ggml : dynamic ggml_sched_max_splits based on graph_size (#9047 )

* ggml : Dynamic ggml_sched_max_splits based on graph_size * Fixed and readded debug code for causes
retrieval : fix memory leak in retrieval query handling (#8955 )
2026-05-16 14:04:05 +00:00 · 2024-08-16 04:22:55 +02:00 · 2024-08-15 10:40:12 +03:00 · 2024-08-15 10:28:05 +03:00 · 2024-08-15 10:23:23 +03:00 · 2024-08-15 10:17:12 +03:00
47 changed files with 914 additions and 973 deletions
--- a/.github/workflows/bench.yml.disabled
+++ b/.github/workflows/bench.yml.disabled
@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark

@@ -129,6 +132,8 @@ jobs:

      - name: Server bench
        id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
        run: |
          set -eux

@@ -137,7 +142,7 @@ jobs:
          python bench.py \
              --runner-label ${{ env.RUNNER_LABEL }} \
              --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
              --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
              --scenario script.js \
              --duration ${{ github.event.inputs.duration || env.DURATION }} \
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -47,7 +47,7 @@ jobs:
          sysctl -a
          mkdir build
          cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)

      - name: Test
@@ -105,7 +105,7 @@ jobs:
          sysctl -a
          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
          # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

      - name: Test
@@ -222,7 +222,7 @@ jobs:
        run: |
          mkdir build
          cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
          cmake --build . --config Release -j $(nproc)

      - name: Test
@@ -696,22 +696,20 @@ jobs:
    strategy:
      matrix:
        include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
          - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
          - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
          - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'llvm-arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
          - build: 'msvc-arm64'
--- a/4
+++ b/4
@@ -763,6 +763,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif

+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2702,12 +2702,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
    return text;
 }

-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // Chat template utils
 //
--- a/common/common.h
+++ b/common/common.h
@@ -380,10 +380,6 @@ std::string llama_detokenize(
        const std::vector<llama_token> & tokens,
                                  bool   special = true);

-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -590,6 +590,12 @@ class Model:
        if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
            # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
            res = "smollm"
+        if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7":
+            # ref: https://huggingface.co/bigscience/bloom
+            res = "bloom"
+        if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
+            # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
+            res = "gpt3-finnish"

        if res is None:
            logger.warning("\n")
@@ -893,7 +899,7 @@ class GPTNeoXModel(Model):
        return tensors


-@Model.register("BloomForCausalLM")
+@Model.register("BloomForCausalLM", "BloomModel")
 class BloomModel(Model):
    model_arch = gguf.MODEL_ARCH.BLOOM

--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -94,6 +94,8 @@ models = [
    {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
    {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
 ]


--- a/examples/cvector-generator/cvector-generator.cpp
+++ b/examples/cvector-generator/cvector-generator.cpp
@@ -271,7 +271,7 @@ struct tokenized_prompt {
    size_t max_seq_len;

    tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
        max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -127,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }

 static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));

    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);

--- a/examples/export-lora/README.md
+++ b/examples/export-lora/README.md
@@ -17,9 +17,9 @@ For example:

 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2-q8_0.gguf \
-    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```

 Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:
--- a/examples/export-lora/export-lora.cpp
+++ b/examples/export-lora/export-lora.cpp
@@ -10,6 +10,12 @@

 static bool g_verbose = false;

+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
 static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
    int id = gguf_find_key(ctx_gguf, key.c_str());
    return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@@ -198,8 +204,7 @@ struct lora_merge_ctx {
        }

        // mapping base tensor to out tensor (same shape with base, but different type)
-        // if out_tensor == nullptr, we only copy it
-        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
+        std::vector<tensor_transformation> trans;
        for (auto & it : base_model.tensors) {
            bool t_a = true;
            bool t_b = true;
@@ -212,14 +217,22 @@ struct lora_merge_ctx {
                // only copy
                struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
                ggml_set_name(cpy_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
                gguf_add_tensor(ctx_out, cpy_tensor);
            } else if (t_a && t_b) {
                // need merging
                struct ggml_tensor * out_tensor = ggml_new_tensor(
                    ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
                ggml_set_name(out_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
                gguf_add_tensor(ctx_out, out_tensor);
            } else {
                throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -234,12 +247,12 @@ struct lora_merge_ctx {

        // process base model tensors
        size_t n_merged = 0;
-        for (auto & it : base_to_out_tensors) {
-            if (it.second != nullptr) {
-                merge_tensor(it.first, it.second);
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
                n_merged++;
            } else {
-                copy_tensor(it.first);
+                copy_tensor(it.in);
            }
        }

@@ -252,7 +265,7 @@ struct lora_merge_ctx {
        }

        printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
    }

    void copy_tensor(struct ggml_tensor * base) {
@@ -285,6 +298,10 @@ struct lora_merge_ctx {
        for (size_t i = 0; i < adapters.size(); ++i) {
            auto t_a = adapters[i]->get_tensor(name_lora_a);
            auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
            inp_a[i] = ggml_dup_tensor(ctx, t_a);
            inp_b[i] = ggml_dup_tensor(ctx, t_b);
        }
--- a/examples/imatrix/imatrix.cpp
+++ b/examples/imatrix/imatrix.cpp
@@ -433,8 +433,8 @@ static void process_logits(
 }

 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
    const int n_ctx = llama_n_ctx(ctx);

    auto tim1 = std::chrono::high_resolution_clock::now();
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -203,8 +203,8 @@ int main(int argc, char ** argv) {
        LOG_TEE("\n");
        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
    }
-    const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
    LOG("add_bos: %d\n", add_bos);

    std::vector<llama_token> embd_inp;
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -267,9 +267,9 @@ int main(int argc, char ** argv) {
        }
    }

-    const bool add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = llama_add_bos_token(model);
    if (!llama_model_has_encoder(model)) {
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
+        GGML_ASSERT(!llama_add_eos_token(model));
    }
    LOG("add_bos: %d\n", add_bos);

--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -340,8 +340,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
    // Output: `perplexity: 13.5106 [114/114]`
    // BOS tokens will be added for each chunk before eval

-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));

    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);

@@ -480,8 +480,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
    // Output: `perplexity: 13.5106 [114/114]`
    // BOS tokens will be added for each chunk before eval

-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));

    std::ofstream logits_stream;
    if (!params.logits_file.empty()) {
@@ -1733,8 +1733,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
    const int n_batch = params.n_batch;
    const int num_batches = (n_ctx + n_batch - 1)/n_batch;
    const int nv = 2*((n_vocab + 1)/2) + 4;
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));

    std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
    std::vector<float>    kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
--- a/examples/retrieval/retrieval.cpp
+++ b/examples/retrieval/retrieval.cpp
@@ -253,6 +253,8 @@ int main(int argc, char ** argv) {
        chunks[i].tokens.clear();
    }

+    struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
+
    // start loop, receive query and return top k similar chunks based on cosine similarity
    std::string query;
    while (true) {
@@ -260,7 +262,6 @@ int main(int argc, char ** argv) {
        std::getline(std::cin, query);
        std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);

-        struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
        batch_add_seq(query_batch, query_tokens, 0);

        std::vector<float> query_emb(n_embd, 0);
@@ -293,6 +294,7 @@ int main(int argc, char ** argv) {
    }

    // clean up
+    llama_batch_free(query_batch);
    llama_print_timings(ctx);
    llama_free(ctx);
    llama_free_model(model);
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -693,9 +693,8 @@ struct server_context {

        n_ctx = llama_n_ctx(ctx);

-        add_bos_token = llama_should_add_bos_token(model);
-        has_eos_token = llama_add_eos_token(model) != 1;
-
+        add_bos_token = llama_add_bos_token(model);
+        has_eos_token = !llama_add_eos_token(model);
        return true;
    }

@@ -754,13 +753,13 @@ struct server_context {
        default_generation_settings_for_props = get_formated_generation(slots.front());
        default_generation_settings_for_props["seed"] = -1;

-        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
        // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
        {
            const int32_t n_batch = llama_n_batch(ctx);

            // only a single seq_id per token is needed
-            batch = llama_batch_init(n_batch, 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
        }

        metrics.init();
@@ -1137,28 +1136,19 @@ struct server_context {
        if (!system_prompt.empty()) {
            system_tokens = ::llama_tokenize(ctx, system_prompt, true);

-            llama_batch_clear(batch);
-
-            for (int i = 0; i < (int)system_tokens.size(); ++i) {
-                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-            }
-
            const int32_t n_batch = llama_n_batch(ctx);
+            const int32_t n_tokens_prompt = system_tokens.size();

-            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                    0, 0, 0, // unused
-                };
+            for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
+                const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);

-                if (llama_decode(ctx, batch_view) != 0) {
+                llama_batch_clear(batch);
+
+                for (int32_t j = 0; j < n_tokens; ++j) {
+                    llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
+                }
+
+                if (llama_decode(ctx, batch) != 0) {
                    LOG_ERROR("llama_decode() failed", {});
                    return;
                }
@@ -1331,7 +1321,7 @@ struct server_context {

        return json {
            {"n_ctx",                     slot.n_ctx},
-            {"n_predict",                 slot.n_predict},
+            {"n_predict",                 slot.n_predict},     // Server configured n_predict
            {"model",                     params.model_alias},
            {"seed",                      slot.sparams.seed},
            {"temperature",               slot.sparams.temp},
@@ -1353,7 +1343,7 @@ struct server_context {
            {"mirostat_eta",              slot.sparams.mirostat_eta},
            {"penalize_nl",               slot.sparams.penalize_nl},
            {"stop",                      slot.params.antiprompt},
-            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
+            {"max_tokens",                slot.params.n_predict}, // User configured n_predict
            {"n_keep",                    slot.params.n_keep},
            {"n_discard",                 slot.params.n_discard},
            {"ignore_eos",                ignore_eos},
@@ -1861,6 +1851,8 @@ struct server_context {
                    llama_lora_adapters_apply(ctx, lora_adapters);
                    server_task_result result;
                    result.id = task.id;
+                    result.stop = true;
+                    result.error = false;
                    result.data = json{{ "success", true }};
                    queue_results.send(result);
                } break;
@@ -2045,7 +2037,7 @@ struct server_context {
                        slot.t_start_generation = 0;

                        if (slot.infill) {
-                            const bool add_bos = llama_should_add_bos_token(model);
+                            const bool add_bos = llama_add_bos_token(model);
                            bool suff_rm_leading_spc = true;
                            if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                                params.input_suffix.erase(0, 1);
--- a/examples/tokenize/tokenize.cpp
+++ b/examples/tokenize/tokenize.cpp
@@ -362,7 +362,7 @@ int main(int raw_argc, char ** raw_argv) {
        prompt = stdin_buffer.str();
    }

-    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool model_wants_add_bos = llama_add_bos_token(model);
    const bool add_bos = model_wants_add_bos && !no_bos;
    const bool parse_special = !no_parse_special;

--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -129,13 +129,13 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)

-option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -244,6 +244,8 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1

+#define GGML_ROPE_TYPE_NEOX 2
+
 #define GGUF_MAGIC "GGUF"

 #define GGUF_VERSION 3
@@ -1453,8 +1455,8 @@ extern "C" {
            struct ggml_tensor  * b);

    // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
-    // if mode & 2 == 1, GPT-NeoX style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
    //
    // b is an int32 vector with size a->ne[2], it contains the positions
    GGML_API struct ggml_tensor * ggml_rope(
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -602,6 +602,10 @@ if (GGML_VULKAN)
            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
        endif()

+        if (GGML_VULKAN_PERF)
+            add_compile_definitions(GGML_VULKAN_PERF)
+        endif()
+
        if (GGML_VULKAN_VALIDATE)
            add_compile_definitions(GGML_VULKAN_VALIDATE)
        endif()
--- a/ggml/src/ggml-backend.c
+++ b/ggml/src/ggml-backend.c
@@ -1018,10 +1018,6 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #define GGML_SCHED_MAX_BACKENDS 16
 #endif

-#ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 2048
-#endif
-
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
 #define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
 #endif
@@ -1125,7 +1121,8 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
 }

 #if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1549,7 +1546,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                    sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                    GGML_ASSERT(sched->splits != NULL);
                }
-                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
                split = &sched->splits[i_split];
                split->backend_id = node_backend_id;
                split->i_start = i;
@@ -1865,13 +1861,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
    sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
    sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));

-    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
    sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
    sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
    sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
    sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));

-    sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
    sched->context_buffer = malloc(sched->context_buffer_size);

    const int initial_splits_capacity = 16;
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -2881,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
                             beta_slow, corr_dims);

-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

    // init cos/sin cache
    ggml_cann_pool_alloc sin_allocator(
--- a/ggml/src/ggml-cuda/rope.cu
+++ b/ggml/src/ggml-cuda/rope.cu
@@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));

-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

    const int32_t * pos = (const int32_t *) src1_d;

--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
@@ -2313,7 +2313,7 @@ static enum ggml_status ggml_metal_graph_compute(
                        memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                        memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));

-                        const bool is_neox = mode & 2;
+                        const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

                        id<MTLComputePipelineState> pipeline = nil;

--- a/ggml/src/ggml-sycl/rope.cpp
+++ b/ggml/src/ggml-sycl/rope.cpp
@@ -226,7 +226,7 @@ void ggml_sycl_op_rope(
    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));

-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

    const int32_t * pos = (const int32_t *) src1_dd;

--- a/ggml/src/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan.cpp
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -14094,7 +14094,7 @@ static void ggml_compute_forward_rope_f32(
    float corr_dims[2];
    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);

-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

    const float * freq_factors = NULL;
    if (src2 != NULL) {
@@ -14219,7 +14219,7 @@ static void ggml_compute_forward_rope_f16(
    float corr_dims[2];
    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);

-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;

    const float * freq_factors = NULL;
    if (src2 != NULL) {
--- a/ggml/src/kompute-shaders/op_rope_f16.comp
+++ b/ggml/src/kompute-shaders/op_rope_f16.comp
@@ -11,7 +11,7 @@ void main() {
    const uint i2 = gl_WorkGroupID.y;
    const uint i1 = gl_WorkGroupID.x;

-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;

    float corr_dims[2];
    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
--- a/ggml/src/kompute-shaders/op_rope_f32.comp
+++ b/ggml/src/kompute-shaders/op_rope_f32.comp
@@ -11,7 +11,7 @@ void main() {
    const uint i2 = gl_WorkGroupID.y;
    const uint i1 = gl_WorkGroupID.x;

-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;

    float corr_dims[2];
    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
--- a/ggml/src/kompute-shaders/rope_common.comp
+++ b/ggml/src/kompute-shaders/rope_common.comp
@@ -1,5 +1,7 @@
 #include "common.comp"

+#define GGML_ROPE_TYPE_NEOX 2
+
 // TODO: use a local size of 32 or more (Metal uses 1024)
 layout(local_size_x = 1) in;

--- a/ggml/src/vulkan-shaders/concat.comp
+++ b/ggml/src/vulkan-shaders/concat.comp
@@ -30,6 +30,10 @@ void main() {
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
 #else
-    data_d[p.d_offset + dst_idx] = is_src0 ? data_a[src0_idx] : data_b[src1_idx];
+    if (is_src0) {
+        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+    } else {
+        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+    }
 #endif
 }
--- a/ggml/src/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec.comp
@@ -39,8 +39,7 @@ void main() {
        vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);

        // matrix multiplication
-        tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) +
-                    FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
    }

    // sum up partial sums and write back result
--- a/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp
@@ -53,7 +53,7 @@ void main() {

        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);

-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
    }

    // sum up partial sums and write back result
--- a/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp
@@ -52,7 +52,7 @@ void main() {
        // y is not transposed but permuted
        const uint iy = channel*nrows_y + row_y;

-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
    }

    // dst is not transposed and not permuted
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -39,24 +39,25 @@ void main() {
        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3);
-            sum2 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF);
+            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
        }
-        tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
    }

    // sum up partial sums and write back result
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -40,16 +40,17 @@ void main() {

        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4));
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
        }
-        tmp[16 * ix + tid] += d * sum;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
    }

    // sum up partial sums and write back result
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
@@ -67,17 +67,17 @@ void main() {
        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66]  >> 4);
        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);

-        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3);
-        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7);
-        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11);
-        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15);
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx    ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx    ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7
-        );
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]),      q4_0,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]),  q4_1,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]),  q4_2,  FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) *  q4_3)));
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6,  FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]),      q4_8,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]),  q4_9,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]),  q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) *  q4_11)));
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6,     FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7)))))))))))))));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
 #else
        const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
        const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
@@ -88,16 +88,19 @@ void main() {
        const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
        const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);

-        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * q4_0  + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
-        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2  + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
-        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * q4_4  + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
-        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
-        );
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+          + fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));

-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
+                        sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
+                       fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
 #endif
    }

--- a/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
@@ -66,35 +66,33 @@ void main() {
        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);

-        const FLOAT_TYPE sx = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sy = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sz = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sw = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            (FLOAT_TYPE(data_b[b_offset + y1_idx]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17])) * sc2 + (FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49])) * sc3
-          + (FLOAT_TYPE(data_b[b_offset + y2_idx]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17])) * sc6 + (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7
-        );
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+        const FLOAT_TYPE sx =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sy =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sz =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sw =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE smin =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
+              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
    }

    // sum up partial sums and write back result
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
@@ -44,22 +44,22 @@ void main() {
        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);

 #if K_QUANTS_PER_ITERATION == 1
-        FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32);
-        tmp[16 * ix + tid] += sum;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
 #else
        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
        [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32);
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
        }
        tmp[16 * ix + tid] += sum;
 #endif
--- a/ggml/src/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/vulkan-shaders/mul_mm.comp
@@ -326,10 +326,10 @@ void main() {
                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
            }
            const float d = loadd.x * sc;
-            const float m = loadd.y * mbyte;
+            const float m = -loadd.y * mbyte;

-            buf_a[buf_idx    ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) - m);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
 #elif defined(DATA_A_Q5_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
@@ -357,10 +357,10 @@ void main() {
                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
            }
            const float d = loadd.x * sc;
-            const float m = loadd.y * mbyte;
+            const float m = -loadd.y * mbyte;

-            buf_a[buf_idx    ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0)) - m);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
 #elif defined(DATA_A_Q6_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
@@ -463,7 +463,8 @@ void main() {
                [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-                            sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]);
+                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
+                            sums[sums_idx] = fma(float(cache_a[wsir * TM + cr]), float(cache_b[wsic * TN + cc]), sums[sums_idx]);
                        }
                    }
                }
--- a/ggml/src/vulkan-shaders/repeat.comp
+++ b/ggml/src/vulkan-shaders/repeat.comp
@@ -0,0 +1,24 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+uint src0_idx_mod(uint idx) {
+    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return (i13 % p.ne03)*p.nb03 + (i12 % p.ne02)*p.nb02 + (i11 % p.ne01)*p.nb01 + (i10 % p.ne00)*p.nb00;
+}
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx_mod(idx)]);
+}
--- a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -380,6 +380,10 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
        string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
    }));

+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
    }));
--- a/include/llama.h
+++ b/include/llama.h
@@ -93,15 +93,14 @@ extern "C" {
        LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
        LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
        LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+        LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
+        LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
    };

-    // note: these values should be synchronized with ggml_rope
-    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
    enum llama_rope_type {
        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM =  0,
-        LLAMA_ROPE_TYPE_NEOX =  2,
-        LLAMA_ROPE_TYPE_GLM  =  4,
+        LLAMA_ROPE_TYPE_NORM = 0,
+        LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
    };

    enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -915,11 +914,8 @@ extern "C" {
    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
    LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding

-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model);
-
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_eos_token(const struct llama_model * model);

    // Codellama infill tokens
    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -410,6 +410,8 @@ struct llm_tokenizer_bpe {
                };
                break;
            case LLAMA_VOCAB_PRE_TYPE_PORO:
+            case LLAMA_VOCAB_PRE_TYPE_BLOOM:
+            case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
                regex_exprs = {
                    " ?[^(\\s|.,!?…。，、।۔،)]+",
                };
@@ -1466,11 +1468,11 @@ llama_token llama_token_pad_impl(const struct llama_vocab & vocab) {
    return vocab.special_pad_id;
 }

-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab) {
    return vocab.tokenizer_add_bos;
 }

-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
    return vocab.tokenizer_add_eos;
 }

--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -95,8 +95,8 @@ llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);

-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab);
-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);

 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3575,13 +3575,8 @@ namespace GGUFMeta {

 using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;

-// TODO: update when needed or think of some clever automatic way to do this
-static size_t llama_model_max_nodes(const llama_model & /*model*/) {
-    //if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
-    //    return 32768;
-    //}
-
-    return 8192;
+static size_t llama_model_max_nodes(const llama_model & model) {
+    return std::max<size_t>(8192, model.tensors_by_name.size()*5);
 }

 struct llama_model_loader {
@@ -5472,6 +5467,12 @@ static void llm_load_vocab(
            } else if (
                tokenizer_pre == "codeshell") {
                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
+            } else if (
+                tokenizer_pre == "bloom") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_BLOOM;
+            } else if (
+                tokenizer_pre == "gpt3-finnish") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH;
            } else {
                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
            }
@@ -18704,11 +18705,11 @@ llama_token llama_token_pad(const struct llama_model * model) {
    return llama_token_pad_impl(model->vocab);
 }

-int32_t llama_add_bos_token(const struct llama_model * model) {
+bool llama_add_bos_token(const struct llama_model * model) {
    return llama_add_bos_token_impl(model->vocab);
 }

-int32_t llama_add_eos_token(const struct llama_model * model) {
+bool llama_add_eos_token(const struct llama_model * model) {
    return llama_add_eos_token_impl(model->vocab);
 }
Author	SHA1	Message	Date
Nico Bosshard	e3f6fd56b1	ggml : dynamic ggml_sched_max_splits based on graph_size (#9047 ) * ggml : Dynamic ggml_sched_max_splits based on graph_size * Fixed and readded debug code for causes	2024-08-16 04:22:55 +02:00
gtygo	4b9afbbe90	retrieval : fix memory leak in retrieval query handling (#8955 ) * retrieval * Reuse querybatch to reduce frequent memory allocation * delete unused white space	2024-08-15 10:40:12 +03:00
Riceball LEE	37501d9c79	server : fix duplicated n_predict key in the generation_settings (#8994 )	2024-08-15 10:28:05 +03:00
Zhenwei Jin	4af8420afb	common : remove duplicate function llama_should_add_bos_token (#8778 )	2024-08-15 10:23:23 +03:00
Esko Toivonen	6bda7ce6c3	llama : add pre-tokenizer regexes for BLOOM and gpt3-finnish (#8850 )	2024-08-15 10:17:12 +03:00
Georgi Gerganov	d5492f0525	ci : disable bench workflow (#9010 )	2024-08-15 10:11:11 +03:00
Jiří Podivín	234b30676a	server : init stop and error fields of the result struct (#9026 ) Signed-off-by: Jiri Podivin <jpodivin@redhat.com>	2024-08-15 09:21:57 +03:00
0cc4m	5fd89a70ea	Vulkan Optimizations and Fixes (#8959 ) * Optimize Vulkan REPEAT performance * Use Vulkan GLSL fused multiply-add instruction where possible * Add GGML_VULKAN_PERF option to output performance data per operator * Rework and fix Vulkan descriptor set and descriptor pool handling * Fix float32 concat f16 shader validation error * Add Vulkan GROUP_NORM eps parameter * Fix validation error with transfer queue memory barrier flags * Remove trailing whitespaces	2024-08-14 18:32:53 +02:00
compilade	98a532d474	server : fix segfault on long system prompt (#8987 ) * server : fix segfault on long system prompt * server : fix parallel generation with very small batch sizes * server : fix typo in comment	2024-08-14 09:51:02 +03:00
Georgi Gerganov	43bdd3ce18	cmake : remove unused option GGML_CURL (#9011 )	2024-08-14 09:14:49 +03:00
Daniel Bevenius	06943a69f6	ggml : move rope type enum to ggml.h (#8949 ) * ggml : move rope type enum to ggml.h This commit moves the `llama_rope_type` enum from `llama.h` to `ggml.h` and changes its name to `ggml_rope_type`. The motivation for this change is to address the TODO in `llama.h` and use the enum in ggml. Note: This commit does not change the `mode` parameter to be of type `enum ggml_rope_type`. The name `mode` and its usage suggest that it might be more generic and possibly used as a bit field for multiple flags. Further investigation/discussion may be needed to determine if `mode` should be restricted to RoPE types. * squash! ggml : move rope type enum to ggml.h This commit removes GGML_ROPE_TYPE_NONE and GGML_ROPE_TYPE_GLM from ggml.h, and back the llama_rope_type enum. I've kept the assert for GGML_ROPE_TYPE_GLM as I'm not sure if it is safe to remove it yet. * squash! ggml : move rope type enum to ggml.h This commit removes the enum ggml_rope_type from ggml.h and replaces it with a define (GGML_ROPE_TYPE_NEOX). This define is used in the code to check if the mode is set to GPT-NeoX. Also the enum llama_rope_type has been updated to reflect this change. * squash! ggml : move rope type enum to ggml.h This commit contains a suggestion enable the GGML_ROPE_TYPE_NEOX macro/define to be passed to the shader compiler. * squash! ggml : move rope type enum to ggml.h This commit fixes the editorconfig-checker warnings. * squash! ggml : move rope type enum to ggml.h Update comment for ggml_rope function. * Revert "squash! ggml : move rope type enum to ggml.h" This reverts commit `6261222bd0`. * squash! ggml : move rope type enum to ggml.h Add GGML_ROPE_TYPE_NEOX to rope_common.comp. * remove extra line --------- Co-authored-by: slaren <slarengh@gmail.com>	2024-08-13 21:13:15 +02:00
Xuan Son Nguyen	828d6ff7d7	export-lora : throw error if lora is quantized (#9002 )	2024-08-13 11:41:14 +02:00
Diogo Teles Sant'Anna	fc4ca27b25	ci : fix github workflow vulnerable to script injection (#9008 ) Signed-off-by: Diogo Teles Sant'Anna <diogoteles@google.com>	2024-08-12 19:28:23 +03:00
Radoslav Gerganov	1f67436c5e	ci : enable RPC in all of the released builds (#9006 ) ref: #8912	2024-08-12 19:17:03 +03:00
Nico Bosshard	0fd93cdef5	llama : model-based max number of graph nodes calculation (#8970 ) * llama : model-based max number of graph nodes calculation * Update src/llama.cpp --------- Co-authored-by: slaren <slarengh@gmail.com>	2024-08-12 17:13:59 +02:00