Merge branch 'upstream' into concedo_experimental

# Conflicts: # .github/workflows/winget.yml # CMakeLists.txt # common/CMakeLists.txt # examples/model-conversion/scripts/causal/run-org-model.py # ggml/src/ggml-cpu/CMakeLists.txt # tools/perplexity/perplexity.cpp # tools/server/CMakeLists.txt
2026-07-10 01:18:32 +00:00 · 2026-02-17 19:41:28 +08:00 · 2026-02-17 19:41:28 +08:00 · b6bb9c914e
commit b6bb9c914e
parent faf322c83f ae2d3f28a8
13 changed files with 80 additions and 54 deletions
--- a/common/common.h
+++ b/common/common.h
@ -667,7 +667,7 @@ static std::vector<T> string_split(const std::string & str, char delim) {
 }

 template<>
-std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
+inline std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
 {
    std::vector<std::string> parts;
    size_t begin_pos = 0;
@ -682,7 +682,7 @@ std::vector<std::string> string_split<std::string>(const std::string & input, ch
    return parts;
 }

-static bool string_starts_with(const std::string & str,
+inline bool string_starts_with(const std::string & str,
                               const std::string & prefix) {  // While we wait for C++20's std::string::starts_with...
    return str.rfind(prefix, 0) == 0;
 }
@ -867,11 +867,11 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

 const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";

-static std::string llm_ffn_exps_block_regex(int idx) {
+inline std::string llm_ffn_exps_block_regex(int idx) {
    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
 }

-static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
    return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
 }

--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@ -1049,6 +1049,9 @@ class TextModel(ModelBase):
        if chkhsh == "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902":
            # ref: https://huggingface.co/zai-org/GLM-4.5-Air
            res = "glm4"
+        if chkhsh == "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267":
+            # ref: https://huggingface.co/zai-org/GLM-4.7-Flash
+            res = "glm4"
        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
            res = "minerva-7b"
@ -1082,9 +1085,6 @@ class TextModel(ModelBase):
        if chkhsh == "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df":
            # ref: https://huggingface.co/aari1995/German_Semantic_V3
            res = "jina-v2-de"
-        if chkhsh == "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267":
-            # ref: https://huggingface.co/zai-org/GLM-4.7-Flash
-            res = "glm4"
        if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
            # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
            res = "llama-bpe"
@ -1124,6 +1124,9 @@ class TextModel(ModelBase):
        if chkhsh == "9c2227e4dd922002fb81bde4fc02b0483ca4f12911410dee2255e4987644e3f8":
            # ref: https://huggingface.co/CohereForAI/c4ai-command-r-v01
            res = "command-r"
+        if chkhsh == "d772b220ace2baec124bed8cfafce0ead7d6c38a4b65ef11261cf9d5d62246d1":
+            # ref: https://huggingface.co/CohereLabs/tiny-aya-base
+            res = "tiny_aya"
        if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
            # ref: https://huggingface.co/Qwen/Qwen1.5-7B
            res = "qwen2"
@ -1265,6 +1268,9 @@ class TextModel(ModelBase):
        if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4":
            # ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct
            res = "qwen35"
+        if chkhsh == "b4b8ca1f9769494fbd956ebc4c249de6131fb277a4a3345a7a92c7dd7a55808d":
+            # ref: https://huggingface.co/jdopensource/JoyAI-LLM-Flash
+            res = "joyai-llm"

        if res is None:
            logger.warning("\n")
@ -7360,6 +7366,17 @@ class Cohere2Model(TextModel):
        self.gguf_writer.add_rope_dimension_count(int(rotary_pct * (hidden_size // num_attention_heads)))
        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Cohere2 runtime in llama.cpp expects no bias tensors;
+        # the actual weight only contains 0-value tensors as bias, we can skip them
+        if name.endswith(".bias"):
+            if torch.any(data_torch != 0):
+                raise ValueError(f"Bias tensor {name!r} is not zero.")
+            logger.debug(f"Skipping bias tensor {name!r} for Cohere2 conversion.")
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+

@ModelBase.register("OlmoForCausalLM")
@ModelBase.register("OLMoForCausalLM")
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@ -99,6 +99,7 @@ models = [
    {"name": "stablelm2",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
    {"name": "refact",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
    {"name": "command-r",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
+    {"name": "tiny_aya",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/tiny-aya-base", },
    {"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", },
@ -148,7 +149,8 @@ models = [
    {"name": "youtu",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
    {"name": "solar-open",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
    {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
-    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", }
+    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", },
+    {"name": "joyai-llm",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jdopensource/JoyAI-LLM-Flash", },
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@ -158,6 +160,7 @@ pre_computed_hashes = [
    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
    {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
@ -171,7 +174,6 @@ pre_computed_hashes = [
    {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
    # jina-v2-de variants
    {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
-    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
 ]


--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@ -770,6 +770,7 @@ extern "C" {
    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
    GGML_API bool ggml_is_empty     (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_view      (const struct ggml_tensor * tensor);
    GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
    GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
    GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@ -17,11 +17,6 @@
 //#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #define AT_PRINTF(...)

-
-static bool ggml_is_view(const struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
 // ops that return true for this function must not use restrict pointers for their backend implementations
 bool ggml_op_can_inplace(enum ggml_op op) {
    switch (op) {
@ -627,7 +622,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
    GGML_ASSERT(buffer_id >= 0);
    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);

-    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
+    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_impl_is_view(node)) {
        hn->allocated = true;
        assert(hn->addr.offset == 0);

@ -658,7 +653,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor

                struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
+                    if (ggml_impl_is_view(parent)) {
                        struct ggml_tensor * view_src = parent->view_src;
                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
@ -739,7 +734,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
        // itself is never used and should not be considered a dependency
-        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+        if (ggml_impl_is_view(node) && node->op != GGML_OP_NONE) {
            struct ggml_tensor * view_src = node->view_src;
            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
        }
@ -806,7 +801,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);

            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                if (ggml_is_view(parent)) {
+                if (ggml_impl_is_view(parent)) {
                    struct ggml_tensor * view_src = parent->view_src;
                    struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
                    view_src_hn->n_views -= 1;
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@ -2286,11 +2286,12 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *

    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;

+    // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
        static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
        if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
            if (ggml_is_quantized(src0->type)) {
-                if (ne2 <= 4) {
+                if (ne2 <= MMVQ_MMID_MAX_BATCH_SIZE) {
                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
                    return;
                }
@ -2313,6 +2314,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
        }
    }

+    // note: this path should not be reached when recording CUDA graphs, because it requires stream synchronization
+    // TODO: add asserts to verify this. should work with CUDA, HIP, etc.
    cudaStream_t stream = ctx.stream();

    GGML_ASSERT(nb12 % nb11 == 0);
@ -2877,15 +2880,6 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
    bool use_cuda_graph = true;
    // Loop over nodes in GGML graph to obtain info needed for CUDA graph

-    const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
-    const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
-    const std::string ffn_moe_gate_bias_prefix = "ffn_moe_gate_biased";
-    const std::string ffn_moe_up_bias_prefix = "ffn_moe_up_biased";
-    const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
-    const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out";
-    const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d";
-    const std::string delta_net_prefix = "dnet_add";
-
    for (int i = 0; i < cgraph->n_nodes; i++) {
        ggml_tensor * node = cgraph->nodes[i];

@ -2900,31 +2894,14 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 #endif
        }

-        if (node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
-            use_cuda_graph = false; // This node type is not supported by CUDA graph capture
-#ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
-#endif
-        }
-
-        if (node->op == GGML_OP_ADD &&
-            node->src[1] && node->src[1]->ne[1] > 1 &&
-            (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) &&
-            (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true) &&
-            strncmp(node->name, ffn_moe_gate_bias_prefix.c_str(), ffn_moe_gate_bias_prefix.size()) != 0 &&
-            strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
-            strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 &&
-            strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 &&
-            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0 &&
-            strncmp(node->name, delta_net_prefix.c_str(), delta_net_prefix.size()) != 0) {
-            // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
-            // by means of matching node names. See
-            // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
-            // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773,
-            // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.
+        // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
+        if (node->op == GGML_OP_MUL_MAT_ID && (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > MMVQ_MMID_MAX_BATCH_SIZE)) {
+            // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
+            // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+            // ref: https://github.com/ggml-org/llama.cpp/pull/18958
            use_cuda_graph = false;
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
        }

--- a/ggml/src/ggml-cuda/mmvq.cuh
+++ b/ggml/src/ggml-cuda/mmvq.cuh
@ -1,6 +1,7 @@
 #include "common.cuh"

 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
+#define MMVQ_MMID_MAX_BATCH_SIZE 4 // Max. batch size for which to use MMVQ kernels for MUL_MAT_ID

 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@ -98,6 +98,10 @@ static bool ggml_op_is_empty(enum ggml_op op) {
    }
 }

+static inline bool ggml_impl_is_view(const struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 static inline float ggml_compute_softplus_f32(float input) {
    return (input > 20.0f) ? input : logf(1 + expf(input));
 }
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@ -1512,6 +1512,10 @@ bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tenso
        (t0->nb[3] == t1->nb[3]);
 }

+bool ggml_is_view(const struct ggml_tensor * t) {
+    return ggml_impl_is_view(t);
+}
+
 // check if t1 can be represented as a repetition of t0
 bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@ -533,6 +533,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                break;
            case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
            case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
+            case LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM:
                regex_exprs = {
                    "\\p{N}{1,3}",
                    "[一-龥぀-ゟ゠-ヿ]+",
@ -647,6 +648,14 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                };
                break;
+            case LLAMA_VOCAB_PRE_TYPE_TINY_AYA:
+                regex_exprs = {
+                    // original regex from tokenizer.json: "\\d{1,3}(?=(?:\\d{3})*\\b)"
+                    "\\d{1,3}(?=(?:\\d{3})*\\b)",
+                    // original regex from tokenizer.json: "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
+                    "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
            case LLAMA_VOCAB_PRE_TYPE_KIMI_K2:
                regex_exprs = {
                    // K2 trigger pattern - this will activate the custom K2 handler in unicode.cpp
@ -2241,10 +2250,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                tokenizer_pre == "megrez") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
            } else if (
-                    tokenizer_pre == "gpt-4o" ||
-                    tokenizer_pre == "llama4") {
+                tokenizer_pre == "gpt-4o" ||
+                tokenizer_pre == "llama4") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "tiny_aya") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
+                clean_spaces = false;
            } else if (
                tokenizer_pre == "superbpe") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE;
@ -2275,6 +2288,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                tokenizer_pre == "hunyuan-dense") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "joyai-llm") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM;
+                clean_spaces = false;
            } else if (
                tokenizer_pre == "kimi-k2") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@ -56,6 +56,8 @@ enum llama_vocab_pre_type {
    LLAMA_VOCAB_PRE_TYPE_YOUTU           = 44,
    LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE      = 45,
    LLAMA_VOCAB_PRE_TYPE_QWEN35          = 46,
+    LLAMA_VOCAB_PRE_TYPE_TINY_AYA        = 47,
+    LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM       = 48,
 };

 struct LLM_KV;
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@ -769,6 +769,12 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
    } else if (regex_expr == "\\p{AFMoE_digits}") {
        // AFMOE digit pattern - use custom implementation for proper splitting
        bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
+    } else if (regex_expr == "\\d{1,3}(?=(?:\\d{3})*\\b)") {
+        // tiny_aya digit grouping pattern from tokenizer.json:
+        //   {"type": "Split", "pattern": {"Regex": "\\d{1,3}(?=(?:\\d{3})*\\b)"}, "behavior": "Isolated"}
+        // Splits digits into groups of 3 from the right (e.g., 1234567 -> 1, 234, 567)
+        // TODO: Revisit this regex, incase there are any subtle tokenization differences with the original regex.
+        bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
    }

    return bpe_offsets;
--- a/vendor/cpp-httplib/CMakeLists.txt
+++ b/vendor/cpp-httplib/CMakeLists.txt
@ -17,7 +17,7 @@ endif()
 target_link_libraries(${TARGET} PRIVATE Threads::Threads)

 if (WIN32 AND NOT MSVC)
-    target_link_libraries(${TARGET} PRIVATE ws2_32)
+    target_link_libraries(${TARGET} PUBLIC ws2_32)
 endif()

 target_compile_features(${TARGET} PRIVATE cxx_std_17)