Merge branch 'master' of https://github.com/ggerganov/llama.cpp into add-paligemma-support

CONTRIBUTING.md

@@ -27,3 +27,8 @@
 
 ![matmul](media/matmul.png)
 
+# Resources
+
+The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
+
+https://github.com/ggerganov/llama.cpp/projects

README.md

@@ -174,6 +174,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 - [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
@@ -442,7 +443,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - Contributors can open PRs
 - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
-- Any help with managing issues and PRs is very appreciated!
+- Any help with managing issues, PRs and projects is very appreciated!
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
 - Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)

ci/run.sh

@@ -712,6 +712,81 @@ function gg_run_embd_bge_small {
     set +e
 }
 
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
+# rerank_tiny
+
+function gg_run_rerank_tiny {
+    cd ${SRC}
+
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+
+    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/rerank-tiny"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+
+    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s><s>hi\nwhat is panda?</s><s>it's a bear\nwhat is panda?</s><s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+
+    # sample output
+    # rerank score 0:    0.029
+    # rerank score 1:    0.029
+    # rerank score 2:    0.135
+
+    # check that the score is in the range [$3, $4]
+    function check_score {
+        qnt="$1"
+        score=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$score < $3" | bc) -eq 1 ] || [ $(echo "$score > $4" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: score not in range [%s, %s])\n' "$qnt" "$score" "$3" "$4"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$score"
+        return 0
+    }
+
+    check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.15" | tee -a $OUT/${ci}-rk-f16.log
+
+    set +e
+}
+
+function gg_sum_rerank_tiny {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Rerank Tiny (Jina):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-rk-f16.log)"
+}
+
 function gg_check_build_requirements {
     if ! command -v cmake &> /dev/null; then
         gg_printf 'cmake not found, please install'
@@ -726,15 +801,6 @@ function gg_check_build_requirements {
     fi
 }
 
-function gg_sum_embd_bge_small {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'BGE Small (BERT):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-}
-
 ## main
 
 export LLAMA_LOG_PREFIX=1
@@ -762,6 +828,7 @@ test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
+    test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
         test $ret -eq 0 && gg_run test_scripts_debug

common/arg.cpp

@@ -284,6 +284,10 @@ static bool gpt_params_parse_ex(int argc, char ** argv, gpt_params_context & ctx
         params.kv_overrides.back().key[0] = 0;
     }
 
+    if (params.reranking && params.embedding) {
+        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
+    }
+
     return true;
 }
 
@@ -391,7 +395,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params) {
             params.verbose_prompt = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ));
     add_opt(llama_arg(
         {"--no-display-prompt"},
         format("don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false"),
@@ -1093,13 +1097,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         }
     ).set_sparam());
     add_opt(llama_arg(
-        {"--pooling"}, "{none,mean,cls,last}",
+        {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",
         [](gpt_params & params, const std::string & value) {
             /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
             else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
-            else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
+            else if (value == "cls")  { params.pooling_type = LLAMA_POOLING_TYPE_CLS;  }
             else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; }
+            else if (value == "rank") { params.pooling_type = LLAMA_POOLING_TYPE_RANK; }
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
@@ -1749,6 +1754,13 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.embedding = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
+    add_opt(llama_arg(
+        {"--reranking", "--rerank"},
+        format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
+        [](gpt_params & params) {
+            params.reranking = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(llama_arg(
         {"--api-key"}, "KEY",
         "API key to use for authentication (default: none)",

common/common.cpp

@@ -1023,6 +1023,11 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
 
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }
+
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
@@ -1432,6 +1437,8 @@ void llama_batch_add(
                           llama_pos   pos,
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits) {
+    GGML_ASSERT(batch.seq_id[batch.n_tokens] && "llama_batch size exceeded");
+
     batch.token   [batch.n_tokens] = id;
     batch.pos     [batch.n_tokens] = pos;
     batch.n_seq_id[batch.n_tokens] = seq_ids.size();

common/common.h

@@ -271,6 +271,7 @@ struct gpt_params {
     int32_t embd_normalize = 2;     // normalisation for embendings (-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 embendings
+    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

@@ -291,8 +291,13 @@ class Model:
                     bid = int(part)
                     break
 
-            for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
+            for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
-                data: np.ndarray  # type hint
+                data = data_torch.squeeze().numpy()
+
+                # if data ends up empty, it means data_torch was a scalar tensor -> restore
+                if len(data.shape) == 0:
+                    data = data_torch.numpy()
+
                 n_dims = len(data.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
@@ -592,6 +597,9 @@ class Model:
         if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
             # ref: https://huggingface.co/databricks/dbrx-base
             res = "dbrx"
+        if chkhsh == "c7699093ba4255a91e702aa38a596aa81669f3525dae06c2953267dde580f448":
+            # ref: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+            res = "jina-v1-en"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
             res = "jina-v2-en"
@@ -640,6 +648,9 @@ class Model:
         if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
             # ref: https://huggingface.co/microsoft/phi-2
             res = "phi-2"
+        if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
+            # ref: https://huggingface.co/facebook/chameleon-7b
+            res = "chameleon"
 
         if res is None:
             logger.warning("\n")
@@ -2598,7 +2609,7 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
-@Model.register("XLMRobertaModel")
+@Model.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -2696,6 +2707,11 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
         # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
         if name == "embeddings.position_embeddings.weight":
             if self._position_offset is not None:
@@ -3107,6 +3123,14 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "bert.", remove the prefix
+        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+        if name.startswith("bert."):
+            name = name[5:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
 
 @Model.register("OpenELMForCausalLM")
 class OpenELMModel(Model):
@@ -4138,6 +4162,48 @@ class GraniteMoeModel(GraniteModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@Model.register("ChameleonForConditionalGeneration")
+@Model.register("ChameleonForCausalLM")  # obsolete
+class ChameleonModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHAMELEON
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # ignore image tokenizer for now
+        # TODO: remove this once image support is implemented for Chameleon
+        if name.startswith("model.vqmodel"):
+            return []
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        hidden_dim = self.hparams.get("hidden_size")
+
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        if name.endswith(("q_norm.weight", "q_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
+        if name.endswith(("k_norm.weight", "k_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
+    @staticmethod
+    def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
+        head_dim = hidden_dim // n_heads
+        data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
+        data_torch = data_torch.repeat_interleave(n_heads, 0)
+        return data_torch
+
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -81,6 +81,7 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v1-en",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
@@ -99,6 +100,7 @@ models = [
     {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
     {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
     {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
+    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
 ]
 
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -201,7 +201,7 @@ static void print_sample_weights(TransformerWeights *w){
 
 //////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
 
-struct llama_vocab {
+struct my_llama_vocab {
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -525,7 +525,7 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
+static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
         LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
@@ -583,13 +583,13 @@ static void load_vocab(const char * filename, const Config * config, struct llam
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (llama_vocab::id id=0; id<n_vocab; ++id) {
+        for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
 
             unsigned char byte_val;
-            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
             if (id == UNKNOWN_TOKEN_ID) {
                 text = "<unk>";
                 type = LLAMA_TOKEN_TYPE_UNKNOWN;
@@ -631,7 +631,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
 }
 
 static void save_as_llama_model(
-    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+    struct my_llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
 ) {
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
@@ -671,7 +671,7 @@ static void save_as_llama_model(
     std::vector<const char*> tokens;
     std::vector<float> scores;
     std::vector<llama_token_type> token_types;
-    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+    for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
         tokens.push_back(token_data.text.c_str());
         scores.push_back(token_data.score);
         token_types.push_back(token_data.type);
@@ -905,7 +905,7 @@ int main(int argc, char ** argv) {
         fclose(file);
     }
 
-    struct llama_vocab vocab;
+    struct my_llama_vocab vocab;
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;

examples/embedding/embedding.cpp

@@ -135,7 +135,7 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true, false);
+        auto inp = ::llama_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -234,6 +234,11 @@ int main(int argc, char ** argv) {
                 }
                 LOG("\n");
             }
+        } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            for (int j = 0; j < n_embd_count; j++) {
+                // NOTE: if you change this log - update the tests in ci/run.sh
+                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+            }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding
             for (int j = 0; j < n_prompts; j++) {

examples/server/README.md

@@ -7,6 +7,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
+ * Reranking endoint (WIP: https://github.com/ggerganov/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
  * Continuous batching
  * Multimodal (wip)
@@ -23,6 +24,7 @@ The project is under active development, and we are [looking for feedback and co
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
 | `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
 | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
 | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
@@ -130,7 +132,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
 | `-sp, --special` | special tokens output enabled (default: false) |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
-| `--pooling {none,mean,cls,last}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
+| `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
 | `-a, --alias STRING` | set alias for model name (to be used by REST API)<br/>(env: LLAMA_ARG_ALIAS) |
@@ -138,6 +140,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
+| `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
@@ -152,6 +155,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 
+
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
 
 Example usage of docker compose with environment variables:
@@ -478,6 +482,39 @@ The same as [the embedding example](../embedding) does.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
+### POST `/reranking`: Rerank documents according to a given query
+
+Similar to https://jina.ai/reranker/ but might change in the future.
+Requires a reranker model (such as [bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3)) and the `--embedding --pooling rank` options.
+
+    *Options:*
+
+    `query`: The query against which the documents will be ranked.
+
+    `documents`: An array strings representing the documents to be ranked.
+
+    *Aliases:*
+      - `/rerank`
+      - `/v1/rerank`
+      - `/v1/reranking`
+
+    *Examples:*
+
+    ```shell
+    curl http://127.0.0.1:8012/v1/rerank \
+        -H "Content-Type: application/json" \
+        -d '{
+            "model": "some-model",
+                "query": "What is panda?",
+                "top_n": 3,
+                "documents": [
+                    "hi",
+                "it is a bear",
+                "The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China."
+                ]
+        }' | jq
+    ```
+
 ### POST `/infill`: For code infilling.
 
 Takes a prefix and a suffix and returns the predicted completion as stream.

examples/server/server.cpp

@@ -92,6 +92,7 @@ enum server_task_type {
 enum server_task_cmpl_type {
     SERVER_TASK_CMPL_TYPE_NORMAL,
     SERVER_TASK_CMPL_TYPE_EMBEDDING,
+    SERVER_TASK_CMPL_TYPE_RERANK,
     SERVER_TASK_CMPL_TYPE_INFILL,
 };
 
@@ -172,6 +173,7 @@ struct server_slot {
     std::vector<completion_token_output> generated_token_probs;
 
     server_task_cmpl_type cmpl_type = SERVER_TASK_CMPL_TYPE_NORMAL;
+
     bool has_next_token = true;
     bool truncated      = false;
     bool stopped_eos    = false;
@@ -954,8 +956,17 @@ struct server_context {
                 slot.prompt = *prompt;
             } else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
                 slot.prompt = prompt->at(0);
+            } else if (prompt->is_array() && prompt->size() > 1) {
+                // array of strings
+                for (const auto & el : *prompt) {
+                    if (!el.is_string()) {
+                        send_error(task, "\"prompt\" must be a string, an array of strings or an array of integers", ERROR_TYPE_INVALID_REQUEST);
+                        return false;
+                    }
+                }
+                slot.prompt = *prompt;
             } else {
-                send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
+                send_error(task, "\"prompt\" must be a string, an array of strings or an array of integers", ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
         }
@@ -1389,6 +1400,7 @@ struct server_context {
 
                 res.data = json {
                     {"embedding", std::vector<float>(n_embd, 0.0f)},
+                    {"index",     slot.index},
                 };
 
                 continue;
@@ -1407,6 +1419,44 @@ struct server_context {
         queue_results.send(res);
     }
 
+    void send_rerank(const server_slot & slot, const llama_batch & batch) {
+        server_task_result res;
+        res.id       = slot.id_task;
+        res.error    = false;
+        res.stop     = true;
+
+        for (int i = 0; i < batch.n_tokens; ++i) {
+            if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) {
+                continue;
+            }
+
+            const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            if (embd == NULL) {
+                embd = llama_get_embeddings_ith(ctx, i);
+            }
+
+            if (embd == NULL) {
+                SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]);
+
+                res.data = json {
+                    {"index", slot.index},
+                    {"score", -1e6},
+                };
+
+                continue;
+            }
+
+            res.data = json {
+                {"index", slot.index},
+                {"score", embd[0]},
+            };
+        }
+
+        SLT_DBG(slot, "sending rerank result, res = '%s'\n", res.data.dump().c_str());
+
+        queue_results.send(res);
+    }
+
     //
     // Functions to create new task(s) and receive result(s)
     //
@@ -1442,13 +1492,27 @@ struct server_context {
         // otherwise, it's a multiple-prompt task, we break it into smaller tasks
         else if (prompt.is_array()) {
             std::vector<json> prompts = prompt;
-            for (size_t i = 0; i < prompts.size(); i++) {
+            if (cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
-                const auto & e = prompts[i];
+                // prompts[0] is the question
-                if (e.is_string() || json_is_array_of_numbers(e)) {
+                // the rest are the answers/documents
-                    data["index"] = i;
+                SRV_DBG("creating rerank tasks, n_prompts = %d\n", (int) prompts.size() - 1);
-                    create_task(data, true, e);
+                for (size_t i = 1; i < prompts.size(); i++) {
-                } else {
+                    json qd;
-                    throw std::runtime_error(error_msg);
+                    qd.push_back(prompts[0]);
+                    qd.push_back(prompts[i]);
+                    data["index"] = i - 1;
+                    create_task(data, true, qd);
+                }
+            } else {
+                SRV_DBG("creating multi-prompt tasks, n_prompts = %d\n", (int) prompts.size());
+                for (size_t i = 0; i < prompts.size(); i++) {
+                    const auto & e = prompts[i];
+                    if (e.is_string() || json_is_array_of_numbers(e)) {
+                        data["index"] = i;
+                        create_task(data, true, e);
+                    } else {
+                        throw std::runtime_error(error_msg);
+                    }
                 }
             }
         }
@@ -1492,7 +1556,9 @@ struct server_context {
                 return;
             }
 
-            size_t idx = result.data["index"];
+            const size_t idx = result.data["index"];
+            GGML_ASSERT(idx < results.size() && "index out of range");
+
             results[idx] = result;
         }
         result_handler(results);
@@ -1903,6 +1969,7 @@ struct server_context {
         // track if this is an embedding or non-embedding batch
         // if we've added sampled tokens above, we are in non-embedding mode
         // -1: none, 0: non-embedding, 1: embedding
+        // TODO: make enum
         int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
 
         // next, batch any pending prompts without exceeding n_batch
@@ -1951,6 +2018,29 @@ struct server_context {
                             }
 
                             prompt_tokens = embd_inp;
+                        } else if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
+                            // require slot.prompt to be array of 2 strings
+                            if (!slot.prompt.is_array() || slot.prompt.size() != 2) {
+                                SLT_ERR(slot, "%s", "invalid prompt for rerank task\n");
+                                slot.release();
+                                send_error(slot, "invalid prompt for rerank task", ERROR_TYPE_INVALID_REQUEST);
+                                continue;
+                            }
+
+                            // prompt: <s>query</s><s>doc</s>
+                            prompt_tokens.clear();
+                            prompt_tokens.push_back(llama_token_bos(model));
+                            {
+                                const auto part = tokenize(slot.prompt[0], false);
+                                prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
+                            }
+                            prompt_tokens.push_back(llama_token_eos(model));
+                            prompt_tokens.push_back(llama_token_bos(model));
+                            {
+                                const auto part = tokenize(slot.prompt[1], false);
+                                prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
+                            }
+                            prompt_tokens.push_back(llama_token_eos(model));
                         } else {
                             prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt
                         }
@@ -1970,7 +2060,7 @@ struct server_context {
                             continue;
                         }
 
-                        if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
+                        if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING || slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
                             // this prompt is too large to process - discard it
                             if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.release();
@@ -2048,7 +2138,8 @@ struct server_context {
                         slot.n_prompt_tokens_processed = 0;
                     }
 
-                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
+                    // non-causal tasks require to fit the entire prompt in the physical batch
+                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING || slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
                         // cannot fit the prompt in the current batch - will try next iter
                         if (batch.n_tokens + slot.n_prompt_tokens > n_batch) {
                             continue;
@@ -2056,7 +2147,10 @@ struct server_context {
                     }
 
                     // check that we are in the right batch_type, if not defer the slot
-                    bool slot_type = slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING ? 1 : 0;
+                    const bool slot_type =
+                        slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING ||
+                        slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK     ? 1 : 0;
+
                     if (batch_type == -1) {
                         batch_type = slot_type;
                     } else if (batch_type != slot_type) {
@@ -2229,6 +2323,13 @@ struct server_context {
                         continue; // continue loop of slots
                     }
 
+                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
+                        send_rerank(slot, batch_view);
+                        slot.release();
+                        slot.i_batch = -1;
+                        continue; // continue loop of slots
+                    }
+
                     // prompt evaluated for next-token prediction
                     slot.state = SLOT_STATE_GENERATING;
                 } else if (slot.state != SLOT_STATE_GENERATING) {
@@ -2787,8 +2888,8 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_cmpl_type cmpl_type, json & data, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params.embedding || ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2848,8 +2949,8 @@ int main(int argc, char ** argv) {
 
     // TODO: maybe merge this function with "handle_completions_generic"
     const auto handle_chat_completions = [&ctx_server, &params, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params.embedding || ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2973,6 +3074,11 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
+        // TODO: somehow clean up this checks in the future
+        if (!ctx_server.params.embedding || ctx_server.params.reranking) {
+            res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings` and without `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
         const json body = json::parse(req.body);
         bool is_openai = false;
 
@@ -3023,6 +3129,79 @@ int main(int argc, char ** argv) {
         res_ok(res, root);
     };
 
+    const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
+        if (!ctx_server.params.reranking) {
+            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
+        const json body = json::parse(req.body);
+
+        // TODO: implement
+        //int top_n = 1;
+        //if (body.count("top_n") != 1) {
+        //    top_n = body.at("top_n");
+        //} else {
+        //    res_error(res, format_error_response("\"top_n\" must be provided", ERROR_TYPE_INVALID_REQUEST));
+        //    return;
+        //}
+
+        json query;
+        if (body.count("query") == 1) {
+            query = body.at("query");
+            if (!query.is_string()) {
+                res_error(res, format_error_response("\"query\" must be a string", ERROR_TYPE_INVALID_REQUEST));
+                return;
+            }
+        } else {
+            res_error(res, format_error_response("\"query\" must be provided", ERROR_TYPE_INVALID_REQUEST));
+            return;
+        }
+
+        std::vector<std::string> documents = json_value(body, "documents", std::vector<std::string>());
+        if (documents.empty()) {
+            res_error(res, format_error_response("\"documents\" must be a non-empty string array", ERROR_TYPE_INVALID_REQUEST));
+            return;
+        }
+
+        // construct prompt object: array of ["query", "doc0", "doc1", ...]
+        json prompt;
+        prompt.push_back(query);
+        for (const auto & doc : documents) {
+            prompt.push_back(doc);
+        }
+
+        LOG_DBG("rerank prompt: %s\n", prompt.dump().c_str());
+
+        // create and queue the task
+        json responses = json::array();
+        bool error = false;
+        {
+            std::vector<server_task> tasks = ctx_server.create_tasks_cmpl({{"prompt", prompt}}, SERVER_TASK_CMPL_TYPE_RERANK);
+            ctx_server.queue_results.add_waiting_tasks(tasks);
+            ctx_server.queue_tasks.post(tasks);
+
+            // get the result
+            std::unordered_set<int> task_ids = server_task::get_list_id(tasks);
+
+            ctx_server.receive_cmpl_results(task_ids, [&](std::vector<server_task_result> & results) {
+                for (const auto & res : results) {
+                    responses.push_back(res.data);
+                }
+            }, [&](const json & error_data) {
+                res_error(res, error_data);
+                error = true;
+            });
+        }
+
+        if (error) {
+            return;
+        }
+
+        // write JSON response
+        json root = format_response_rerank(body, responses);
+        res_ok(res, root);
+    };
+
     const auto handle_lora_adapters_list = [&](const httplib::Request &, httplib::Response & res) {
         json result = json::array();
         for (size_t i = 0; i < ctx_server.loras.size(); ++i) {
@@ -3119,6 +3298,10 @@ int main(int argc, char ** argv) {
     svr->Post("/embedding",           handle_embeddings); // legacy
     svr->Post("/embeddings",          handle_embeddings);
     svr->Post("/v1/embeddings",       handle_embeddings);
+    svr->Post("/rerank",              handle_rerank);
+    svr->Post("/reranking",           handle_rerank);
+    svr->Post("/v1/rerank",           handle_rerank);
+    svr->Post("/v1/reranking",        handle_rerank);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
     // LoRA adapters hotswap

examples/server/tests/features/embeddings.feature

@@ -15,7 +15,7 @@ Feature: llama.cpp server
     And   128 as batch size
     And   128 as ubatch size
     And   512 KV cache size
-    And   embeddings extraction
+    And   enable embeddings endpoint
     Then  the server is starting
     Then  the server is healthy
 

examples/server/tests/features/rerank.feature

@@ -0,0 +1,42 @@
+@llama.cpp
+@rerank
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/models/resolve/main/jina-reranker-v1-tiny-en/ggml-model-f16.gguf
+    And   a model file jina-reranker-v1-tiny-en.gguf
+    And   a model alias jina-reranker-v1-tiny-en
+    And   42 as server seed
+    And   2 slots
+    And   512 as batch size
+    And   512 as ubatch size
+    And   512 KV cache size
+    And   enable reranking endpoint
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Rerank
+    Given a rerank query:
+      """
+      Machine learning is
+      """
+    And   a rerank document:
+      """
+      A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.
+      """
+    And   a rerank document:
+      """
+      Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.
+      """
+    And   a rerank document:
+      """
+      Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.
+      """
+    And   a rerank document:
+      """
+      Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine.
+      """
+    When  reranking request
+    Then  reranking results are returned
+    Then  reranking highest score is index 2 and lowest score is index 3

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

@@ -68,6 +68,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.server_api_key = None
     context.server_continuous_batching = False
     context.server_embeddings = False
+    context.server_reranking = False
     context.server_metrics = False
     context.server_process = None
     context.seed = None
@@ -83,6 +84,10 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.concurrent_tasks = []
     context.prompts = []
 
+    context.reranking_query = None
+    context.reranking_documents = []
+    context.reranking_results = None
+
 
 @step('a model file {hf_file} from HF repo {hf_repo}')
 def step_download_hf_model(context, hf_file: str, hf_repo: str):
@@ -172,10 +177,13 @@ def step_server_continuous_batching(context):
     context.server_continuous_batching = True
 
 
-@step('embeddings extraction')
+@step('enable embeddings endpoint')
 def step_server_embeddings(context):
     context.server_embeddings = True
 
+@step('enable reranking endpoint')
+def step_server_reranking(context):
+    context.server_reranking = True
 
 @step('prometheus compatible metrics exposed')
 def step_server_metrics(context):
@@ -452,6 +460,14 @@ def step_impl(context, n_ga_w):
 def step_prompt_passkey(context):
     context.prompt_passkey = context_text(context)
 
+@step('a rerank query')
+def step_set_rerank_query(context):
+    context.reranking_query = context_text(context)
+    context.reranking_documents = []
+
+@step('a rerank document')
+def step_set_rerank_document(context):
+    context.reranking_documents.append(context_text(context))
 
 @step('{n_prompts:d} fixed prompts')
 def step_fixed_prompts(context, n_prompts):
@@ -619,6 +635,22 @@ async def step_compute_embedding(context):
     context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url)
 
 
+@step('reranking request')
+@async_run_until_complete
+async def step_compute_reranking(context):
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
+        async with session.post(f'{context.base_url}/reranking',
+                                json={
+                                    "query": context.reranking_query,
+                                    "documents": context.reranking_documents,
+                                }) as response:
+            if response.status == 200:
+                response_json = await response.json()
+                context.reranking_results = response_json['results']
+            else:
+                context.reranking_results = response.status
+
+
 @step('all embeddings are the same')
 @async_run_until_complete
 async def step_all_embeddings_are_the_same(context):
@@ -704,6 +736,24 @@ async def all_embeddings_are_generated(context):
     for i in range(n_embedding_requests):
         assert_embeddings(context.tasks_result.pop().pop())
 
+@step('reranking results are returned')
+def reranking_results_are_returned(context):
+    assert len(context.reranking_results) == len(context.reranking_documents)
+
+@step('reranking highest score is index {idx_high:d} and lowest score is index {idx_low:d}')
+def reranking_results_are_returned(context, idx_high: int, idx_low: int):
+    max_score, max_idx = 0, 0
+    min_score, min_idx = 0, 0
+    for res in context.reranking_results:
+        if max_score < res['relevance_score']:
+            max_score = res['relevance_score']
+            max_idx   = res['index']
+        if min_score > res['relevance_score']:
+            min_score = res['relevance_score']
+            min_idx   = res['index']
+    print(context.reranking_results)
+    assert max_idx == idx_high
+    assert min_idx == idx_low
 
 @step('adding special tokens')
 def step_tokenize_set_add_special(context):
@@ -1362,6 +1412,8 @@ def start_server_background(context):
         server_args.append('--cont-batching')
     if context.server_embeddings:
         server_args.append('--embedding')
+    if context.server_reranking:
+        server_args.append('--reranking')
     if context.server_metrics:
         server_args.append('--metrics')
     if context.model_alias:

examples/server/utils.hpp

@@ -537,7 +537,7 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
     json res = json {
         {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
         {"object", "list"},
-        {"usage", json {
+        {"usage", json { // TODO: fill
             {"prompt_tokens", 0},
             {"total_tokens", 0}
         }},
@@ -547,6 +547,29 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
     return res;
 }
 
+static json format_response_rerank(const json & request, const json & ranks) {
+    json data = json::array();
+    int i = 0;
+    for (const auto & rank : ranks) {
+        data.push_back(json{
+            {"index",    i++},
+            {"relevance_score", json_value(rank, "score", 0.0)},
+        });
+    }
+
+    json res = json {
+        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+        {"object", "list"},
+        {"usage", json { // TODO: fill
+            {"prompt_tokens", 0},
+            {"total_tokens", 0}
+        }},
+        {"results", data}
+    };
+
+    return res;
+}
+
 static bool is_valid_utf8(const std::string & str) {
     const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
     const unsigned char* end = bytes + str.length();

ggml/include/ggml.h

@@ -229,14 +229,16 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
-#ifndef GGML_MAX_NAME
-#define GGML_MAX_NAME           64
 #define GGML_MAX_N_THREADS      512
-
-#endif
 #define GGML_MAX_OP_PARAMS      64
+
+#ifndef GGML_MAX_NAME
+#   define GGML_MAX_NAME        64
+#endif
+
 #define GGML_DEFAULT_N_THREADS  4
 #define GGML_DEFAULT_GRAPH_SIZE 2048
+
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
 #else
@@ -259,21 +261,21 @@
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
 #ifndef NDEBUG
-#define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
+#   define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
-#define GGML_UNREACHABLE() __builtin_unreachable()
+#   define GGML_UNREACHABLE() __builtin_unreachable()
 #elif defined(_MSC_VER)
-#define GGML_UNREACHABLE() __assume(0)
+#   define GGML_UNREACHABLE() __assume(0)
 #else
-#define GGML_UNREACHABLE() ((void) 0)
+#   define GGML_UNREACHABLE() ((void) 0)
 #endif
 
 #ifdef __cplusplus
-#define GGML_NORETURN [[noreturn]]
+#   define GGML_NORETURN [[noreturn]]
 #elif defined(_MSC_VER)
-#define GGML_NORETURN __declspec(noreturn)
+#   define GGML_NORETURN __declspec(noreturn)
 #else
-#define GGML_NORETURN _Noreturn
+#   define GGML_NORETURN _Noreturn
 #endif
 
 #define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
@@ -2507,6 +2509,9 @@ extern "C" {
     GGML_API int ggml_cpu_has_cann       (void);
     GGML_API int ggml_cpu_has_llamafile  (void);
 
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
+
     //
     // Internal types and functions exposed for tests and benchmarks
     //

ggml/src/ggml-aarch64.c

@@ -598,15 +598,6 @@ size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_
     return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
 }
 
-// Return the number of byte lanes in the SVE vector if SVE is supported; otherwise, returns 0 if SVE is not supported.
-static int sve_lane_count(void) {
-#if defined(__ARM_FEATURE_SVE)
-    return ggml_sve_cnt_b;
-#else
-    return 0;
-#endif
-}
-
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -843,7 +834,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_cpu_has_sve() && sve_lane_count() == QK8_0) {
+    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2020,7 +2011,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && sve_lane_count() == QK8_0) {
+    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;

ggml/src/ggml-cuda/im2col.cu

@@ -69,7 +69,6 @@ void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 

ggml/src/ggml-quants.c

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

ggml/src/ggml-quants.h

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

ggml/src/ggml-vulkan.cpp

@@ -20,6 +20,8 @@
 #include <unordered_map>
 #include <memory>
 #include <mutex>
+#include <future>
+#include <thread>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -607,13 +609,16 @@ typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx
 
 GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend);
 
-static void ggml_vk_create_pipeline(vk_device& device, vk_pipeline& pipeline, const std::string& name, size_t spv_size, const void* spv_data, const std::string& entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
+// variables to track number of compiles in progress
+static uint32_t compile_count = 0;
+static std::mutex compile_count_mutex;
+static std::condition_variable compile_count_cond;
+
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align) {
     VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
-    std::lock_guard<std::mutex> guard(device->mutex);
-
     pipeline = std::make_shared<vk_pipeline_struct>();
     pipeline->name = name;
     pipeline->parameter_count = parameter_count;
@@ -681,7 +686,17 @@ static void ggml_vk_create_pipeline(vk_device& device, vk_pipeline& pipeline, co
         pipeline->layout);
     pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
 
-    device->pipelines.insert({ pipeline->name, pipeline });
+    {
+        std::lock_guard<std::mutex> guard(device->mutex);
+        device->pipelines.insert({ pipeline->name, pipeline });
+    }
+
+    {
+        std::lock_guard<std::mutex> guard(compile_count_mutex);
+        assert(compile_count > 0);
+        compile_count--;
+    }
+    compile_count_cond.notify_all();
 }
 
 static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) {
@@ -1079,7 +1094,8 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
             // Fall back to host memory type
             buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
         } else {
-            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
+            // use rebar if available, otherwise fallback to device only visible memory
+            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
         }
     } catch (const vk::SystemError& e) {
         std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
@@ -1193,6 +1209,20 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
+    std::vector<std::future<void>> compiles;
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
+        {
+            // wait until fewer than N compiles are in progress
+            uint32_t N = std::max(1u, std::thread::hardware_concurrency());
+            std::unique_lock<std::mutex> guard(compile_count_mutex);
+            while (compile_count >= N) {
+                compile_count_cond.wait(guard);
+            }
+            compile_count++;
+        }
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align));
+    };
+
     if (device->fp16) {
         ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
         ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
@@ -1742,6 +1772,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
+
+    for (auto &c : compiles) {
+        c.wait();
+    }
 }
 
 static vk_device ggml_vk_get_device(size_t idx) {
@@ -2806,7 +2840,11 @@ static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t
 
 static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")");
-    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+
+    // If the device is not an UMA device the memory is host-accessible through rebar. While writing
+    // through PCIe is sufficient fast reading back data from PCIe is slower than going through
+    // the HW device to host copy path.
+    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
         GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
         memcpy(dst, (uint8_t *) src->ptr + offset, size);
@@ -5008,6 +5046,8 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         }
     }
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -5124,7 +5164,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
 
     avg_err /= m * n;
 
-    std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms avg_err=" << avg_err << std::endl;
+    double tflops = 2.0*m*n*k*batch*num_it / (time / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
 
     if (avg_err > 0.1) {
         std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
@@ -5246,12 +5288,14 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 
     ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
 
     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
-    ggml_vk_dispatch_pipeline(ctx, subctx, p, { { qx_buf, 0, qx_sz }, { x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
     ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
@@ -5378,6 +5422,8 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         }
     }
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
@@ -5445,7 +5491,9 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 
     avg_err /= m * n;
 
-    std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms avg_err=" << avg_err << std::endl;
+    double tflops = 2.0*m*n*k*batch*num_it / (time_ms / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
 
     if (avg_err > 0.01 || std::isnan(avg_err)) {
         std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
@@ -5497,9 +5545,6 @@ static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor)
 
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 #if defined(GGML_VULKAN_RUN_TESTS)
-    ctx->staging = ggml_vk_create_buffer_check(ctx->device, 100ul * 1024ul * 1024ul,
-        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
-        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);

ggml/src/ggml.c

@@ -39,9 +39,6 @@
 #include <unistd.h>
 #endif
 
-#if defined(__ARM_FEATURE_SVE)
-int ggml_sve_cnt_b = 0;
-#endif
 #if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
@@ -455,6 +452,15 @@ static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];
 
+#if defined(__ARM_ARCH)
+struct ggml_arm_arch_features_type {
+    int has_neon;
+    int has_i8mm;
+    int has_sve;
+    int sve_cnt;
+} ggml_arm_arch_features = {-1, -1, -1, 0};
+#endif
+
 GGML_CALL const char * ggml_status_to_string(enum ggml_status status) {
     switch (status) {
         case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
@@ -3673,6 +3679,70 @@ static inline int ggml_up(int n, int m) {
 
 ////////////////////////////////////////////////////////////////////////////////
 
+#if defined(__ARM_ARCH)
+
+#if defined(__linux__) && defined(__aarch64__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM 0
+#endif
+
+static void ggml_init_arm_arch_features(void) {
+#if defined(__linux__) && defined(__aarch64__)
+    uint32_t hwcap = getauxval(AT_HWCAP);
+    uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+    ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#endif
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_neon = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_i8mm = oldp;
+
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#else
+// Run-time CPU feature detection not implemented for this platform, fallback to compile time
+#if defined(__ARM_NEON)
+    ggml_arm_arch_features.has_neon = 1;
+#else
+    ggml_arm_arch_features.has_neon = 0;
+#endif
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_arm_arch_features.has_i8mm = 1;
+#else
+    ggml_arm_arch_features.has_i8mm = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.has_sve = 1;
+    ggml_arm_arch_features.sve_cnt = 16;
+#else
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#endif
+#endif
+}
+#endif
+
 struct ggml_context * ggml_init(struct ggml_init_params params) {
     // make this function thread safe
     ggml_critical_section_start();
@@ -3723,6 +3793,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
+#if defined(__ARM_ARCH)
+        ggml_init_arm_arch_features();
+#endif
+
         is_first_call = false;
     }
 
@@ -3771,12 +3845,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_ASSERT_ALIGNED(ctx->mem_buffer);
 
-#if defined(__ARM_FEATURE_SVE)
-    if (!ggml_sve_cnt_b) {
-        ggml_sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-    }
-#endif
-
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
     ggml_critical_section_end();
@@ -23578,16 +23646,16 @@ int ggml_cpu_has_fma(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_NEON)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
 }
 
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_sve;
 #else
     return 0;
 #endif
@@ -23734,11 +23802,18 @@ int ggml_cpu_has_vsx(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_ARCH)
-    return 1;
+    return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
 #endif
 }
 
+int ggml_cpu_get_sve_cnt(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.sve_cnt;
+#else
+    return 0;
+#endif
+}
 ////////////////////////////////////////////////////////////////////////////////

ggml/src/vulkan-shaders/argsort.comp

@@ -29,20 +29,18 @@ void main() {
     const int col = int(gl_LocalInvocationID.x);
     const uint row = gl_WorkGroupID.y;
 
-    if (col >= p.ncols_pad) {
-        return;
-    }
-
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    dst_row[col] = col;
+    if (col < p.ncols_pad) {
+        dst_row[col] = col;
+    }
     barrier();
 
     for (uint k = 2; k <= p.ncols_pad; k *= 2) {
         for (uint j = k / 2; j > 0; j /= 2) {
             const uint ixj = col ^ j;
-            if (ixj > col) {
+            if (col < p.ncols_pad && ixj > col) {
                 if ((col & k) == 0) {
                     if (dst_row[col] >= p.ncols ||
                         (dst_row[ixj] < p.ncols && (p.order == ASC ?

gguf-py/gguf/constants.py

@@ -94,6 +94,7 @@ class Keys:
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
         ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
         FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
+        SWIN_NORM                         = "{arch}.swin_norm"
         RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
         TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
         TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
@@ -236,6 +237,7 @@ class MODEL_ARCH(IntEnum):
     EXAONE       = auto()
     GRANITE      = auto()
     GRANITE_MOE  = auto()
+    CHAMELEON    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -343,6 +345,8 @@ class MODEL_TENSOR(IntEnum):
     ENC_FFN_DOWN         = auto()
     ENC_FFN_UP           = auto()
     ENC_OUTPUT_NORM      = auto()
+    CLS                  = auto() # classifier
+    CLS_OUT              = auto() # classifier output projection
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -394,6 +398,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.EXAONE:         "exaone",
     MODEL_ARCH.GRANITE:        "granite",
     MODEL_ARCH.GRANITE_MOE:    "granitemoe",
+    MODEL_ARCH.CHAMELEON:      "chameleon",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -501,6 +506,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ENC_FFN_DOWN:              "enc.blk.{bid}.ffn_down",
     MODEL_TENSOR.ENC_FFN_UP:                "enc.blk.{bid}.ffn_up",
     MODEL_TENSOR.ENC_OUTPUT_NORM:           "enc.output_norm",
+    MODEL_TENSOR.CLS:                       "cls",
+    MODEL_TENSOR.CLS_OUT:                   "cls.output",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -610,6 +617,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
     ],
     MODEL_ARCH.NOMIC_BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -641,6 +650,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
     ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -1260,6 +1270,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.CHAMELEON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -670,6 +670,9 @@ class GGUFWriter:
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 
+    def add_swin_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
+
     def add_rescale_every_n_layers(self, count: int) -> None:
         self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
 

gguf-py/gguf/tensor_mapping.py

@@ -380,7 +380,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -389,7 +389,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
@@ -679,6 +679,15 @@ class TensorNameMap:
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
         ),
+
+        MODEL_TENSOR.CLS: (
+            "classifier",       # jina
+            "classifier.dense", # roberta
+        ),
+
+        MODEL_TENSOR.CLS_OUT: (
+            "classifier.out_proj", # roberta
+        ),
     }
 
     # architecture-specific block mappings

include/llama.h

@@ -102,6 +102,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
     };
 
     enum llama_rope_type {
@@ -192,6 +193,7 @@ extern "C" {
         LLAMA_POOLING_TYPE_MEAN = 1,
         LLAMA_POOLING_TYPE_CLS  = 2,
         LLAMA_POOLING_TYPE_LAST = 3,
+        LLAMA_POOLING_TYPE_RANK = 4, // used by reranking models to attach the classification head to the graph
     };
 
     enum llama_attention_type {
@@ -201,9 +203,9 @@ extern "C" {
     };
 
     enum llama_split_mode {
-        LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
+        LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
-        LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
+        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -871,7 +873,8 @@ extern "C" {
 
     // Get the embeddings for a sequence id
     // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // shape: [n_embd] (1-dimensional)
+    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+    // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 
     //
@@ -910,6 +913,8 @@ extern "C" {
     //
     // Tokenization
     //
+    // The API is thread-safe.
+    //
 
     /// @details Convert the provided text into tokens.
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.

models/ggml-vocab-chameleon.gguf.inp

@@ -0,0 +1,112 @@
+ied 4 ½ months
+__ggml_vocab_test__
+Führer
+__ggml_vocab_test__
+
+__ggml_vocab_test__
+ 
+__ggml_vocab_test__
+  
+__ggml_vocab_test__
+   
+__ggml_vocab_test__
+	
+__ggml_vocab_test__
+
+
+__ggml_vocab_test__
+
+
+
+__ggml_vocab_test__
+
+
+
+
+__ggml_vocab_test__
+	
+
+__ggml_vocab_test__
+Hello world
+__ggml_vocab_test__
+ Hello world
+__ggml_vocab_test__
+Hello World
+__ggml_vocab_test__
+ Hello World
+__ggml_vocab_test__
+ Hello World!
+__ggml_vocab_test__
+Hello, world!
+__ggml_vocab_test__
+ Hello, world!
+__ggml_vocab_test__
+ this is 🦙.cpp
+__ggml_vocab_test__
+w048 7tuijk dsdfhu
+__ggml_vocab_test__
+нещо на Български
+__ggml_vocab_test__
+កាន់តែពិសេសអាចខលចេញ
+__ggml_vocab_test__
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
+__ggml_vocab_test__
+Hello
+__ggml_vocab_test__
+ Hello
+__ggml_vocab_test__
+  Hello
+__ggml_vocab_test__
+   Hello
+__ggml_vocab_test__
+    Hello
+__ggml_vocab_test__
+    Hello
+    Hello
+__ggml_vocab_test__
+ (
+__ggml_vocab_test__
+
+ =
+__ggml_vocab_test__
+' era
+__ggml_vocab_test__
+Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
+__ggml_vocab_test__
+!!!!!!
+__ggml_vocab_test__
+3
+__ggml_vocab_test__
+33
+__ggml_vocab_test__
+333
+__ggml_vocab_test__
+3333
+__ggml_vocab_test__
+33333
+__ggml_vocab_test__
+333333
+__ggml_vocab_test__
+3333333
+__ggml_vocab_test__
+33333333
+__ggml_vocab_test__
+333333333
+__ggml_vocab_test__
+Cửa Việt
+__ggml_vocab_test__
+ discards
+__ggml_vocab_test__
+
+ 
+
+ 
+
+
+ 	 		 	
+  
+   
+    
+     
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
+__ggml_vocab_test__

models/ggml-vocab-chameleon.gguf.out

@@ -0,0 +1,46 @@
+ 17245 16604 16403 16604 33583 18355
+ 16421 51153
+
+ 16604
+ 16650
+ 16650 16604
+ 16581
+ 16582
+ 16582 16582
+ 16582 16582 16582
+ 16581 16582
+ 31596 17394
+ 34926 17394
+ 31596 18671
+ 34926 18671
+ 34926 18671 16384
+ 31596 16395 17394 16384
+ 34926 16395 17394 16384
+ 16811 16704 20410 16483 16631 16397 52854
+ 16470 16399 16403 16407 16604 16406 35764 38185 51595 22592 26639
+ 29479 23955 17012 20103 25527 27670 17408 19005 21473 24774
+ 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 21954 16607 21954 16633 21954 16611 29409 16607 21954 16615
+ 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 16604 16391 24664 17153 57169 16721 16872 17073 17304 28729 16392
+ 31596
+ 34926
+ 16650 31596
+ 16650 34926
+ 16696 31596
+ 16696 31596 16582 16696 31596
+ 16604 16391
+ 16582 16604 16412
+ 16390 22623
+ 31596 16395 16712 16390 16828 16384 17674 16769 16732 23686 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636
+ 16384 16384 16384 16384 16384 16384
+ 16402
+ 16402 16402
+ 16402 16402 16402
+ 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402 16402 16402
+ 16418 19038 16639 16448 24315 33727 16467
+ 18765 17981
+ 16582 16604 16582 16582 16604 16582 16582 16582 16604 16581 16604 16581 16581 16604 16581 16582 16650 16582 16650 16604 16582 16696 16582 16696 16604 16582 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 20410 16483 16631 18885 16483 16631 16604 16402 16604 16402 16402 16604 16402 16402 16402 16604 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16402 16604 16402 16397 16402 16604 16402 16397 16397 16402 16604 16402 16397 16397 16397 16402 16604 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 27683 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636 16604 16396 16396 16396 16396 16396 16396 16412 16412 16412 16412 16412 16412 16412 27268 23955 17012 20103 25527 27670 17408 19005 21473 24774 16604 16390 16390 16390 16390 16390 16390 16447 16447 16447 16447 16447 16447 16447 16385 16385 16385 16385 16397 16397 16397 16397 16397 16397 16384 16384 16384 16384 16384 16384 16414 16414 16414 16414 16414 16414 16687 16390 16690 16992 16604 16390 61797 16733 16390 16466 16986 16395 16604 16390 17879 16732 17811 16414 16604 16390 16428 16804 17811 16687 16390 16683 17190 16728 16395 16604 16390 16419 16732 16945 16991 25251 16414 17119 16390 38127 16641 16390 16459 16427

scripts/sync-ggml.last

@@ -1 +1 @@
-336c10a4c3c8ec99af484b25a0cddd397a09cdb2
+9a24b8c8c40eab7262d067e91d08df160678df8d

src/llama-vocab.cpp

@@ -50,7 +50,7 @@ struct naive_trie {
             res.first->second.insert(key + 1, len - 1, value);
         }
     }
-    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) {
+    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) const {
         if (len == 0 || offset == len) {
             return std::make_pair(key, offset);
         }
@@ -79,6 +79,15 @@ struct naive_trie {
 // impl
 //
 
+struct llm_tokenizer {
+   llm_tokenizer() {}
+   virtual ~llm_tokenizer() = default;
+};
+
+llama_vocab::~llama_vocab() {
+    delete tokenizer;
+}
+
 int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
     GGML_ASSERT(token_left.find(' ')   == std::string::npos);
     GGML_ASSERT(token_left.find('\n')  == std::string::npos);
@@ -187,10 +196,15 @@ struct llm_bigram_spm {
     size_t size;
 };
 
-struct llm_tokenizer_spm {
+struct llm_tokenizer_spm : llm_tokenizer {
-    llm_tokenizer_spm(const llama_vocab & vocab) : vocab(vocab) {}
+    llm_tokenizer_spm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
+
+struct llm_tokenizer_spm_session {
+    llm_tokenizer_spm_session(const llama_vocab & vocab) : vocab(vocab) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+
         // split string into utf8 chars
         int index = 0;
         size_t offs = 0;
@@ -271,7 +285,7 @@ private:
             return;
         }
 
-        resegment(symbols[p->second.first],  output);
+        resegment(symbols[p->second.first], output);
         resegment(symbols[p->second.second], output);
     }
 
@@ -279,7 +293,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n);
         auto token = vocab.token_to_id.find(text);
 
@@ -306,10 +319,11 @@ private:
     }
 
     const llama_vocab & vocab;
+    // currently unused
+    // const llm_tokenizer_spm * spm_tokenizer;
 
     std::vector<llm_symbol> symbols;
     llm_bigram_spm::queue work_queue;
-
     std::map<std::string, std::pair<int, int>> rev_merge;
 };
 
@@ -352,8 +366,8 @@ struct llm_bigram_bpe {
     size_t size;
 };
 
-struct llm_tokenizer_bpe {
+struct llm_tokenizer_bpe : llm_tokenizer {
-    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {
+    llm_tokenizer_bpe(const llama_vocab & vocab) : llm_tokenizer() {
         GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE);
         switch (vocab.type_pre) {
             case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
@@ -450,6 +464,20 @@ struct llm_tokenizer_bpe {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_CHAMELEON:
+                // Note: in theory, the special token (sentinel and image token) regex_exprs below
+                // are unnecessary, as they are split in `tokenizer_st_partition` anyway.
+                // However, since the upstream pre-tokenizer uses them, they are also
+                // included here (see https://huggingface.co/facebook/chameleon-7b).
+                regex_exprs = {
+                    "<sentinel:[0-9]+>",  // Sentinel tokens
+                    "(IMGIMG)((A|B|C|D|E|F|G|H|I){1,4})Z",  // Image tokens
+                    "([\\t\\n]|    |  )",  // directly from tokenizer.json
+                    "\\p{N}", // Individual digits
+                    "[\\p{P}!-/:-@\\[-`{-~]",  // Punctuation, Isolated
+                    "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -462,7 +490,14 @@ struct llm_tokenizer_bpe {
         }
     }
 
-    void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output) const {
+    std::vector<std::string> regex_exprs;
+};
+
+struct llm_tokenizer_bpe_session {
+    llm_tokenizer_bpe_session(const llama_vocab & vocab) : vocab(vocab),
+        bpe_tokenizer(static_cast<const llm_tokenizer_bpe *>(vocab.tokenizer)) {}
+
+    static void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output)  {
         output.push_back(token_id);
     }
 
@@ -501,12 +536,11 @@ struct llm_tokenizer_bpe {
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         int final_prev_index = -1;
-
+        const auto word_collection = unicode_regex_split(text, bpe_tokenizer->regex_exprs);
-        const auto word_collection = unicode_regex_split(text, regex_exprs);
 
         symbols_final.clear();
 
-        for (auto & word : word_collection) {
+        for (const auto & word : word_collection) {
             work_queue = llm_bigram_bpe::queue();
             symbols.clear();
 
@@ -609,7 +643,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         std::string left_token  = std::string(symbols[left].text,  symbols[left].n);
         std::string right_token = std::string(symbols[right].text, symbols[right].n);
 
@@ -633,12 +666,10 @@ private:
     }
 
     const llama_vocab & vocab;
-
+    const llm_tokenizer_bpe * bpe_tokenizer;
-    std::vector<std::string> regex_exprs;
 
     std::vector<llm_symbol> symbols;
     std::vector<llm_symbol> symbols_final;
-
     llm_bigram_bpe::queue work_queue;
 };
 
@@ -646,15 +677,17 @@ private:
 // WPM tokenizer
 //
 
-struct llm_tokenizer_wpm {
+struct llm_tokenizer_wpm : llm_tokenizer {
-    llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
+    llm_tokenizer_wpm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
 
-    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) const {
+struct llm_tokenizer_wpm_session {
+    llm_tokenizer_wpm_session(const llama_vocab & vocab) : vocab(vocab) {}
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         const auto & token_map = vocab.token_to_id;
-
         // normalize and split by whitespace
         std::vector<std::string> words = preprocess(text);
-
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -699,7 +732,7 @@ struct llm_tokenizer_wpm {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    std::vector<std::string> preprocess(const std::string & text) const {
+    static std::vector<std::string> preprocess(const std::string & text)  {
         const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
         std::vector<std::string> words(1, "");
 
@@ -751,15 +784,18 @@ struct llm_tokenizer_wpm {
             //(cpt >= 0xFF00  && cpt <= 0xFFEF);
     }
 
+private:
     const llama_vocab & vocab;
+    // currently unused
+    // const llm_tokenizer_wpm * wpm_tokenizer;
 };
 
 //
 // UGM tokenizer
 //
 
-struct llm_tokenizer_ugm {
+struct llm_tokenizer_ugm : llm_tokenizer {
-    llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) {
+    llm_tokenizer_ugm(const llama_vocab & vocab) : llm_tokenizer() {
         if (vocab.precompiled_charsmap.size() > 0) {
             size_t charsmap_offset = 0;
 
@@ -805,6 +841,30 @@ struct llm_tokenizer_ugm {
         unknown_token_score = min_score - unknown_token_score_penalty;
     }
 
+    // escaped space symbol - U+2581 (Lower One Eighth Block)
+    const std::string escaped_space = "\xE2\x96\x81";
+
+    const char * prefix_replacements = NULL;
+    size_t prefix_replacements_size = 0;
+
+    const uint32_t * xcda_array = NULL;
+    size_t xcda_array_size = 0;
+
+    struct naive_trie user_defined_token_matcher;
+
+    float min_score = FLT_MAX;
+    float max_score = -FLT_MAX;
+
+    float unknown_token_score_penalty = 10.0;
+    float unknown_token_score;
+
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_ugm_session {
+    llm_tokenizer_ugm_session(const llama_vocab & vocab) : vocab(vocab),
+        ugm_tokenizer(static_cast<const llm_tokenizer_ugm *>(vocab.tokenizer)) {}
+
     /* This implementation is based on SentencePiece optimized Viterbi algorithm for
      * unigram language models. The general idea is to:
      * - move along the input sequence in steps of one UTF code point,
@@ -843,7 +903,7 @@ struct llm_tokenizer_ugm {
             // traverse the token matcher trie to find a matching token
             bool single_codepoint_token_found = false;
             const struct best_tokenization & current_best = tokenization_results[input_offset];
-            const struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]);
+            const struct naive_trie * node = ugm_tokenizer->token_matcher.traverse(normalized[prefix_offset++]);
 
             while (prefix_offset <= input_len && node != NULL) {
                 // check if we found valid token in prefix
@@ -873,7 +933,7 @@ struct llm_tokenizer_ugm {
             // if we didn't find a valid token corresponding to the whole UTF code point
             // then use unknown token as the tokenization of this UTF code point
             if (!single_codepoint_token_found) {
-                const double challenger_score = current_best.score_sum + unknown_token_score;
+                const double challenger_score = current_best.score_sum + ugm_tokenizer->unknown_token_score;
                 prefix_offset = input_offset + n_utf8_code_units;
                 struct best_tokenization & current_champ = tokenization_results[prefix_offset];
                 if (challenger_score > current_champ.score_sum) {
@@ -905,7 +965,6 @@ struct llm_tokenizer_ugm {
     }
 
 private:
-    const llama_vocab & vocab;
 
     // helper structure for returning normalization results
     struct normalization_result {
@@ -918,7 +977,7 @@ private:
         normalized->clear();
         normalized->reserve(input.size() * 3);
 
-        const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " ";
+        const std::string space = vocab.tokenizer_escape_whitespaces ? ugm_tokenizer->escaped_space : " ";
 
         bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
         bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
@@ -1000,13 +1059,21 @@ private:
         size_t xcda_array_size;
     };
 
+    // this structure stores the best tokenization so far at input_offset
+    struct best_tokenization {
+        llama_token token_id;
+        size_t input_offset;
+        float score_sum;
+    };
+
     struct normalization_result normalize_prefix(const std::string & input, size_t input_offset) {
         if (input_offset == input.size()) {
             return { &input[input_offset], 0, 0 };
         }
 
         // if input prefix matches some user-defined token return this token as normalization result
-        auto user_defined_token_match = user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
+        auto user_defined_token_match =
+           ugm_tokenizer->user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
         if (user_defined_token_match.second > 0) {
             return { &input[input_offset], user_defined_token_match.second, user_defined_token_match.second };
         }
@@ -1014,8 +1081,8 @@ private:
         size_t longest_prefix_length = 0;
         size_t longest_prefix_offset = 0;
 
-        if (xcda_array_size > 0) {
+        if (ugm_tokenizer->xcda_array_size > 0) {
-            struct xcda_array_view xcda_view(xcda_array, xcda_array_size);
+            struct xcda_array_view xcda_view(ugm_tokenizer->xcda_array, ugm_tokenizer->xcda_array_size);
 
             // Find the longest normalized sequence matching the input prefix by walking
             // the XOR-compressed compact double array (XCDA) starting from the root node
@@ -1051,50 +1118,27 @@ private:
 
         if (longest_prefix_length > 0) {
             // we have a match, so return the replacement sequence
-            if (longest_prefix_offset >= prefix_replacements_size) {
+            if (longest_prefix_offset >= ugm_tokenizer->prefix_replacements_size) {
                 throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
             }
-            const char * prefix_replacement = &prefix_replacements[longest_prefix_offset];
+            const char * prefix_replacement = &(ugm_tokenizer->prefix_replacements)[longest_prefix_offset];
             return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
-        } else {
+        }
-            // check if the input prefix contains a valid sequence of UTF-8 code units
+
-            try {
+        // check if the input prefix contains a valid sequence of UTF-8 code units
-                // if yes, return this sequence unmodified
+        try {
-                size_t prefix_offset = input_offset;
+            // if yes, return this sequence unmodified
-                unicode_cpt_from_utf8(input, prefix_offset);
+            size_t prefix_offset = input_offset;
-                return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
+            unicode_cpt_from_utf8(input, prefix_offset);
-            } catch (std::invalid_argument & /*ex*/) {
+            return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
-                // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
+        } catch (std::invalid_argument & /*ex*/) {
-                return { "\xEF\xBF\xBD", 3, 1 };
+            // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
-            }
+            return { "\xEF\xBF\xBD", 3, 1 };
         }
     }
 
-    // escaped space symbol - U+2581 (Lower One Eighth Block)
+    const llama_vocab & vocab;
-    const std::string escaped_space = "\xE2\x96\x81";
+    const llm_tokenizer_ugm * ugm_tokenizer;
-
-    const char * prefix_replacements = NULL;
-    size_t prefix_replacements_size = 0;
-
-    const uint32_t * xcda_array = NULL;
-    size_t xcda_array_size = 0;
-
-    struct naive_trie user_defined_token_matcher;
-
-    // this structure stores the best tokenization so far at input_offset
-    struct best_tokenization {
-        llama_token token_id;
-        size_t input_offset;
-        float score_sum;
-    };
-
-    float min_score = FLT_MAX;
-    float max_score = -FLT_MAX;
-
-    float unknown_token_score_penalty = 10.0;
-    float unknown_token_score;
-
-    struct naive_trie token_matcher;
 };
 
 //
@@ -1155,8 +1199,8 @@ static std::vector<uint8_t> llama_unescape_rwkv_token(const std::string & escape
     return output;
 }
 
-struct llm_tokenizer_rwkv {
+struct llm_tokenizer_rwkv : llm_tokenizer {
-    llm_tokenizer_rwkv(const llama_vocab & vocab): vocab(vocab) {
+    llm_tokenizer_rwkv(const llama_vocab & vocab) : llm_tokenizer() {
         // RWKV supports arbitrary byte tokens, but the vocab struct only supports string tokens.
         // For now, we decode the vocab here into the lookup we'll use for tokenization.
 
@@ -1168,11 +1212,17 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_rwkv_session {
+    llm_tokenizer_rwkv_session(const llama_vocab & vocab) : vocab(vocab),
+        rwkv_tokenizer(static_cast<const llm_tokenizer_rwkv &>(*vocab.tokenizer)) {}
+
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         uint32_t position = 0;
-
         while (position < text.size()) {
-            const struct naive_trie * node = token_matcher.traverse(text[position]);
+            const struct naive_trie * node = rwkv_tokenizer.token_matcher.traverse(text[position]);
             if (node == NULL) {
                 // no matching token found, add unknown token
                 output.push_back(vocab.special_unk_id);
@@ -1197,11 +1247,33 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+private:
     const llama_vocab & vocab;
-
+    const llm_tokenizer_rwkv & rwkv_tokenizer;
-    struct naive_trie token_matcher;
 };
 
+void llama_vocab::init_tokenizer() {
+    switch (type) {
+        case LLAMA_VOCAB_TYPE_SPM:
+            tokenizer = new llm_tokenizer_spm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_BPE:
+            tokenizer = new llm_tokenizer_bpe(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_WPM:
+            tokenizer = new llm_tokenizer_wpm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_UGM:
+            tokenizer = new llm_tokenizer_ugm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_RWKV:
+            tokenizer = new llm_tokenizer_rwkv(*this);
+            break;
+        default:
+            GGML_ABORT("unsupported vocab type");
+    }
+}
+
 //
 // (de-) tokenize
 //
@@ -1263,7 +1335,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
             // if a fragment is text ( not yet processed )
             if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                auto & raw_text = fragment.raw_text;
+                const auto & raw_text = fragment.raw_text;
 
                 auto raw_text_base_offset = fragment.offset;
                 auto raw_text_base_length = fragment.length;
@@ -1362,7 +1434,13 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
     }
 }
 
-std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) {
+std::vector<llama_vocab::id> llama_tokenize_internal(
+        const llama_vocab & vocab,
+        std::string raw_text,
+        bool add_special,
+        bool parse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     std::vector<llama_vocab::id> output;
     std::forward_list<fragment_buffer_variant> fragment_buffer;
 
@@ -1399,9 +1477,9 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        llm_tokenizer_spm tokenizer(vocab);
                         llama_escape_whitespace(raw_text);
-                        tokenizer.tokenize(raw_text, output);
+                        llm_tokenizer_spm_session session(vocab);
+                        session.tokenize(raw_text, output);
                         is_prev_special = false;
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
@@ -1423,10 +1501,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                llm_tokenizer_bpe tokenizer(vocab);
+                llm_tokenizer_bpe_session session(vocab);
-
+                // it calls some other methods that are not exist in llm_tokenizer,
+                // here just cast it to bpe tokenizer object
                 if (add_special) {
-                    tokenizer.append_bos(output);
+                    session.append_bos(output);
                 }
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1435,15 +1514,15 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
-                        tokenizer.append(fragment.token, output);
+                        session.append(fragment.token, output);
                     }
                 }
 
                 if (add_special) {
-                    tokenizer.append_eos(output);
+                    session.append_eos(output);
-                    tokenizer.check_double_bos_eos(output);
+                    session.check_double_bos_eos(output);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
@@ -1453,7 +1532,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                     output.push_back(vocab.special_cls_id);
                 }
 
-                llm_tokenizer_wpm tokenizer(vocab);
+                llm_tokenizer_wpm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1462,7 +1541,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1475,12 +1554,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_UGM:
             {
-                llm_tokenizer_ugm tokenizer(vocab);
+                if (add_special && vocab.tokenizer_add_bos) {
-
-                if (add_special && vocab.tokenizer_add_bos != 0) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
                     output.push_back(vocab.special_bos_id);
                 }
+                llm_tokenizer_ugm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1488,26 +1566,27 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
                 }
 
-                if (add_special && vocab.tokenizer_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+                if (add_special && vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) {
                     LLAMA_LOG_WARN(
                         "%s: Added a BOS token to the prompt as specified by the model but the prompt "
                         "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
                         "Are you sure this is what you want?\n", __FUNCTION__);
                 }
 
-                if (add_special && vocab.tokenizer_add_eos == 1) {
+                if (add_special && vocab.tokenizer_add_eos) {
                     GGML_ASSERT(vocab.special_eos_id != -1);
                     output.push_back(vocab.special_eos_id);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_RWKV:
             {
+                llm_tokenizer_rwkv_session session(vocab);
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@@ -1516,8 +1595,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
 
-                        llm_tokenizer_rwkv tokenizer(vocab);
+                        session.tokenize(raw_text, output);
-                        tokenizer.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1630,13 +1708,13 @@ llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
 }
 
 int32_t llama_tokenize_impl(
-    const struct llama_vocab & vocab,
+        const struct llama_vocab & vocab,
-                  const char * text,
+                      const char * text,
-                     int32_t   text_len,
+                         int32_t   text_len,
-                 llama_token * tokens,
+                     llama_token * tokens,
-                     int32_t   n_tokens_max,
+                         int32_t   n_tokens_max,
-                        bool   add_special,
+                            bool   add_special,
-                        bool   parse_special) {
+                            bool   parse_special) {
     auto res = llama_tokenize_internal(vocab, std::string(text, text_len), add_special, parse_special);
     if (n_tokens_max < (int) res.size()) {
         // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
@@ -1713,11 +1791,13 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
                 // suppressing them like CONTROL tokens.
                 if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) {
                     return _try_copy(token_text.data(), token_text.size());
-                } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
+                }
+                if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
                     std::string result = token_text;
                     llama_unescape_whitespace(result);
                     return _try_copy(result.data(), result.size());
-                } else if (attr & LLAMA_TOKEN_ATTR_BYTE) {
+                }
+                if (attr & LLAMA_TOKEN_ATTR_BYTE) {
                     char byte = (char) llama_token_to_byte(vocab, token);
                     return _try_copy((char*) &byte, 1);
                 }
@@ -1728,7 +1808,8 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
                 // suppressing them like CONTROL tokens.
                 if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) {
                     return _try_copy(token_text.data(), token_text.size());
-                } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
+                }
+                if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
                     std::string result = llama_decode_text(token_text);
                     return _try_copy(result.data(), result.size());
                 }
@@ -1761,6 +1842,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     int32_t avail = text_len_max;
     int32_t total = 0;
 

src/llama-vocab.h

@@ -8,6 +8,8 @@
 #include <map>
 #include <set>
 
+struct llm_tokenizer;
+
 struct llama_vocab {
     using id    = llama_token;
     using token = std::string;
@@ -65,7 +67,14 @@ struct llama_vocab {
 
     std::vector<char> precompiled_charsmap;
 
+    llm_tokenizer * tokenizer = nullptr;
+
+    llama_vocab() = default;
+    ~llama_vocab();
+
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+
+    void init_tokenizer();
 };
 
 //

src/llama.cpp

@@ -216,6 +216,7 @@ enum llm_arch {
     LLM_ARCH_RWKV6,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_CHAMELEON,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -268,6 +269,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RWKV6,           "rwkv6"        },
     { LLM_ARCH_GRANITE,         "granite"      },
     { LLM_ARCH_GRANITE_MOE,     "granitemoe"   },
+    { LLM_ARCH_CHAMELEON,       "chameleon"    },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -304,6 +306,7 @@ enum llm_kv {
     LLM_KV_DECODER_START_TOKEN_ID,
     LLM_KV_ATTN_LOGIT_SOFTCAPPING,
     LLM_KV_FINAL_LOGIT_SOFTCAPPING,
+    LLM_KV_SWIN_NORM,
     LLM_KV_RESCALE_EVERY_N_LAYERS,
     LLM_KV_TIME_MIX_EXTRA_DIM,
     LLM_KV_TIME_DECAY_EXTRA_DIM,
@@ -411,6 +414,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
     { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
     { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
+    { LLM_KV_SWIN_NORM,                         "%s.swin_norm"                         },
     { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
     { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
@@ -602,6 +606,8 @@ enum llm_tensor {
     LLM_TENSOR_ENC_FFN_DOWN,
     LLM_TENSOR_ENC_FFN_UP,
     LLM_TENSOR_ENC_OUTPUT_NORM,
+    LLM_TENSOR_CLS,
+    LLM_TENSOR_CLS_OUT,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -789,6 +795,8 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_CLS,             "cls" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
         },
     },
     {
@@ -824,6 +832,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
             { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_CLS,             "cls" },
         },
     },
     {
@@ -1499,6 +1508,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_CHAMELEON,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { 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_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2362,6 +2390,7 @@ struct llama_hparams {
     bool vocab_only;
     bool rope_finetuned;
     bool use_par_res;
+    bool swin_norm;
 
     uint32_t n_vocab;
     uint32_t n_ctx_train; // context size the model was trained on
@@ -2870,6 +2899,7 @@ struct llama_model {
     llama_hparams hparams = {};
     llama_vocab   vocab;
 
+    // TODO: should init all tensors to nullptr
     struct ggml_tensor * tok_embd;
     struct ggml_tensor * type_embd;
     struct ggml_tensor * pos_embd;
@@ -2882,6 +2912,12 @@ struct llama_model {
     struct ggml_tensor * output_b;
     struct ggml_tensor * output_norm_enc;
 
+    // classifier
+    struct ggml_tensor * cls;
+    struct ggml_tensor * cls_b;
+    struct ggml_tensor * cls_out   = nullptr;
+    struct ggml_tensor * cls_out_b = nullptr;
+
     std::vector<llama_layer> layers;
 
     llama_split_mode split_mode;
@@ -5580,11 +5616,11 @@ static void llm_load_hparams(
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type);
-                ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type);
+                ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type, false);
                 hparams.f_max_alibi_bias = 8.0f;
 
                 switch (hparams.n_layer) {
-                    case 4: model.type = e_model::MODEL_33M; break; // jina-embeddings-small
+                    case 4:  model.type = e_model::MODEL_33M;  break; // jina-embeddings-small
                     case 12: model.type = e_model::MODEL_137M; break; // jina-embeddings-base
                 }
             } break;
@@ -6084,6 +6120,18 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_CHAMELEON:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                hparams.f_norm_eps = 1e-5;  // eps for qk-norm, torch default
+                ml.get_key(LLM_KV_SWIN_NORM, hparams.swin_norm);
+
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 48: model.type = e_model::MODEL_34B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+               }
+            } break;
         default: (void)0;
     }
 
@@ -6277,6 +6325,7 @@ static void llm_load_vocab(
                     tokenizer_pre == "phi-2"   ||
                     tokenizer_pre == "jina-es" ||
                     tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
                     tokenizer_pre == "jina-v2-code") {
@@ -6341,6 +6390,11 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "exaone") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
+            } else if (
+                tokenizer_pre == "chameleon") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
+                vocab.tokenizer_add_bos = true;
+                vocab.tokenizer_clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -6398,7 +6452,12 @@ static void llm_load_vocab(
 
     for (uint32_t i = 0; i < n_vocab; i++) {
         std::string word = gguf_get_arr_str(ctx, token_idx, i);
-        GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);
+
+        //GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);
+        if (word.empty()) {
+            LLAMA_LOG_WARN("%s: empty token at index %u\n", __func__, i);
+            word = "[EMPTY_" + std::to_string(i) + "]";
+        }
 
         vocab.token_to_id[word] = i;
         vocab.max_token_len = std::max(vocab.max_token_len, (int) word.size());
@@ -6423,6 +6482,8 @@ static void llm_load_vocab(
     }
     GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
 
+    vocab.init_tokenizer();
+
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         // For Fill-In-the-Middle (FIM)/infill models which where converted
@@ -6477,8 +6538,14 @@ static void llm_load_vocab(
         vocab.linefeed_id = ids[0];
     } else {
         const std::vector<int> ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
-        GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+
-        vocab.linefeed_id = ids[0];
+        //GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+        if (ids.empty()) {
+            LLAMA_LOG_WARN("%s: model vocab missing newline token, using special_pad_id instead\n", __func__);
+            vocab.linefeed_id = vocab.special_pad_id;
+        } else {
+            vocab.linefeed_id = ids[0];
+        }
     }
 
     // special tokens
@@ -7351,6 +7418,12 @@ static bool llm_load_tensors(
 
                     if (model.arch == LLM_ARCH_BERT) {
                         model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train});
+
+                        model.cls   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        model.cls_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        model.cls_out   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        model.cls_out_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS_OUT, "bias"),   {1},         llama_model_loader::TENSOR_NOT_REQUIRED);
                     }
 
                     model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});
@@ -7403,6 +7476,8 @@ static bool llm_load_tensors(
                     model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); // LayerNorm
                     model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}); //LayerNorm bias
 
+                    model.cls   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "weight"), {n_embd, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    model.cls_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "bias"),   {1},         llama_model_loader::TENSOR_NOT_REQUIRED);
                     for (int i = 0; i < n_layer; ++i) {
                         ggml_context * ctx_layer = ctx_for_layer(i);
                         ggml_context * ctx_split = ctx_for_layer_split(i);
@@ -8728,6 +8803,45 @@ static bool llm_load_tensors(
                     }
 
                 } break;
+            case LLM_ARCH_CHAMELEON:
+                {
+                 model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                 // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head});
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv});
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i),  {n_embd_head_k, n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i),  {n_embd_head_k, n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -10197,6 +10311,10 @@ struct llm_build_context {
         struct ggml_tensor * cur;
 
         switch (pooling_type) {
+            case LLAMA_POOLING_TYPE_NONE:
+                {
+                    cur = inp;
+                } break;
             case LLAMA_POOLING_TYPE_MEAN:
                 {
                     struct ggml_tensor * inp_mean = build_inp_mean();
@@ -10208,9 +10326,26 @@ struct llm_build_context {
                     struct ggml_tensor * inp_cls = build_inp_cls();
                     cur = ggml_get_rows(ctx0, inp, inp_cls);
                 } break;
-            case LLAMA_POOLING_TYPE_NONE:
+            case LLAMA_POOLING_TYPE_RANK:
                 {
-                    cur = inp;
+                    struct ggml_tensor * inp_cls = build_inp_cls();
+                    inp = ggml_get_rows(ctx0, inp, inp_cls);
+
+                    // classification head
+                    // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
+                    GGML_ASSERT(model.cls       != nullptr);
+                    GGML_ASSERT(model.cls_b     != nullptr);
+
+                    cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls, inp), model.cls_b);
+                    cur = ggml_tanh(ctx0, cur);
+
+                    // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+                    // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
+                    if (model.cls_out) {
+                        GGML_ASSERT(model.cls_out_b != nullptr);
+
+                        cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls_out, cur), model.cls_out_b);
+                    }
                 } break;
             default:
                 {
@@ -11439,8 +11574,8 @@ struct llm_build_context {
             inpL = cur;
         }
 
-        // final output
         cur = inpL;
+
         cb(cur, "result_embd", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -15872,6 +16007,184 @@ struct llm_build_context {
 
         return gf;
     }
+
+    // ref: https://github.com/facebookresearch/chameleon
+    // based on the original build_llama() function, changes:
+    //   * qk-norm
+    //   * swin-norm
+    //   * removed bias
+    //   * removed MoE
+    struct ggml_cgraph * build_chameleon() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        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);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            if (hparams.swin_norm) {
+                cur = inpL;
+            } else {
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_norm", il);
+            }
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                                ggml_element_size(Qcur) * n_embd_head,
+                                ggml_element_size(Qcur) * n_embd_head * n_head,
+                                0);
+                    cb(Qcur, "Qcur", il);
+
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                                model.layers[il].attn_q_norm,
+                                model.layers[il].attn_q_norm_b,
+                                LLM_NORM, cb, il);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                                ggml_element_size(Kcur) * n_embd_head,
+                                ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                                0);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                               model.layers[il].attn_k_norm,
+                               model.layers[il].attn_k_norm_b,
+                               LLM_NORM, cb, il);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, nullptr,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+                if (hparams.swin_norm) {
+                    cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].attn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                }
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            if (!hparams.swin_norm) {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+            }
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            if (hparams.swin_norm) {
+                cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output_with_img_logits", -1);
+
+        // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+        // Needs to be removed once image outputs are supported.
+        int img_token_end_idx = 8196;
+        int img_token_start_idx = 4;
+        int num_img_tokens = img_token_end_idx - img_token_start_idx;
+        // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+        // which ensures that text token values are always at least larger than image token values
+        struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+        img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+        cb(img_logits, "img_logits", -1);
+        cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -16132,6 +16445,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_rwkv6();
             } break;
+        case LLM_ARCH_CHAMELEON:
+            {
+                result = llm.build_chameleon();
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -16418,7 +16735,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         }
     }
 
-    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
+    if (cparams.embeddings && (
+                cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
+                cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
         const int64_t n_tokens     = batch.n_tokens;
         const int64_t n_seq_tokens = batch.n_seq_tokens;
         const int64_t n_seqs       = batch.n_seqs;
@@ -16433,7 +16752,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             const llama_seq_id seq_id = batch.seq_id[s][0];
 
             // TODO: adapt limits to n_seqs when batch.equal_seqs is true
-            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
                 const llama_pos pos = batch.pos[s*n_seq_tokens + i];
@@ -16973,6 +17292,20 @@ static int llama_decode_internal(
                             ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                         }
                     } break;
+                case LLAMA_POOLING_TYPE_RANK:
+                    {
+                        // extract the rerank score - a single float per sequence
+                        auto & embd_seq_out = lctx.embd_seq;
+
+                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
+                            embd_seq_out[seq_id].resize(1);
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
+                        }
+                    } break;
                 case LLAMA_POOLING_TYPE_UNSPECIFIED:
                     {
                         GGML_ABORT("unknown pooling type");
@@ -17179,6 +17512,13 @@ static int llama_encode_internal(
                             ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                         }
                     } break;
+                case LLAMA_POOLING_TYPE_RANK:
+                    {
+                        // TODO: this likely should be the same logic as in llama_decoder_internal, but better to
+                        //       wait for an encoder model that requires this pooling type in order to test it
+                        //       https://github.com/ggerganov/llama.cpp/pull/9510
+                        GGML_ABORT("RANK pooling not implemented yet");
+                    }
                 case LLAMA_POOLING_TYPE_UNSPECIFIED:
                     {
                         GGML_ABORT("unknown pooling type");
@@ -19257,6 +19597,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
+        case LLM_ARCH_CHAMELEON:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

tests/test-backend-ops.cpp

@@ -32,63 +32,52 @@
 #include <stdlib.h>
 #include <string>
 #include <thread>
+#include <future>
 #include <vector>
 
 static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
-    // static RNG initialization (revisit if n_threads stops being constant)
+    size_t nels = ggml_nelements(tensor);
-    static const size_t n_threads = std::thread::hardware_concurrency();
+    std::vector<float> data(nels);
-    static std::vector<std::default_random_engine> generators = []() {
+    {
-        std::random_device rd;
+        // parallel initialization
-        std::vector<std::default_random_engine> vec;
+        static const size_t n_threads = std::thread::hardware_concurrency();
-        vec.reserve(n_threads);
+        // static RNG initialization (revisit if n_threads stops being constant)
-        //for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(1234 + i); } // fixed seed
+        static std::vector<std::default_random_engine> generators = []() {
-        for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(rd()); }
+            std::random_device rd;
-        return vec;
+            std::vector<std::default_random_engine> vec;
-    }();
+            vec.reserve(n_threads);
+            //for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(1234 + i); } // fixed seed
+            for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(rd()); }
+            return vec;
+        }();
 
-    size_t size = ggml_nelements(tensor);
+        auto init_thread = [&](size_t ith, size_t start, size_t end) {
-    std::vector<float> data(size);
+            std::uniform_real_distribution<float> distribution(min, max);
+            auto & gen = generators[ith];
+            for (size_t i = start; i < end; i++) {
+                data[i] = distribution(gen);
+            }
+        };
 
-    auto init_thread = [&](size_t ith, size_t start, size_t end) {
+        std::vector<std::future<void>> tasks;
-        std::uniform_real_distribution<float> distribution(min, max);
+        tasks.reserve(n_threads);
-        for (size_t i = start; i < end; i++) {
+        for (size_t i = 0; i < n_threads; i++) {
-            data[i] = distribution(generators[ith]);
+            size_t start =     i*nels/n_threads;
+            size_t end   = (i+1)*nels/n_threads;
+            tasks.push_back(std::async(std::launch::async, init_thread, i, start, end));
         }
-    };
+        for (auto & t : tasks) {
-
+            t.get();
-    std::vector<std::thread> threads;
-    threads.reserve(n_threads);
-    for (size_t i = 0; i < n_threads; i++) {
-        size_t start =     i*size/n_threads;
-        size_t end   = (i+1)*size/n_threads;
-        threads.emplace_back(init_thread, i, start, end);
-    }
-    for (auto & t : threads) {
-        t.join();
-    }
-
-#if 0
-    const char * val_str = getenv("GGML_TEST_EPS");
-    float val = 1e-9f;
-    if (val_str != nullptr) {
-        val = std::stof(val_str);
-        printf("GGML_TEST_EPS=%e\n", val);
-    }
-
-    // test quantization with very small values that may result in nan scales due to division by zero
-    if (ggml_is_quantized(tensor->type)) {
-        for (int i = 0; i < 256; i++) {
-            data[i] = val;
         }
     }
-#endif
 
     if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
-        ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
+        ggml_backend_tensor_set(tensor, data.data(), 0, nels * sizeof(float));
     } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
-        GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
+        GGML_ASSERT(nels % ggml_blck_size(tensor->type) == 0);
-        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
+
-        std::vector<float> imatrix(tensor->ne[0], 1.0f); // dummy importance matrix
+         // dummy importance matrix
+        std::vector<float> imatrix(tensor->ne[0], 1.0f);
         const float * im = imatrix.data();
         if (!ggml_quantize_requires_imatrix(tensor->type)) {
             // when the imatrix is optional, we want to test both quantization with and without imatrix
@@ -98,15 +87,31 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
             }
         }
 
-        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], im);
+        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, nels));
-        GGML_ASSERT(ggml_validate_row_data(tensor->type, dataq.data(), dataq.size()));
+        {
-        // TODO: other cases
+            // parallel quantization by block
-        //#pragma omp parallel for
+            size_t blck_size = ggml_blck_size(tensor->type);
-        //for (int i = 0; i < tensor->ne[1]; i++) {
+            size_t n_blocks = nels / blck_size;
-        //    ggml_quantize_chunk(tensor->type, data.data(), dataq.data(),
-        //        i * tensor->ne[0], 1, tensor->ne[0], im);
-        //}
 
+            auto quantize_thread = [&](size_t start, size_t end) {
+                ggml_quantize_chunk(tensor->type, data.data(), dataq.data(),
+                    start * blck_size, end - start, blck_size, im);
+            };
+
+            const size_t min_blocks_per_thread = 1;
+            const size_t n_threads = std::min<size_t>(std::thread::hardware_concurrency()/2,
+                                                      std::max<size_t>(1, n_blocks / min_blocks_per_thread));
+            std::vector<std::future<void>> tasks;
+            tasks.reserve(n_threads);
+            for (size_t i = 0; i < n_threads; i++) {
+                size_t start =     i*n_blocks/n_threads;
+                size_t end   = (i+1)*n_blocks/n_threads;
+                tasks.push_back(std::async(std::launch::async, quantize_thread, start, end));
+            }
+            for (auto & t : tasks) {
+                t.get();
+            }
+        }
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
     } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
         // This is going to create some weird integers though.
@@ -160,60 +165,6 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
     return tv;
 }
 
-/*
-static double cosine_similarity(const float * v1, const float * v2, size_t n) {
-    double dot = 0.0;
-    double mag1 = 0.0;
-    double mag2 = 0.0;
-
-    for (size_t i = 0; i < n; i++) {
-        if (std::isnan(v1[i]) || std::isnan(v2[i])) {
-            return -1.0f;
-        }
-        if (std::isinf(v1[i]) && std::isinf(v2[i])) {
-            continue;
-        }
-        dot  += v1[i]*v2[i];
-        mag1 += v1[i]*v1[i];
-        mag2 += v2[i]*v2[i];
-    }
-
-    return dot/sqrt(mag1*mag2);
-}
-
-static float distance(const float * v1, const float * v2, size_t n) {
-    double d = 0.0;
-
-    for (size_t i = 0; i < n; i++) {
-        if (std::isnan(v1[i]) || std::isnan(v2[i])) {
-            return INFINITY;
-        }
-        if (std::isinf(v1[i]) && std::isinf(v2[i])) {
-            continue;
-        }
-        d += (v1[i] - v2[i])*(v1[i] - v2[i]);
-    }
-
-    return sqrt(d);
-}
-
-static float vec_len(const float * v, size_t n) {
-    double d = 0.0;
-
-    for (size_t i = 0; i < n; i++) {
-        if (std::isnan(v[i])) {
-            return INFINITY;
-        }
-        if (std::isinf(v[i])) {
-            continue;
-        }
-        d += v[i]*v[i];
-    }
-
-    return sqrt(d);
-}
-*/
-
 // normalized mean squared error = mse(a, b) / mse(a, 0)
 static double nmse(const float * a, const float * b, size_t n) {
     double mse_a_b = 0.0;
@@ -264,7 +215,6 @@ static double mean_abs_asymm(const float * a, const float * b, const size_t n, c
 }
 
 // utils for printing the variables of the test cases
-#define VAR_TO_STR(x) (#x "=" + var_to_str(x))
 
 template<typename T>
 static std::string var_to_str(const T & x) {
@@ -297,10 +247,6 @@ static std::string var_to_str(const std::array<T, N> & x) {
     return s;
 }
 
-//static std::string var_to_str(ggml_unary_op unary_op) {
-//    return ggml_unary_op_name(unary_op);
-//}
-
 static std::string var_to_str(ggml_type type) {
     return ggml_type_name(type);
 }
@@ -313,6 +259,8 @@ static std::string var_to_str(ggml_op_pool pool) {
     }
 }
 
+#define VAR_TO_STR(x) (#x "=" + var_to_str(x))
+
 #define VARS_TO_STR1(a) VAR_TO_STR(a)
 #define VARS_TO_STR2(a, b) VAR_TO_STR(a) + "," + VAR_TO_STR(b)
 #define VARS_TO_STR3(a, b, c) VAR_TO_STR(a) + "," + VARS_TO_STR2(b, c)
@@ -370,13 +318,13 @@ struct test_case {
         return 1e-4;
     }
 
-    virtual float grad_eps(){
+    virtual float grad_eps() {
         return 1e-1f;
     }
 
     // If false, estimate gradient with 2 points, neglects 3rd order derivative and higher.
     // If true,  estimate gradient with 4 points, neglects 5th order derivative and higher.
-    virtual bool grad_precise(){
+    virtual bool grad_precise() {
         return false;
     }
 
@@ -409,6 +357,11 @@ struct test_case {
         return size;
     }
 
+    virtual uint64_t op_flops(ggml_tensor * t) {
+        GGML_UNUSED(t);
+        return 0;
+    }
+
     ggml_cgraph * gf = nullptr;
     ggml_cgraph * gb = nullptr;
 
@@ -651,12 +604,11 @@ struct test_case {
         }
 
         // align while also leaving some margin for variations in parameters
-        int align = 20;
+        int align = 8;
         int last = (len + align - 1) / align * align;
         if (last - len < 5) {
             last += align;
         }
-        last = std::max(last, 60);
         printf("%*s", last - len, "");
 
         // allocate
@@ -677,9 +629,25 @@ struct test_case {
         // warmup run
         ggml_backend_graph_compute(backend, gf);
 
+        // determine number of runs
+        int n_runs;
+        if (op_flops(out) > 0) {
+            // based on flops
+            const uint64_t GFLOP = 1000 * 1000 * 1000;
+            const uint64_t target_flops_cpu =   8ULL * GFLOP;
+            const uint64_t target_flops_gpu = 100ULL * GFLOP;
+            uint64_t target_flops = ggml_backend_is_cpu(backend) ? target_flops_cpu : target_flops_gpu;
+            n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_flops / op_flops(out)) + 1;
+        } else {
+            // based on memory size
+            const size_t GB = 1ULL << 30;
+            const size_t target_size_cpu =  8 * GB;
+            const size_t target_size_gpu = 32 * GB;
+            size_t target_size = ggml_backend_is_cpu(backend) ? target_size_cpu : target_size_gpu;
+            n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
+        }
+
         // duplicate the op
-        size_t target_size = ggml_backend_is_cpu(backend) ? 1ULL << 33 : 1ULL << 35; // 8 GB CPU, 32 GB GPU
-        int n_runs = std::min((size_t) ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
         for (int i = 1; i < n_runs; i++) {
             ggml_graph_add_node(gf, out);
         }
@@ -706,17 +674,47 @@ struct test_case {
         // run
         ggml_backend_synchronize(backend);
 
-        int64_t start_time = ggml_time_us();
+        int64_t total_time_us = 0;
-        ggml_backend_graph_compute(backend, gf);
+        int total_runs = 0;
-        ggml_backend_synchronize(backend);
+        do {
-        int64_t end_time = ggml_time_us();
+            int64_t start_time = ggml_time_us();
-        double time_us = end_time - start_time;
+            ggml_backend_graph_compute(backend, gf);
+            ggml_backend_synchronize(backend);
+            int64_t end_time = ggml_time_us();
 
-        printf("    %5d runs - %8.2f us/run - %8zu kB/run - \033[1;34m%7.2f GB/s\033[0m\n",
+            total_time_us += end_time - start_time;
-            n_runs,
+            total_runs += n_runs;
-            time_us / n_runs,
+        } while (total_time_us < 1000*1000); // run for at least 1 second
-            op_size(out) / 1024,
+
-            mem / (time_us/1e6) / 1024.0 / 1024.0 / 1024.0);
+        printf("    %8d runs - %8.2f us/run - ",
+            total_runs,
+            (double)total_time_us / total_runs);
+
+        if (op_flops(out) > 0) {
+            double flops_per_sec = (op_flops(out) * total_runs) / (total_time_us / 1e6);
+            auto format_flops = [](double flops) -> std::string {
+                char buf[256];
+                if (flops >= 1e12) {
+                    snprintf(buf, sizeof(buf), "%6.2f TFLOP", flops / 1e12);
+                } else if (flops >= 1e9) {
+                    snprintf(buf, sizeof(buf), "%6.2f GFLOP", flops / 1e9);
+                } else if (flops >= 1e6) {
+                    snprintf(buf, sizeof(buf), "%6.2f MFLOP", flops / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%6.2f KFLOP", flops / 1e3);
+                }
+                return buf;
+            };
+            printf("%s/run - \033[1;34m%sS\033[0m",
+                format_flops(op_flops(out)).c_str(),
+                format_flops(flops_per_sec).c_str());
+
+        } else {
+            printf("%8zu kB/run - \033[1;34m%7.2f GB/s\033[0m",
+                op_size(out) / 1024,
+                mem / (total_time_us / 1e6) / 1024.0 / 1024.0 / 1024.0);
+        }
+        printf("\n");
 
         ggml_backend_buffer_free(buf);
 
@@ -1591,13 +1589,9 @@ struct test_mul_mat : public test_case {
         return 5e-4;
     }
 
-    size_t op_size(ggml_tensor * t) override {
+    uint64_t op_flops(ggml_tensor * t) override {
-        size_t a = ggml_nbytes(t->src[0]) * n * nr[0] * nr[1];
-        size_t b = ggml_nbytes(t->src[1]) * m;
-        size_t c  = ggml_nbytes(t);
-        return a + b + c;
-
         GGML_UNUSED(t);
+        return 2 * m * n * k * bs[0] * nr[0] * bs[1] * nr[1];
     }
 
     test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
@@ -1641,13 +1635,9 @@ struct test_mul_mat_id : public test_case {
         return 5e-4;
     }
 
-    size_t op_size(ggml_tensor * t) override {
+    uint64_t op_flops(ggml_tensor * t) override {
-        size_t a = ggml_nbytes(t->src[2]) * n;
-        size_t b = ggml_nbytes(t->src[1]) * m;
-        size_t c  = ggml_nbytes(t);
-        return a + b + c;
-
         GGML_UNUSED(t);
+        return 2 * m * k * n * n_used;
     }
 
     test_mul_mat_id(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
@@ -3163,47 +3153,46 @@ struct test_falcon : public test_llm {
 // ###########################################
 // ## Section 3: GGML Op Test Instantiation ##
 // ###########################################
+static const ggml_type all_types[] = {
+    GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16,
+    GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
+    GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
+    GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
+    GGML_TYPE_Q6_K,
+    // GGML_TYPE_TQ1_0, GGML_TYPE_TQ2_0, // TODO: implement for all backends
+    GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
+    GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
+    GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
+};
 
+static const ggml_type base_types[] = {
+    GGML_TYPE_F32, GGML_TYPE_F16,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_K,
+    GGML_TYPE_IQ2_XXS
+};
 
-static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
+static const ggml_type other_types[] = {
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
+    GGML_TYPE_Q5_K,
+    GGML_TYPE_Q6_K,
+    // GGML_TYPE_TQ1_0, GGML_TYPE_TQ2_0, // TODO: implement for all backends
+    GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
+    GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
+    GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
+    GGML_TYPE_BF16,
+};
+
+// Test cases for evaluation: should try to cover edge cases while using small input sizes to keep the runtime low
+static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     std::vector<std::unique_ptr<test_case>> test_cases;
     std::default_random_engine rng(0);
 
-    const ggml_type all_types[] = {
-        GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16,
-        GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
-        GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
-        GGML_TYPE_Q8_0,
-        GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
-        GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
-        GGML_TYPE_Q6_K,
-        // GGML_TYPE_TQ1_0, GGML_TYPE_TQ2_0, // TODO: implement for all backends
-        GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
-        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
-        GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
-    };
-
-    const ggml_type base_types[] = {
-        GGML_TYPE_F32, GGML_TYPE_F16,
-        GGML_TYPE_Q4_0,
-        GGML_TYPE_Q4_K,
-        GGML_TYPE_IQ2_XXS
-    };
-
-    const ggml_type other_types[] = {
-        GGML_TYPE_Q4_1,
-        GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
-        GGML_TYPE_Q8_0,
-        GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
-        GGML_TYPE_Q5_K,
-        GGML_TYPE_Q6_K,
-        // GGML_TYPE_TQ1_0, GGML_TYPE_TQ2_0, // TODO: implement for all backends
-        GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
-        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
-        GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
-        GGML_TYPE_BF16,
-    };
-
     // unary ops
     for (int v : {0, 1}) {
         for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
@@ -3392,6 +3381,14 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
         }
     }
+    for (ggml_type type_a : other_types) {
+        for (ggml_type type_b : {GGML_TYPE_F32}) {
+            if (ggml_blck_size(type_a) != 256) {
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1,  1}, {1, 1}));
+            }
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1,  1}, {1, 1}));
+        }
+    }
 #else
     // m = a rows
     // n = b rows
@@ -3411,15 +3408,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     }
 #endif
 
-    for (ggml_type type_a : other_types) {
-        for (ggml_type type_b : {GGML_TYPE_F32}) {
-            if (ggml_blck_size(type_a) != 256) {
-                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1,  1}, {1, 1}));
-            }
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1,  1}, {1, 1}));
-        }
-    }
-
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,  128, { 8,  1}, {1, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,  128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,   64, { 8,  1}, {4, 1}));
@@ -3624,20 +3612,30 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_falcon(2));
 #endif
 
-    // run tests
+    return test_cases;
-    if (mode == MODE_GRAD) {
+}
-        size_t n_ok = 0;
+
-        for (auto & test : test_cases) {
+// Test cases for performance evaluation: should be representative of real-world use cases
-            if (test->eval_grad(backend, op_name)) {
+static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
-                n_ok++;
+    std::vector<std::unique_ptr<test_case>> test_cases;
+
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1,   1, 1, 1}));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 1}));
+
+    for (int bs : {1, 512}) {
+        for (ggml_type type_a : all_types) {
+            for (ggml_type type_b : {GGML_TYPE_F32}) {
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, bs, 14336, {1,  1}, {1, 1}));
             }
         }
-        printf("  %zu/%zu tests passed\n", n_ok, test_cases.size());
-
-        return n_ok == test_cases.size();
     }
 
+    return test_cases;
+}
+
+static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     if (mode == MODE_TEST) {
+        auto test_cases = make_test_cases_eval();
         ggml_backend_t backend_cpu = ggml_backend_cpu_init();
 
         size_t n_ok = 0;
@@ -3653,7 +3651,21 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         return n_ok == test_cases.size();
     }
 
+    if (mode == MODE_GRAD) {
+        auto test_cases = make_test_cases_eval();
+        size_t n_ok = 0;
+        for (auto & test : test_cases) {
+            if (test->eval_grad(backend, op_name)) {
+                n_ok++;
+            }
+        }
+        printf("  %zu/%zu tests passed\n", n_ok, test_cases.size());
+
+        return n_ok == test_cases.size();
+    }
+
     if (mode == MODE_PERF) {
+        auto test_cases = make_test_cases_perf();
         for (auto & test : test_cases) {
             test->eval_perf(backend, op_name);
         }
@@ -3667,9 +3679,9 @@ static void usage(char ** argv) {
     printf("Usage: %s [mode] [-o op] [-b backend]\n", argv[0]);
     printf("    valid modes:\n");
     printf("      - test (default, compare with CPU backend for correctness)\n");
-    printf("      - perf (performance evaluation)\n");
     printf("      - grad (compare gradients from backpropagation with method of finite differences)\n");
-    printf("    op names are as given by ggml_op_desc() (e.g. GGML_ADD)\n");
+    printf("      - perf (performance evaluation)\n");
+    printf("    op names for -o are as given by ggml_op_desc() (e.g. ADD, MUL_MAT, etc)\n");
 }
 
 int main(int argc, char ** argv) {
@@ -3728,6 +3740,11 @@ int main(int argc, char ** argv) {
             continue;
         }
 
+        if (ggml_backend_is_cpu(backend)) {
+            // TODO: better value for n_threads
+            ggml_backend_cpu_set_n_threads(backend, std::thread::hardware_concurrency() / 2);
+        }
+
         printf("  Backend name: %s\n", ggml_backend_name(backend));
 
         bool ok = test_backend(backend, mode, op_name_filter);

tests/test-tokenizer-0.cpp

@@ -7,6 +7,7 @@
 #include <map>
 #include <vector>
 #include <fstream>
+#include <thread>
 
 //static const std::map<std::string, std::vector<llama_token>> & k_tests() {
 //    static std::map<std::string, std::vector<llama_token>> _k_tests = {
@@ -194,45 +195,64 @@ int main(int argc, char **argv) {
 
     const bool add_special = false;
 
-    for (const auto & test_kv : k_tests) {
+    // multi-threaded tokenization
-        const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
+    const int nthread = std::thread::hardware_concurrency();
+    std::vector<std::thread> threads(nthread);
 
-        printf("\n");
+    for (int i = 0; i < nthread; i++) {
-        printf("src: '%s'\n", test_kv.first.c_str());
+        threads[i] = std::thread([&, i]() {
-        printf("res: '%s'\n", llama_detokenize(ctx, res).c_str());
+            for (const auto & test_kv : k_tests) {
-        printf("tok: ");
+                const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
-        for (const auto & tok : res) {
-            printf("%d ", tok);
-        }
-        printf("\n");
 
-        bool correct = res.size() == test_kv.second.size();
+                // here only print the result of the first thread
-        for (int i = 0; i < (int) res.size() && correct; ++i) {
+                // because the other threads are running the same tests
-            if (test_kv.second[i] != res[i]) {
+                if (i != 0) {
-                correct = false;
+                    continue;
+                }
+
+                printf("\n");
+                printf("src: '%s'\n", test_kv.first.c_str());
+                printf("res: '%s'\n", llama_detokenize(ctx, res).c_str());
+                printf("tok: ");
+                for (const auto & tok : res) {
+                    printf("%d ", tok);
+                }
+                printf("\n");
+
+                bool correct = res.size() == test_kv.second.size();
+                for (int i = 0; i < (int) res.size() && correct; ++i) {
+                    if (test_kv.second[i] != res[i]) {
+                        correct = false;
+                    }
+                }
+
+                if (!correct) {
+                    fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+                    fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                        llama_detokenize(ctx, res).c_str(),
+                        llama_detokenize(ctx, test_kv.second).c_str());
+                    fprintf(stderr, "%s : expected tokens: ", __func__);
+                    for (const auto & t : test_kv.second) {
+                        fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+                    fprintf(stderr, "%s : got tokens:      ", __func__);
+                    for (const auto & t : res) {
+                        fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+
+                    success = false;
+                }
             }
-        }
+        });
-
-        if (!correct) {
-            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
-                llama_detokenize(ctx, res).c_str(),
-                llama_detokenize(ctx, test_kv.second).c_str());
-            fprintf(stderr, "%s : expected tokens: ", __func__);
-            for (const auto & t : test_kv.second) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s : got tokens:      ", __func__);
-            for (const auto & t : res) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
-
-            success = false;
-        }
     }
 
+    for (int i = 0; i < nthread; i++) {
+        threads[i].join();
+    }
+
+    // single threaded tokenization
     if (!fname_text.empty()) {
         fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());