ggml: ggml-cpu: force-no-lto-for-cpu-feats (#19609)

When LTO enabled in build environments it forces all builds to have LTO
in place. But feature detection logic is fragile, and causing Illegal
instruction errors with lto. This disables LTO for the feature
detection code to prevent cross-module optimization from inlining
architecture-specific instructions into the score function. Without this,
LTO can cause SIGILL when loading backends on older CPUs (e.g., loading
power10 backend on power9 crashes before feature check runs).

.github/workflows/winget.yml

@@ -17,7 +17,7 @@ jobs:
 
       - name: Install komac
         run: |
-          cargo binstall komac@2.11.2 -y
+          cargo binstall komac@2.15.0 -y
 
       - name: Find latest release
         id: find_latest_release

common/CMakeLists.txt

@@ -5,7 +5,6 @@ find_package(Threads REQUIRED)
 llama_add_compile_flags()
 
 # Build info header
-#
 
 if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
@@ -110,29 +109,16 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-# TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
-set(LLAMA_COMMON_EXTRA_LIBS build_info)
-set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
+target_link_libraries(${TARGET} PRIVATE
+    build_info
+    cpp-httplib
+)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)
     set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
     set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
-
-    # Set the correct library file extension based on platform
-    if (WIN32)
-        set(LLGUIDANCE_LIB_NAME "llguidance.lib")
-        # Add Windows-specific libraries
-        set(LLGUIDANCE_PLATFORM_LIBS
-            ws2_32    # Windows Sockets API
-            userenv   # For GetUserProfileDirectoryW
-            ntdll     # For NT functions
-            bcrypt    # For BCryptGenRandom
-        )
-    else()
-        set(LLGUIDANCE_LIB_NAME "libllguidance.a")
-        set(LLGUIDANCE_PLATFORM_LIBS "")
-    endif()
+    set(LLGUIDANCE_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}llguidance${CMAKE_STATIC_LIBRARY_SUFFIX}")
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
@@ -154,8 +140,10 @@ if (LLAMA_LLGUIDANCE)
     add_dependencies(llguidance llguidance_ext)
 
     target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
-    # Add platform libraries to the main target
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
-endif ()
+    target_link_libraries(${TARGET} PRIVATE llguidance)
+    if (WIN32)
+        target_link_libraries(${TARGET} PRIVATE ws2_32 userenv ntdll bcrypt)
+    endif()
+endif()
 
-target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+target_link_libraries(${TARGET} PUBLIC llama Threads::Threads)

common/common.h

@@ -670,7 +670,7 @@ 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;
@@ -685,7 +685,7 @@ 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;
 }
@@ -870,11 +870,11 @@ 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() };
 }
 

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")

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"},
 ]
 
 

examples/model-conversion/scripts/causal/run-org-model.py

@@ -42,11 +42,15 @@ def load_model_and_tokenizer(model_path, device="auto"):
         config = config.text_config
         multimodal = True
 
-    print("Vocab size:       ", config.vocab_size)
-    print("Hidden size:      ", config.hidden_size)
-    print("Number of layers: ", config.num_hidden_layers)
-    print("BOS token id:     ", config.bos_token_id)
-    print("EOS token id:     ", config.eos_token_id)
+    def print_if_exists(label, obj, attr, default="N/A"):
+        val = getattr(obj, attr) if hasattr(obj, attr) else default
+        print(f"{label}", val)
+
+    print_if_exists("Vocab size:       ", config, "vocab_size")
+    print_if_exists("Hidden size:      ", config, "hidden_size")
+    print_if_exists("Number of layers: ", config, "num_hidden_layers")
+    print_if_exists("BOS token id:     ", config, "bos_token_id")
+    print_if_exists("EOS token id:     ", config, "eos_token_id")
 
     unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
     if unreleased_model_name:

ggml/include/ggml.h

@@ -752,6 +752,7 @@ 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);

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;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -9,6 +9,11 @@ function(ggml_add_cpu_backend_features cpu_name arch)
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
     set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    # Disable LTO for the feature detection code to prevent cross-module optimization
+    # from inlining architecture-specific instructions into the score function.
+    # Without this, LTO can cause SIGILL when loading backends on older CPUs
+    # (e.g., loading power10 backend on power9 crashes before feature check runs).
+    target_compile_options(${GGML_CPU_FEATS_NAME} PRIVATE -fno-lto)
     target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
 endfunction()
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2278,11 +2278,12 @@ 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;
                 }
@@ -2305,6 +2306,8 @@ 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);
@@ -2865,15 +2868,6 @@ 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];
 
@@ -2888,31 +2882,14 @@ 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
         }
 

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);

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));
 }

ggml/src/ggml.c

@@ -1496,6 +1496,10 @@ 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");

src/llama-vocab.cpp

@@ -308,6 +308,7 @@ 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]+",
@@ -422,6 +423,14 @@ 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
@@ -2005,10 +2014,14 @@ 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;
@@ -2039,6 +2052,10 @@ 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;

src/llama-vocab.h

@@ -55,6 +55,8 @@ 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;

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;

tools/perplexity/perplexity.cpp

@@ -347,7 +347,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
     int count = 0;
     double nll = 0.0;
 
-    LOG_INF("%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+    const int n_seq = std::max(1, n_batch / n_ctx);
+    LOG_INF("%s: computing over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
 
     for (int i = 0; i < n_chunk; ++i) {
         const int start =     i * params.ppl_stride;
@@ -1737,11 +1738,21 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
     }
 
     const int n_batch = params.n_batch;
-    const int num_batches = (n_ctx + n_batch - 1)/n_batch;
+    const int num_batches = (static_cast<int>(n_ctx) + n_batch - 1) / n_batch;
+    // Calculate n_seq based on the logits file's n_ctx, but cap it at what the context supports
+    const int n_seq_max = llama_n_seq_max(ctx);
+    int n_seq = std::max(1, n_batch / static_cast<int>(n_ctx));
+    if (n_seq > n_seq_max) {
+        LOG_WRN("%s: calculated n_seq=%d exceeds context's n_seq_max=%d, capping at %d\n",
+                __func__, n_seq, n_seq_max, n_seq_max);
+        n_seq = n_seq_max;
+    }
     const int nv = 2*((n_vocab + 1)/2) + 4;
     const bool add_bos = llama_vocab_get_add_bos(vocab);
     GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
 
+    llama_batch batch = llama_batch_init(std::min(n_batch, static_cast<int>(n_ctx)*n_seq), 0, 1);
+
     std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
     std::vector<float>    kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
     std::vector<float> p_diff_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
@@ -1750,6 +1761,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         logits.reserve(size_t(n_ctx) * n_vocab);
     }
 
+    LOG_INF("%s: computing over %d chunks, n_ctx=%u, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
+
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
     auto mean_and_uncertainty = [] (double sum, double sum2, size_t count) {
@@ -1774,107 +1787,122 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
     auto    kld_ptr =    kld_values.data();
     auto p_diff_ptr = p_diff_values.data();
 
-    for (int i = 0; i < n_chunk; ++i) {
+    const int first = n_ctx/2;
+
+    for (int i = 0; i < n_chunk; i += n_seq) {
         const int start =     i * n_ctx;
         const int end   = start + n_ctx;
 
-        const auto t_start = std::chrono::high_resolution_clock::now();
+        const int n_seq_batch = std::min(n_seq, n_chunk - i);
 
-        if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) {
-            LOG_ERR("%s: failed reading log-probs for chunk %d\n", __func__, i);
-            return;
-        }
+        const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
         llama_memory_clear(llama_get_memory(ctx), true);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
-
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
-            // save original token and restore it after eval
-            const auto token_org = tokens[batch_start];
-
-            // add BOS token for the first batch of each chunk
-            if (add_bos && j == 0) {
-                tokens[batch_start] = llama_vocab_bos(vocab);
-            }
+            int n_outputs = 0;
 
             common_batch_clear(batch);
-            for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            for (int seq = 0; seq < n_seq_batch; seq++) {
+                int seq_start = batch_start + seq*n_ctx;
+
+                // save original token and restore it after eval
+                const auto token_org = tokens[seq_start];
+
+                // add BOS token for the first batch of each chunk
+                if (add_bos && j == 0) {
+                    tokens[seq_start] = llama_vocab_bos(vocab);
+                }
+
+                for (int k = 0; k < batch_size; ++k) {
+                    const int pos = j*n_batch + k;
+                    const bool need_logits = pos >= first;
+                    common_batch_add(batch, tokens[seq_start + k], pos, { seq }, need_logits);
+                    n_outputs += need_logits;
+                }
+
+                // restore the original token in case it was set to BOS
+                tokens[seq_start] = token_org;
             }
 
             if (llama_decode(ctx, batch)) {
-                LOG_ERR("%s : failed to eval\n", __func__);
+                LOG_ERR("%s : failed to decode\n", __func__);
                 llama_batch_free(batch);
                 return;
             }
 
-            // restore the original token in case it was set to BOS
-            tokens[batch_start] = token_org;
-
-            if (num_batches > 1) {
+            if (num_batches > 1 && n_outputs > 0) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + size_t(batch_size) * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + size_t(n_outputs) * n_vocab);
             }
         }
 
-        llama_batch_free(batch);
-
-        const auto t_end = std::chrono::high_resolution_clock::now();
-
         if (i == 0) {
+            llama_synchronize(ctx);
+            const auto t_end = std::chrono::high_resolution_clock::now();
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
             LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
-            int total_seconds = (int)(t_total * n_chunk);
+            int total_seconds = (int)(t_total * n_chunk / n_seq);
             if (total_seconds >= 60*60) {
                 LOG("%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
             LOG("%.2f minutes\n", total_seconds / 60.0);
+            LOG("\n");
+            LOG("chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p\n");
         }
-        LOG("\n");
-        LOG("chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p\n");
 
-        const int first = n_ctx/2;
-        const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
-        process_logits(n_vocab, all_logits + size_t(first)*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
-                workers, log_probs_uint16, kld, kld_ptr, p_diff_ptr);
-        p_diff_ptr += n_ctx - 1 - first;
-        kld_ptr    += n_ctx - 1 - first;
+        // Read log probs for each sequence in the batch
+        for (int seq = 0; seq < n_seq_batch; seq++) {
+            if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) {
+                LOG_ERR("%s: failed reading log-probs for chunk %d\n", __func__, i + seq);
+                llama_batch_free(batch);
+                return;
+            }
 
-        LOG("%4d", i+1);
+            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx + first);
 
-        auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
-        const double ppl_val = exp(log_ppl.first);
-        const double ppl_unc = ppl_val * log_ppl.second; // ppl_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl.second ** 2 )
-        LOG("    %9.4lf ± %9.4lf", ppl_val, ppl_unc);
+            process_logits(n_vocab, all_logits, tokens.data() + start + seq*n_ctx + first, n_ctx - 1 - first,
+                    workers, log_probs_uint16, kld, kld_ptr, p_diff_ptr);
+            p_diff_ptr += n_ctx - 1 - first;
+            kld_ptr    += n_ctx - 1 - first;
 
-        auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count);
-        const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count);
-        const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first;
-        const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov);
-        LOG("    %10.5lf ± %10.5lf", log_ppl_ratio_val, log_ppl_ratio_unc);
+            LOG("%4d", i + seq + 1);
 
-        auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
-        LOG("    %10.5lf ± %10.5lf", kl_div.first, kl_div.second);
+            auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
+            const double ppl_val = exp(log_ppl.first);
+            const double ppl_unc = ppl_val * log_ppl.second;
+            LOG("    %9.4lf ± %9.4lf", ppl_val, ppl_unc);
 
-        auto p_diff_mse   = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count);
-        const double p_diff_rms_val = sqrt(p_diff_mse.first);
-        const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second;
-        LOG("    %6.3lf ± %6.3lf %%", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
+            auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count);
+            const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count);
+            const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first;
+            const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov);
+            LOG("    %10.5lf ± %10.5lf", log_ppl_ratio_val, log_ppl_ratio_unc);
 
-        double p_top_val = 1.*kld.n_same_top/kld.count;
-        double p_top_unc = sqrt(p_top_val*(1 - p_top_val)/(kld.count - 1));
-        LOG("    %6.3lf ± %6.3lf %%", 100.0*p_top_val, 100.0*p_top_unc);
+            auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
+            LOG("    %10.5lf ± %10.5lf", kl_div.first, kl_div.second);
 
-        LOG("\n");
+            auto p_diff_mse   = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count);
+            const double p_diff_rms_val = sqrt(p_diff_mse.first);
+            const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second;
+            LOG("    %6.3lf ± %6.3lf %%", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
+
+            double p_top_val = 1.*kld.n_same_top/kld.count;
+            double p_top_unc = sqrt(p_top_val*(1 - p_top_val)/(kld.count - 1));
+            LOG("    %6.3lf ± %6.3lf %%", 100.0*p_top_val, 100.0*p_top_unc);
+
+            LOG("\n");
+        }
 
         logits.clear();
     }
+
+    llama_batch_free(batch);
     LOG("\n");
 
     if (kld.count < 100) return; // we do not wish to do statistics on so few values
@@ -1996,7 +2024,7 @@ int main(int argc, char ** argv) {
 
     const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence;
 
-    if (ppl) {
+    if (ppl || params.kl_divergence) {
         const int32_t n_seq = std::max(1, params.n_batch / n_ctx);
         const int32_t n_kv = n_seq * n_ctx;
 
@@ -2006,12 +2034,8 @@ int main(int argc, char ** argv) {
         params.n_batch = std::min(params.n_batch, n_kv);
     } else {
         params.n_batch = std::min(params.n_batch, params.n_ctx);
-        if (params.kl_divergence) {
-            params.n_parallel = 1;
-        } else {
-            // ensure there's at least enough seq_ids for HellaSwag
-            params.n_parallel = std::max(4, params.n_parallel);
-        }
+        // ensure there's at least enough seq_ids for HellaSwag
+        params.n_parallel = std::max(4, params.n_parallel);
     }
 
     if (params.ppl_stride > 0) {

tools/server/CMakeLists.txt

@@ -59,8 +59,4 @@ target_include_directories(${TARGET} PRIVATE ../mtmd)
 target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE server-context PUBLIC common cpp-httplib ${CMAKE_THREAD_LIBS_INIT})
 
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-endif()
-
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

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)