i = 0; i < dev_count; ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        if (dev && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            gpu_dev_count++;
+        }
+    }
+    const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split
+    //const int num_models = std::max(1, gpu_dev_count);
+    const int num_contexts = std::max(1, params.n_parallel);
+
+    std::vector<llama_model_ptr> models;
+    std::vector<std::thread> threads;
+    std::atomic<bool> failed = false;
+
+    for (int m = 0; m < num_models; ++m) {
+        auto mparams = common_model_params_to_llama(params);
+
+        if (m < gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = m;
+        } else if (m == gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = -1; // CPU model
+        } else {
+            mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;;
+        }
+
+        llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
+        if (model == NULL) {
+            LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+            return 1;
+        }
+
+        models.emplace_back(model);
+    }
+
+    for  (int m = 0; m < num_models; ++m) {
+        auto * model = models[m].get();
+        for (int c = 0; c < num_contexts; ++c) {
+            threads.emplace_back([&, m, c, model]() {
+                LOG_INF("Creating context %d/%d for model %d/%d\n", c + 1, num_contexts, m + 1, num_models);
+
+                llama_context_ptr ctx { llama_init_from_model(model, cparams) };
+                if (ctx == NULL) {
+                    LOG_ERR("failed to create context\n");
+                    failed.store(true);
+                    return;
+                }
+
+                std::unique_ptr<common_sampler, decltype(&common_sampler_free)> sampler { common_sampler_init(model, params.sampling), common_sampler_free };
+                if (sampler == NULL) {
+                    LOG_ERR("failed to create sampler\n");
+                    failed.store(true);
+                    return;
+                }
+
+                llama_batch batch = {};
+                {
+                    auto prompt = common_tokenize(ctx.get(), params.prompt, true);
+                    if (prompt.empty()) {
+                        LOG_ERR("failed to tokenize prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                    batch = llama_batch_get_one(prompt.data(), prompt.size());
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("failed to decode prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                const auto * vocab = llama_model_get_vocab(model);
+                std::string result = params.prompt;
+
+                for (int i = 0; i < params.n_predict; i++) {
+                    llama_token token;
+                    if (batch.n_tokens > 0) {
+                        token = common_sampler_sample(sampler.get(), ctx.get(), batch.n_tokens - 1);
+                    } else {
+                        token = llama_vocab_bos(vocab);
+                    }
+
+                    result += common_token_to_piece(ctx.get(), token);
+
+                    if (llama_vocab_is_eog(vocab, token)) {
+                        break;
+                    }
+
+                    batch = llama_batch_get_one(&token, 1);
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("Model %d/%d, Context %d/%d: failed to decode\n", m + 1, num_models, c + 1, num_contexts);
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                LOG_INF("Model %d/%d, Context %d/%d: %s\n\n", m + 1, num_models, c + 1, num_contexts, result.c_str());
+            });
+        }
+    }
+
+    for (auto & thread : threads) {
+        thread.join();
+    }
+
+    if (failed) {
+        LOG_ERR("One or more threads failed.\n");
+        return 1;
+    }
+
+    LOG_INF("All threads finished without errors.\n");
+    return 0;
+}

tools/server/server.cpp

@@ -88,6 +88,26 @@ enum error_type {
     ERROR_TYPE_NOT_SUPPORTED, // custom error
 };
 
+static bool server_task_type_need_embd(server_task_type task_type) {
+    switch (task_type) {
+        case SERVER_TASK_TYPE_EMBEDDING:
+        case SERVER_TASK_TYPE_RERANK:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static bool server_task_type_need_logits(server_task_type task_type) {
+    switch (task_type) {
+        case SERVER_TASK_TYPE_COMPLETION:
+        case SERVER_TASK_TYPE_INFILL:
+            return true;
+        default:
+            return false;
+    }
+}
+
 struct slot_params {
     bool stream        = true;
     bool cache_prompt  = true; // remember the prompt to avoid reprocessing all prompt
@@ -1330,13 +1350,24 @@ struct server_slot {
         n_draft_accepted = 0;
     }
 
-    bool is_non_causal() const {
-        return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK;
+    bool need_embd() const {
+        return server_task_type_need_embd(task_type);
+    }
+
+    bool need_logits() const {
+        return server_task_type_need_logits(task_type);
+    }
+
+    // if the context does not have a memory module then all embeddings have to be computed within a single ubatch
+    // also we cannot split if the pooling would require any past tokens
+    bool can_split() const {
+        return
+            !need_embd() ||
+            (llama_get_memory(ctx) && llama_pooling_type(ctx) == LLAMA_POOLING_TYPE_LAST);
     }
 
     bool can_batch_with(server_slot & other_slot) const {
-        return is_non_causal() == other_slot.is_non_causal()
-            && are_lora_equal(lora, other_slot.lora);
+        return task_type == other_slot.task_type && are_lora_equal(lora, other_slot.lora);
     }
 
     bool has_budget(const common_params & global_params) {
@@ -1480,7 +1511,6 @@ struct server_slot {
             {"n_ctx",         n_ctx},
             {"speculative",   can_speculate()},
             {"is_processing", is_processing()},
-            {"non_causal",    is_non_causal()},
             {"params",        params.to_json()},
             {"prompt",        prompt_tokens.detokenize(ctx, true)},
             {"next_token",
@@ -2730,6 +2760,7 @@ struct server_context {
                         queue_tasks.defer(std::move(task));
                         break;
                     }
+
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
@@ -3092,7 +3123,14 @@ struct server_context {
                             continue;
                         }
 
-                        if (slot.is_non_causal()) {
+                        // TODO: support memory-less logits computation
+                        if (slot.need_logits() && !llama_get_memory(ctx)) {
+                            slot.release();
+                            send_error(slot, "the current context does not logits computation. skipping", ERROR_TYPE_SERVER);
+                            continue;
+                        }
+
+                        if (!slot.can_split()) {
                             if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.release();
                                 send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
@@ -3227,8 +3265,7 @@ struct server_context {
                         }
 
                         if (slot.n_past == slot.n_prompt_tokens && slot.n_past > 0) {
-                            // we have to evaluate at least 1 token to generate logits.
-                            SLT_WRN(slot, "need to evaluate at least 1 token to generate logits, n_past = %d, n_prompt_tokens = %d\n", slot.n_past, slot.n_prompt_tokens);
+                            SLT_WRN(slot, "need to evaluate at least 1 token for each active slot, n_past = %d, n_prompt_tokens = %d\n", slot.n_past, slot.n_prompt_tokens);
 
                             slot.n_past--;
                         }
@@ -3236,8 +3273,7 @@ struct server_context {
                         slot.n_prompt_tokens_processed = 0;
                     }
 
-                    // non-causal tasks require to fit the entire prompt in the physical batch
-                    if (slot.is_non_causal()) {
+                    if (!slot.can_split()) {
                         // cannot fit the prompt in the current batch - will try next iter
                         if (batch.n_tokens + slot.n_prompt_tokens > n_batch) {
                             continue;
@@ -3259,8 +3295,7 @@ struct server_context {
                     slot.cache_tokens.keep_first(slot.n_past);
 
                     // check if we should process the image
-                    if (slot.n_past < slot.n_prompt_tokens
-                            && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
+                    if (slot.n_past < slot.n_prompt_tokens && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
                         // process the image
                         int32_t new_n_past;
                         int32_t res = slot.prompt_tokens.process_chunk(ctx, mctx, slot.n_past, slot.id, new_n_past);
@@ -3291,8 +3326,8 @@ struct server_context {
                             break; // end of text chunk
                         }
 
-                        // without pooling, we want to output the embeddings for all the tokens in the batch
-                        const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE;
+                        // embedding requires all tokens in the batch to be output
+                        const bool need_embd = server_task_type_need_embd(slot.task_type);
 
                         common_batch_add(batch, cur_tok, slot.n_past, { slot.id }, need_embd);
                         slot.cache_tokens.push_back(cur_tok);
@@ -3346,17 +3381,15 @@ struct server_context {
         SRV_DBG("decoding batch, n_tokens = %d\n", batch.n_tokens);
 
         if (slot_batched) {
-            // make sure we're in the right embedding mode
-            llama_set_embeddings(ctx, slot_batched->is_non_causal());
             // apply lora, only need to do it once per batch
             common_set_adapter_lora(ctx, slot_batched->lora);
-        }
 
-        const bool do_encode = (params_base.embedding || params_base.reranking);
+            llama_set_embeddings(ctx, slot_batched->need_embd());
+        }
 
         // pad the batch so that batch.n_tokens >= n_slots
         // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689
-        if (do_encode) {
+        if (slot_batched->need_embd()) {
             const int n_slots = slots.size();
 
             if (batch.n_tokens < n_slots) {
@@ -3378,8 +3411,11 @@ struct server_context {
                 SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id);
 
                 for (int j = 0; j < n_add; ++j) {
-                    common_batch_add(batch, 0, j, { seq_id }, false);
+                    common_batch_add(batch, 0, j, { seq_id }, true);
                 }
+
+                slots[seq_id].cache_tokens.clear();
+                llama_memory_seq_rm(llama_get_memory(ctx), seq_id, -1, -1);
             }
         }
 
@@ -4174,11 +4210,6 @@ int main(int argc, char ** argv) {
             oaicompat_type oaicompat) -> void {
         GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
 
-        if (ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
-            return;
-        }
-
         auto completion_id = gen_chatcmplid();
         std::unordered_set<int> task_ids;
         try {
@@ -4433,12 +4464,8 @@ int main(int argc, char ** argv) {
             OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
     };
 
-    const auto handle_chat_completions = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_chat_completions = [&ctx_server, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         LOG_DBG("request: %s\n", req.body.c_str());
-        if (ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
-            return;
-        }
 
         auto body = json::parse(req.body);
         std::vector<raw_buffer> files;
@@ -4566,13 +4593,18 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
-        const json body = json::parse(req.body);
+        if (!ctx_server.params_base.embedding) {
+            res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
 
         if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
             res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
+        const json body = json::parse(req.body);
+
         // for the shape of input/content, see tokenize_input_prompts()
         json prompt;
         if (body.count("input") != 0) {
@@ -4662,8 +4694,8 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
+        if (!ctx_server.params_base.embedding || ctx_server.params_base.pooling_type != LLAMA_POOLING_TYPE_RANK) {
+            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -4878,7 +4910,9 @@ int main(int argc, char ** argv) {
     };
 
     bool was_bound = false;
+    bool is_sock = false;
     if (string_ends_with(std::string(params.hostname), ".sock")) {
+        is_sock = true;
         LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
         svr->set_address_family(AF_UNIX);
         // bind_to_port requires a second arg, any value other than 0 should
@@ -4956,7 +4990,9 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
+    LOG_INF("%s: server is listening on %s - starting the main loop\n", __func__,
+            is_sock ? string_format("unix://%s", params.hostname.c_str()).c_str() :
+                      string_format("http://%s:%d", params.hostname.c_str(), params.port).c_str());
 
     // this call blocks the main thread until queue_tasks.terminate() is called
     ctx_server.queue_tasks.start_loop();