convert : add JoyAI-LLM-Flash (#19651)

* convert_hf_to_gguf: add JoyAI-LLM-Flash tokenizer hash mapping to deepseek-v3

* llama-vocab: create a new pre-tokenizer name for joyai-llm.

* add missing vocab type section

* Update convert_hf_to_gguf_update.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

CMakeLists.txt

@@ -112,15 +112,9 @@ option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_
 option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)
 
 # 3rd party libs
-option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
-# deprecated
-option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
-if (LLAMA_CURL)
-    message(WARNING "LLAMA_CURL option is deprecated and will be ignored")
-endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
@@ -148,10 +142,15 @@ if (NOT DEFINED GGML_CUDA_GRAPHS)
 endif()
 
 # transition helpers
-function (llama_option_depr TYPE OLD NEW)
+function (llama_option_depr TYPE OLD)
     if (${OLD})
-        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
-        set(${NEW} ON PARENT_SCOPE)
+        set(NEW "${ARGV2}")
+        if(NEW)
+            message(${TYPE} "${OLD} is deprecated, use ${NEW} instead")
+            set(${NEW} ON PARENT_SCOPE)
+        else()
+            message(${TYPE} "${OLD} is deprecated and will be ignored")
+        endif()
     endif()
 endfunction()
 
@@ -164,6 +163,7 @@ llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
 llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
+llama_option_depr(WARNING     LLAMA_CURL)
 
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")
@@ -197,9 +197,7 @@ add_subdirectory(src)
 
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
-    if (LLAMA_HTTPLIB)
-        add_subdirectory(vendor/cpp-httplib)
-    endif()
+    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)

build-xcframework.sh

@@ -449,10 +449,9 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
@@ -465,10 +464,9 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet

common/CMakeLists.txt

@@ -112,11 +112,7 @@ endif()
 
 # TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
 set(LLAMA_COMMON_EXTRA_LIBS build_info)
-
-if (LLAMA_HTTPLIB)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
-endif()
+set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)

common/common.cpp

@@ -879,7 +879,8 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
+        defined(__OpenBSD__) || defined(__NetBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else if (std::getenv("HOME")) {

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() };
 }
 

common/download.cpp

@@ -19,9 +19,7 @@
 #include <thread>
 #include <vector>
 
-#if defined(LLAMA_USE_HTTPLIB)
 #include "http.h"
-#endif
 
 #ifndef __EMSCRIPTEN__
 #ifdef __linux__
@@ -142,8 +140,6 @@ std::pair<std::string, std::string> common_download_split_repo_tag(const std::st
     return {hf_repo, tag};
 }
 
-#if defined(LLAMA_USE_HTTPLIB)
-
 class ProgressBar {
     static inline std::mutex mutex;
     static inline std::map<const ProgressBar *, int> lines;
@@ -768,30 +764,6 @@ std::string common_docker_resolve_model(const std::string & docker) {
     }
 }
 
-#else
-
-common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-bool common_download_model(const common_params_model &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-std::string common_docker_resolve_model(const std::string &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-int common_download_file_single(const std::string &,
-                                const std::string &,
-                                const std::string &,
-                                bool,
-                                const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-#endif // defined(LLAMA_USE_HTTPLIB)
-
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();

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")
@@ -2726,8 +2732,6 @@ class AfmoeModel(LlamaModel):
         super().set_gguf_parameters()
 
         # MoE parameters
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (n_shared_experts := self.hparams.get("num_shared_experts")) is not None:
             self.gguf_writer.add_expert_shared_count(n_shared_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
@@ -2749,7 +2753,7 @@ class AfmoeModel(LlamaModel):
         # Handle expert weights - they're already merged in the HF format
         # process the experts separately
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -4074,6 +4078,87 @@ class InternVisionModel(MmprojModel):
                 yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register(
+    "NemotronH_Nano_VL_V2",
+    "RADIOModel",
+)
+class NemotronNanoV2VLModel(MmprojModel):
+    # ViT-Huge architecture parameters for RADIO v2.5-h
+    _vit_hidden_size = 1280
+    _vit_intermediate_size = 5120
+    _vit_num_layers = 32
+    _vit_num_heads = 16
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        # RADIO config doesn't have standard ViT parameters, so they need to be constructed manually
+        vision_config = self.global_config.get("vision_config")
+        if vision_config is None:
+            return None
+        # Add ViT-H parameters
+        vision_config = {
+            **vision_config,
+            "hidden_size": self._vit_hidden_size,
+            "intermediate_size": self._vit_intermediate_size,
+            "num_hidden_layers": self._vit_num_layers,
+            "num_attention_heads": self._vit_num_heads,
+            "image_size": self.global_config.get("force_image_size", 512),
+        }
+        return vision_config
+
+    def set_gguf_parameters(self):
+        if "image_mean" not in self.preprocessor_config:
+            self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
+        if "image_std" not in self.preprocessor_config:
+            self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225]
+
+        super().set_gguf_parameters()
+        hparams = self.global_config
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
+        self.gguf_writer.add_vision_use_gelu(True)
+        downsample_ratio = hparams.get("downsample_ratio", 0.5)
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ".position_embd." in new_name or "pos_embed" in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "input_conditioner" in name:
+            return
+
+        # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
+        if "patch_generator.pos_embed" in name:
+            if not name.endswith(".weight"):
+                name += ".weight"
+            # Downsample position embeddings for fixed 512x512 image size
+            import torch.nn.functional as F
+            n_embd = self.hparams["hidden_size"]
+            image_size = self.global_config.get("force_image_size", 512)
+            patch_size = self.hparams["patch_size"]
+            target_patches_per_side = image_size // patch_size  # 32
+            max_patches_per_side = int((data_torch.shape[1]) ** 0.5)  # 128
+            if target_patches_per_side != max_patches_per_side:
+                # Reshape to grid, interpolate, flatten back
+                data_torch = data_torch.reshape(1, max_patches_per_side, max_patches_per_side, n_embd)
+                data_torch = data_torch.permute(0, 3, 1, 2).float()  # [1, n_embd, 128, 128]
+                data_torch = F.interpolate(data_torch, size=(target_patches_per_side, target_patches_per_side),
+                                           mode='bilinear', align_corners=True)
+                data_torch = data_torch.permute(0, 2, 3, 1)  # [1, 32, 32, n_embd]
+                data_torch = data_torch.reshape(1, target_patches_per_side * target_patches_per_side, n_embd)
+
+        # Reshape linear patch embedding to conv2d format for ggml_conv_2d
+        # From [n_embd, patch_size*patch_size*3] to [n_embd, 3, patch_size, patch_size]
+        if "patch_generator.embedder" in name:
+            patch_size = self.hparams["patch_size"]
+            n_embd = self.hparams["hidden_size"]
+            data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)
+
+        if name.startswith("vision_model.radio_model.model.") or name.startswith("mlp1."):
+            yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -4116,8 +4201,6 @@ class Qwen2MoeModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
             logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
@@ -4162,7 +4245,7 @@ class Qwen2MoeModel(TextModel):
             return
 
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -4913,13 +4996,13 @@ class PhiMoeModel(Phi3MiniModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
-        self.gguf_writer.add_expert_count(self.hparams["num_local_experts"])
+        self.gguf_writer.add_expert_used_count(self.find_hparam(["num_experts_per_tok", "num_experts_per_token"]))
+        self.gguf_writer.add_expert_count(self.find_hparam(["num_local_experts", "num_experts"]))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("block_sparse_moe.experts") != -1:
-            n_experts = self.hparams["num_local_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -5331,7 +5414,7 @@ class KimiLinearModel(TextModel):
 
         # process the experts separately
         if name.find("block_sparse_moe.experts") != -1:
-            n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -5926,12 +6009,13 @@ class NomicBertModel(BertModel):
         if "mlp.experts.bias" in name:
             return # Explicitly return.
 
+        n_experts = self.find_hparam(["num_local_experts", "num_experts"])
         if "mlp.experts.mlp.w1" in name:
-            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.view(n_experts, self.hparams["n_inner"], self.hparams["n_embd"])
             name += ".weight"
 
         if "mlp.experts.mlp.w2" in name:
-            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.view(n_experts, self.hparams["n_inner"], self.hparams["n_embd"])
             data_torch = data_torch.transpose(1, 2)
             name += ".weight"
 
@@ -5941,7 +6025,6 @@ class NomicBertModel(BertModel):
         super().set_gguf_parameters()
         if self.is_moe:
             self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"])
-            self.gguf_writer.add_expert_count(self.hparams["num_experts"])
             self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
 
     def _is_tokenizer_xlmroberta(self) -> bool:
@@ -7055,6 +7138,8 @@ class Mamba2Model(TextModel):
         if hparams is None:
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 hparams = json.load(f)
+        if "llm_config" in hparams:
+            hparams["text_config"] = hparams["llm_config"]
         super().__init__(dir_model, *args, hparams=hparams, **kwargs)
         self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
         self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
@@ -7176,8 +7261,8 @@ class JambaModel(TextModel):
         self.gguf_writer.add_ssm_state_size(d_state)
         self.gguf_writer.add_ssm_time_step_rank(dt_rank)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
-        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
+        self.gguf_writer.add_expert_count(self.find_hparam(["num_local_experts", "num_experts"]))
+        self.gguf_writer.add_expert_used_count(self.find_hparam(["num_experts_per_tok", "num_experts_per_token"]))
         self.gguf_writer.add_file_type(self.ftype)
 
     _experts: list[dict[str, Tensor]] | None = None
@@ -7195,7 +7280,7 @@ class JambaModel(TextModel):
 
         # process the experts separately
         if ".feed_forward.experts." in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
 
             assert bid is not None
 
@@ -7281,6 +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")
@@ -7343,8 +7439,6 @@ class OlmoeModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_layer_norm_rms_eps(1e-5)
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -7352,7 +7446,7 @@ class OlmoeModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -7933,10 +8027,6 @@ class MiniMaxM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.MINIMAXM2
     _experts_cache: dict[int, dict[str, Tensor]] = {}
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.hparams["num_experts"] = self.hparams["num_local_experts"]
-
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
@@ -7949,7 +8039,7 @@ class MiniMaxM2Model(TextModel):
 
         # merge expert weights
         if 'experts' in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             expert_cache = self._experts_cache.setdefault(bid, {})
@@ -9154,7 +9244,6 @@ class ExaoneMoEModel(Exaone4Model):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
         moe_intermediate_size = self.hparams["moe_intermediate_size"]
         num_shared_experts = self.hparams["num_shared_experts"]
         self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
@@ -9195,7 +9284,7 @@ class ExaoneMoEModel(Exaone4Model):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
 
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9346,7 +9435,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         # case, the model architecture needs to be updated to a standard
         # "granite" or "granitemoe" model
         if not self._ssm_layers:
-            has_experts = self.find_hparam(["num_experts_per_tok"], optional=True)
+            has_experts = self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)
             new_arch = (
                 gguf.MODEL_ARCH.GRANITE_MOE
                 if has_experts else
@@ -9542,6 +9631,14 @@ class NemotronHModel(GraniteHybridModel):
             self.gguf_writer.add_add_bos_token(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision model and projector tensors for VLM models (handled by mmproj) (e.g., Nemotron Nano 12B v2 VL)
+        if name.startswith(("vision_model.", "mlp1.")):
+            return
+
+        # Strip language_model. prefix for VLM models (e.g., Nemotron Nano 12B v2 VL)
+        if name.startswith("language_model."):
+            name = name[len("language_model."):]
+
         if self.is_moe and bid is not None:
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -9636,7 +9733,6 @@ class BailingMoeModel(TextModel):
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_weights_scale(1.0)
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
@@ -9670,7 +9766,7 @@ class BailingMoeModel(TextModel):
             yield from super().modify_tensors(v,self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)
             return
         elif name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9741,7 +9837,6 @@ class BailingMoeV2Model(TextModel):
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_shared_feed_forward_length(hparams.get("moe_shared_expert_intermediate_size", hparams["moe_intermediate_size"] * hparams["num_shared_experts"]))
         self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
@@ -9752,7 +9847,7 @@ class BailingMoeV2Model(TextModel):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if "mlp.experts" in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9798,8 +9893,6 @@ class GroveMoeModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
             logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
@@ -9820,7 +9913,7 @@ class GroveMoeModel(TextModel):
 
         # process the experts separately
         if name.find("chunk_experts") != -1:
-            n_experts = self.hparams["num_experts"] // 2 # see add_experts_per_group
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"]) // 2 # see add_experts_per_group
             assert bid is not None
 
             if self._chunk_experts is None:
@@ -9847,7 +9940,7 @@ class GroveMoeModel(TextModel):
             else:
                 return
         elif name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10240,7 +10333,6 @@ class HunYuanMoEModel(TextModel):
         super().set_gguf_parameters()
         hparams = self.hparams
 
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_feed_forward_length(hparams["intermediate_size"])
 
         moe_intermediate_size = hparams["moe_intermediate_size"]
@@ -10283,7 +10375,7 @@ class HunYuanMoEModel(TextModel):
                 return
 
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10325,16 +10417,9 @@ class LLaDAMoEModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
-
         if (expert_intermediate_size := self.hparams.get("expert_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
 
-        # number of experts used per token (top-k)
-        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
-            self.gguf_writer.add_expert_used_count(n_experts_used)
-
         self.gguf_writer.add_mask_token_id(156895)
         self.gguf_writer.add_causal_attention(False)
         self.gguf_writer.add_diffusion_shift_logits(False)
@@ -10345,7 +10430,7 @@ class LLaDAMoEModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10682,7 +10767,6 @@ class LFM2MoeModel(TextModel):
 
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
         self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
         self.gguf_writer.add_leading_dense_block_count(self.hparams["num_dense_layers"])
         self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
@@ -10703,7 +10787,7 @@ class LFM2MoeModel(TextModel):
 
         # merge expert weights
         if 'experts' in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             expert_cache = self._experts_cache.setdefault(bid, {})
@@ -10813,9 +10897,9 @@ class SmallThinkerModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))) is not None:
+        if (n_experts := self.hparams.get("moe_num_primary_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
-        if (n_experts_used := self.hparams.get("num_experts_per_tok", self.hparams.get("moe_num_active_primary_experts"))) is not None:
+        if (n_experts_used := self.hparams.get("moe_num_active_primary_experts")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
         if (moe_intermediate_size := self.hparams.get("moe_ffn_hidden_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
@@ -10840,7 +10924,7 @@ class SmallThinkerModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))
+            n_experts = self.hparams.get("moe_num_primary_experts") or self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:

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"},
 ]
 
 

docs/build-s390x.md

@@ -242,10 +242,10 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 |------------|-------------|------|-------|
 | FP32       | ✅           | ✅    | ❓     |
 | FP16       | ✅           | ✅    | ❓     |
-| BF16       | 🚫           | ✅    | ❓     |
+| BF16       | ✅           | ✅    | ❓     |
 | Q4_0       | ✅           | ❓    | ❓     |
 | Q4_1       | ✅           | ❓    | ❓     |
-| MXFP4      | 🚫           | ❓    | ❓     |
+| MXFP4      | ✅           | ❓    | ❓     |
 | Q5_0       | ✅           | ❓    | ❓     |
 | Q5_1       | ✅           | ❓    | ❓     |
 | Q8_0       | ✅           | ❓    | ❓     |
@@ -272,4 +272,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Feb 15, 2026.

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 5)
+set(GGML_VERSION_PATCH 7)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)

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

@@ -569,27 +569,24 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             cmake_policy(SET CMP0135 NEW)
         endif()
 
+        # TODO: Use FetchContent_MakeAvailable with EXCLUDE_FROM_ALL after bumping minimum CMake version to 3.28+
+        # Using FetchContent_Populate instead to avoid EXCLUDE_FROM_ALL which requires CMake 3.28
         FetchContent_Declare(KleidiAI_Download
             URL ${KLEIDIAI_DOWNLOAD_URL}
             DOWNLOAD_EXTRACT_TIMESTAMP NEW
             URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
 
-        FetchContent_MakeAvailable(KleidiAI_Download)
         FetchContent_GetProperties(KleidiAI_Download
             SOURCE_DIR  KLEIDIAI_SRC
             POPULATED   KLEIDIAI_POPULATED)
 
         if (NOT KLEIDIAI_POPULATED)
-            message(FATAL_ERROR "KleidiAI source downloaded failed.")
+            FetchContent_Populate(KleidiAI_Download)
+            FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
         endif()
 
         add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
 
-        # Remove kleidiai target after fetching it
-        if (TARGET kleidiai)
-            set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
-        endif()
-
         list(APPEND GGML_CPU_SOURCES
             ggml-cpu/kleidiai/kleidiai.cpp
             ggml-cpu/kleidiai/kernels.cpp

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -3226,6 +3226,316 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (svcntb() * 8 == 256) {
+        constexpr int    q8_k_blocklen = 4;
+        const svuint8_t m4b_1          = svdup_n_u8(0x0f);
+        // 8 accumulators: 2 row pairs × 4 col pairs
+        svfloat32_t acc_f32_01, acc_f32_23, acc_f32_45, acc_f32_67;
+        uint32_t idx_arr[8] = { 0, 2, 4, 6,  1, 3, 5, 7 };
+        svbool_t pg = svptrue_pat_b32(SV_VL8);
+        svuint32_t idx = svld1(pg, idx_arr);
+
+        static const uint32_t idx_data[8] = {0, 4, 2, 6, 1, 5, 3, 7};
+        svuint32_t idx1 = svld1_u32(svptrue_b32(), idx_data);
+
+        for (int y = 0; y < nr / q8_k_blocklen; y++) {
+            const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+                acc_f32_01 = svdup_n_f32(0);
+                acc_f32_23 = svdup_n_f32(0);
+                acc_f32_45 = svdup_n_f32(0);
+                acc_f32_67 = svdup_n_f32(0);
+
+                for (int b = 0; b < nb; b++) {
+                    // bsums pairs belongs to the same q8_k subblock
+                    // 64 elemnts loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
+                    const int16x8_t bsums[4]{
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                    };
+
+                    int32_t bsums_arr32[4][8];
+
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        int16x8_t v16 = bsums[q8_row];
+
+                        // low 4
+                        int32x4_t v32_lo = vmovl_s16(vget_low_s16(v16));
+                        vst1q_s32(&bsums_arr32[q8_row][0], v32_lo);
+
+                        // high 4
+                        int32x4_t v32_hi = vmovl_s16(vget_high_s16(v16));
+                        vst1q_s32(&bsums_arr32[q8_row][4], v32_hi);
+                    }
+
+                    svint32_t sb_acc_0 = svdup_n_s32(0);
+                    svint32_t sb_acc_2 = svdup_n_s32(0);
+
+                    svint32_t acc_00 = svdup_n_s32(0);
+                    svint32_t acc_11 = svdup_n_s32(0);
+                    svint32_t acc_22 = svdup_n_s32(0);
+                    svint32_t acc_33 = svdup_n_s32(0);
+                    svint32_t acc_44 = svdup_n_s32(0);
+                    svint32_t acc_55 = svdup_n_s32(0);
+                    svint32_t acc_66 = svdup_n_s32(0);
+                    svint32_t acc_77 = svdup_n_s32(0);
+
+                    svint32_t bias_acc_00 = svdup_n_s32(0);
+                    svint32_t bias_acc_22 = svdup_n_s32(0);
+                    svint32_t bias_acc_44 = svdup_n_s32(0);
+                    svint32_t bias_acc_66 = svdup_n_s32(0);
+
+                    for (int sb = 0; sb < QK_K / 64; sb++) {
+                        // Need scales for the low and high nibbles
+                        // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                        svint32_t block_scale_0, block_scale_1, block_scale_2, block_scale_3;
+                        svint32_t q4sb_mins_0, q4sb_mins_1;
+                        {
+                            // 2-superblock I am working on
+                            const int offset = sb * 24 + 0 * 12;
+                            const uint8_t * scales_in = &q4_ptr[b].scales[offset];
+
+                            const int offset1 = sb * 24 + 12;
+                            const uint8_t * scales_in1 = &q4_ptr[b].scales[offset1];
+
+                            constexpr uint32_t kmask1 = 0x3f3f3f3f;
+                            constexpr uint32_t kmask2 = 0x0f0f0f0f;
+                            constexpr uint32_t kmask3 = 0x03030303;
+                            constexpr uint8_t  scales_size = 12;
+
+                            uint32_t sm[3];
+                            memcpy(sm, scales_in, scales_size);
+
+                            uint32_t sm1[3];
+                            memcpy(sm1, scales_in1, scales_size);
+
+                            const uint32_t mins_0_3 = sm[1] & kmask1;
+                            const uint32_t mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+
+                            const uint32_t mins_0_3_1 = sm1[1] & kmask1;
+                            const uint32_t mins_4_7_1 = ((sm1[2] >> 4) & kmask2) | (((sm1[1] >> 6) & kmask3) << 4);
+
+                            svuint32_t mins_u32_temp = svzip1_u32(svdup_n_u32(mins_0_3), svdup_n_u32(mins_4_7));
+                            svuint32_t mins_u32_temp_1 = svzip1_u32(svdup_n_u32(mins_0_3_1), svdup_n_u32(mins_4_7_1));
+
+                            /* reinterpret u32 → u8 */
+                            svuint8_t mins_u8 = svreinterpret_u8_u32(mins_u32_temp);
+                            svuint8_t mins_u8_1 = svreinterpret_u8_u32(mins_u32_temp_1);
+
+                            /* widen u8 → u16->u32 (lower half only) */
+                            svuint32_t mins_u16 = svunpklo_u32(svunpklo_u16(mins_u8));
+                            svuint32_t mins_u16_1 = svunpklo_u32(svunpklo_u16(mins_u8_1));
+
+                            q4sb_mins_0 = svreinterpret_s32_u32(mins_u16);
+                            q4sb_mins_1 = svreinterpret_s32_u32(mins_u16_1);
+
+                            uint32_t scales_u32_0 = sm[0] & kmask1;
+                            uint32_t scales_u32_1 = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+                            uint32_t scales_u32_2 = sm1[0] & kmask1;
+                            uint32_t scales_u32_3 = (sm1[2] & kmask2) | (((sm1[0] >> 6) & kmask3) << 4);
+
+                            svuint32_t S01 = svdup_n_u32(scales_u32_0);
+                            svuint32_t S23 = svdup_n_u32(scales_u32_1);
+                            svuint32_t R01 = svdup_n_u32(scales_u32_2);
+                            svuint32_t R23 = svdup_n_u32(scales_u32_3);
+
+                            svint8_t S01_b = svreinterpret_s8_u32(S01);
+                            svint8_t S23_b = svreinterpret_s8_u32(S23);
+                            svint8_t R01_b = svreinterpret_s8_u32(R01);
+                            svint8_t R23_b = svreinterpret_s8_u32(R23);
+
+                            svint32_t S01_d = svunpklo_s32(svunpklo_s16(svzip1_s8(S01_b, S01_b)));
+                            svint32_t R01_d = svunpklo_s32(svunpklo_s16(svzip1_s8(R01_b, R01_b)));
+                            svint32_t S23_d = svunpklo_s32(svunpklo_s16(svzip1_s8(S23_b, S23_b)));
+                            svint32_t R23_d = svunpklo_s32(svunpklo_s16(svzip1_s8(R23_b, R23_b)));
+
+                            block_scale_0 = svtbl_s32(svzip1_s32(S01_d, R01_d), idx);
+                            block_scale_1 = svtbl_s32(svzip2_s32(S01_d, R01_d), idx);
+                            block_scale_2 = svtbl_s32(svzip1_s32(S23_d, R23_d), idx);
+                            block_scale_3 = svtbl_s32(svzip2_s32(S23_d, R23_d), idx);
+                        }
+
+                        const int8_t * q8_base_1 = q8_ptr[b].qs + sb * 256;
+
+                        // Load 32-byte per row pair, 1 subblock each time
+                        // predicate for activating higher lanes for 16 int8 elements
+                        const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                        // predicate for activating lower lanes for  16 int8 elements
+                        const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                        svint8_t q8_qs_0 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 0), svld1_s8(pl16, q8_base_1 + 112));
+                        svint8_t q8_qs_2 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 32), svld1_s8(pl16, q8_base_1 + 144));
+                        svint8_t q8_qs_4 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 64), svld1_s8(pl16, q8_base_1 + 176));
+                        svint8_t q8_qs_6 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 96), svld1_s8(pl16, q8_base_1 + 208));
+
+                        svint8_t q8_qs_1 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 16), svld1_s8(pl16, q8_base_1 + 128));
+                        svint8_t q8_qs_3 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 48), svld1_s8(pl16, q8_base_1 + 160));
+                        svint8_t q8_qs_5 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 80), svld1_s8(pl16, q8_base_1 + 192));
+                        svint8_t q8_qs_7 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 112), svld1_s8(pl16, q8_base_1 + 224));
+
+                        // Q4s columns iterated in pairs (01, 23, 45, 67)
+                        for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+
+                            sb_acc_0 = svdup_n_s32(0);
+                            sb_acc_2 = svdup_n_s32(0);
+
+                            svuint8_t q4_qs_cp_00 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);
+                            svuint8_t q4_qs_cp_01 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);
+                            svuint8_t q4_qs_cp_02 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);
+                            svuint8_t q4_qs_cp_03 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);
+
+                            svint8_t q4_nibbles_00 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_00, m4b_1), 4));
+                            svint8_t q4_nibbles_01 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_01, m4b_1), 4));
+                            svint8_t q4_nibbles_02 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_02, m4b_1), 4));
+                            svint8_t q4_nibbles_03 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_03, m4b_1), 4));
+
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_00, q8_qs_0);
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_01, q8_qs_2);
+
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_02, q8_qs_4);
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_03, q8_qs_6);
+
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_00, q8_qs_1);
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_01, q8_qs_3);
+
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_02, q8_qs_5);
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_03, q8_qs_7);
+
+                            if(cp == 0) {
+                                acc_00 = svmla_s32_m(svptrue_b32(), acc_00, sb_acc_0, block_scale_0);
+                                acc_44 = svmla_s32_m(svptrue_b32(), acc_44, sb_acc_2, block_scale_0);
+                            }
+                            if(cp == 1) {
+                                acc_11 = svmla_s32_m(svptrue_b32(), acc_11, sb_acc_0, block_scale_1);
+                                acc_55 = svmla_s32_m(svptrue_b32(), acc_55, sb_acc_2, block_scale_1);
+                            }
+                            if(cp == 2) {
+                                acc_22 = svmla_s32_m(svptrue_b32(), acc_22, sb_acc_0, block_scale_2);
+                                acc_66 = svmla_s32_m(svptrue_b32(), acc_66, sb_acc_2, block_scale_2);
+                            }
+                            if(cp == 3) {
+                                acc_33 = svmla_s32_m(svptrue_b32(), acc_33, sb_acc_0, block_scale_3);
+                                acc_77 = svmla_s32_m(svptrue_b32(), acc_77, sb_acc_2, block_scale_3);
+                            }
+                        }
+
+                        bias_acc_00 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_00, svdup_n_s32(bsums_arr32[sb][0]), q4sb_mins_0);
+                        bias_acc_00 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_00, svdup_n_s32(bsums_arr32[sb][1]), q4sb_mins_1);
+
+                        bias_acc_22 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_22, svdup_n_s32(bsums_arr32[sb][2]), q4sb_mins_0);
+                        bias_acc_22 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_22, svdup_n_s32(bsums_arr32[sb][3]), q4sb_mins_1);
+
+                        bias_acc_44 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_44, svdup_n_s32(bsums_arr32[sb][4]), q4sb_mins_0);
+                        bias_acc_44 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_44, svdup_n_s32(bsums_arr32[sb][5]), q4sb_mins_1);
+
+                        bias_acc_66 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_66, svdup_n_s32(bsums_arr32[sb][6]), q4sb_mins_0);
+                        bias_acc_66 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_66, svdup_n_s32(bsums_arr32[sb][7]), q4sb_mins_1);
+                    }  // for sb
+
+
+                    acc_00 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_00, svext_s32(acc_00, acc_00, 4));
+                    acc_11 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_11, svext_s32(acc_11, acc_11, 4));
+                    acc_22 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_22, svext_s32(acc_22, acc_22, 4));
+                    acc_33 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_33, svext_s32(acc_33, acc_33, 4));
+                    acc_44 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_44, svext_s32(acc_44, acc_44, 4));
+                    acc_55 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_55, svext_s32(acc_55, acc_55, 4));
+                    acc_66 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_66, svext_s32(acc_66, acc_66, 4));
+                    acc_77 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_77, svext_s32(acc_77, acc_77, 4));
+
+                    svint32_t reorder_acc_01 = svtbl_s32( svzip1_s32( svtrn1_s32(acc_00, acc_11), svtrn1_s32(acc_22, acc_33)), idx1);
+                    svint32_t reorder_acc_23 = svtbl_s32( svzip1_s32( svtrn2_s32(acc_00, acc_11), svtrn2_s32(acc_22, acc_33)), idx1);
+
+                    svint32_t reorder_acc_45 = svtbl_s32( svzip1_s32( svtrn1_s32(acc_44, acc_55), svtrn1_s32(acc_66, acc_77)), idx1);
+                    svint32_t reorder_acc_67 = svtbl_s32( svzip1_s32( svtrn2_s32(acc_44, acc_55), svtrn2_s32(acc_66, acc_77)), idx1);
+
+                    // Broadcast q8 scalar
+                    svfloat32_t q8_d = svdup_f32(q8_ptr[b].d[0]);
+
+                    svfloat32_t q4_dmin_temp = svcvt_f32_f16_x(svptrue_b32(), svzip1_f16( svld1_f16(svptrue_pat_b16(SV_VL8), (const __fp16 *)q4_ptr[b].dmin), svdup_f16(0)));
+
+                    svfloat32_t q4_d_temp = svcvt_f32_f16_x(svptrue_b32(), svzip1_f16( svld1_f16(svptrue_pat_b16(SV_VL8), (const __fp16 *)q4_ptr[b].d), svdup_f16(0)));
+
+                    svfloat32_t scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    svfloat32_t dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_01 = svmls_f32_m(svptrue_b32(), acc_f32_01, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_00), dmins1);
+                    acc_f32_01 = svmla_f32_m(svptrue_b32(), acc_f32_01, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_01), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[1]);
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_23 = svmls_f32_m(svptrue_b32(), acc_f32_23, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_22), dmins1);
+                    acc_f32_23 = svmla_f32_m(svptrue_b32(), acc_f32_23, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_23), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[2]);
+
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_45 = svmls_f32_m(svptrue_b32(), acc_f32_45, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_44), dmins1);
+                    acc_f32_45 = svmla_f32_m(svptrue_b32(), acc_f32_45, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_45), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[3]);
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_67 = svmls_f32_m(svptrue_b32(), acc_f32_67, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_66), dmins1);
+                    acc_f32_67 = svmla_f32_m(svptrue_b32(), acc_f32_67, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_67), scale1);
+
+                }  // for b
+
+                // With the previous reorder, the tile is already in the correct memory layout.
+                // Predicate for exactly 4 lanes
+                svbool_t pg4 = svptrue_pat_b32(SV_VL4);
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    int row = y * q8_k_blocklen + i;
+                    for (int j = 0; j < 2; j++) {
+                        int col    = x * ncols_interleaved + j * 4;
+                        int offset = row * bs + col;
+
+                        if (i == 0 && j == 0) {
+                            // acc_f32_0 → lower half of acc_f32_01
+                            svst1_f32(pg4, s + offset, acc_f32_01);
+                        } else if (i == 0 && j == 1) {
+                            // acc_f32_1 → upper half of acc_f32_01
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_01, acc_f32_01, 4));
+                        } else if (i == 1 && j == 0) {
+                            // acc_f32_2
+                            svst1_f32(pg4, s + offset, acc_f32_23);
+                        } else if (i == 1 && j == 1) {
+                            // acc_f32_3
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_23, acc_f32_23, 4));
+                        } else if (i == 2 && j == 0) {
+                            // acc_f32_4
+                            svst1_f32(pg4, s + offset, acc_f32_45);
+                        } else if (i == 2 && j == 1) {
+                            // acc_f32_5
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_45, acc_f32_45, 4));
+                        } else if (i == 3 && j == 0) {
+                            // acc_f32_6
+                            svst1_f32(pg4, s + offset, acc_f32_67);
+                        } else if (i == 3 && j == 1) {
+                            // acc_f32_7
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_67, acc_f32_67, 4));
+                        }
+                    }
+                }
+            }  // for x
+        }  // for y
+        return;
+    }
+#endif  // SVE compile-time end
+
 #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     constexpr int    q8_k_blocklen = 4;
     const uint8x16_t m4b           = vdupq_n_u8(0x0f);

ggml/src/ggml-cpu/common.h

@@ -6,8 +6,8 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 
-#define GGML_FA_TILE_Q  32
-#define GGML_FA_TILE_KV 16
+#define GGML_FA_TILE_Q  64
+#define GGML_FA_TILE_KV 64
 
 #ifdef __cplusplus
 

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

@@ -2874,8 +2874,8 @@ struct ggml_cplan ggml_graph_plan(
                         const int64_t DV = node->src[2]->ne[0];
 
                         // Tiled flash attention scratch (tile sizes defined in common.h)
-                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
-                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + K_f32 + padding
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV + GGML_FA_TILE_KV*DK)*n_tasks;
 
                         // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
                         // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
@@ -2947,7 +2947,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         /*.use_ref    =*/ cplan->use_ref,
     };
 
-    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
@@ -2974,7 +2978,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
-    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     ggml_barrier(state->threadpool);
 

ggml/src/ggml-cpu/ops.cpp

@@ -3,6 +3,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "binary-ops.h"
+#include "simd-gemm.h"
 #include "ggml.h"
 #include "unary-ops.h"
 #include "vec.h"
@@ -8389,10 +8390,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     GGML_ASSERT(k->type == v->type);
     const ggml_type kv_type = k->type;
 
-    const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type);
-    const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float;
-    const ggml_vec_dot_t    kv_vec_dot    = kv_type_traits_cpu->vec_dot;
-    const size_t kv_type_size = ggml_type_size(kv_type);
 
     // broadcast factors
     const int64_t rk2 = neq2/nek2;
@@ -8424,8 +8421,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
     static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
 
-    GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ");
-
     int ir = ir0;
     while (ir < ir1) {
         // q indices for the start of this tile
@@ -8452,18 +8447,20 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         }
 
         // Per-thread scratch layout:
-        // Q_q:    Q_TILE_SZ * DK (converted Q tile in KV type)
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile — F32 for GEMM, KV type for scalar)
         // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
         // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
         // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
-        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion)
-        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32);
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile)
+        // K_f32:  KV_TILE_SZ * DK (F32 buffer for K tile — GEMM path)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + KV_TILE_SZ*DK + CACHE_LINE_SIZE_F32);
 
         void  * Q_q    = base;
         float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
         float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
         float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
-        float * V32    = VKQ32 + Q_TILE_SZ * DV;  // F32 buffer for V tile
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;
+        float * K_f32  = V32 + KV_TILE_SZ * DV;
 
         memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
         memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
@@ -8476,28 +8473,38 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         const int iv3 = iq3 / rv3;
         const int iv2 = iq2 / rv2;
 
-        for (int tq = 0; tq < tile_rows; tq++) {
-            const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
-            kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK);
-        }
-        // Zero-pad remaining rows
-        for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
-            memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size);
+        {
+            float * Q_f32 = (float *)Q_q;
+            for (int tq = 0; tq < tile_rows; tq++) {
+                const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+                memcpy(Q_f32 + tq * DK, pq, DK * sizeof(float));
+            }
+            for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+                memset(Q_f32 + tq * DK, 0, DK * sizeof(float));
+            }
         }
 
+        memset(K_f32, 0, DK * KV_TILE_SZ * sizeof(float));
+        memset(V32,   0, KV_TILE_SZ * DV * sizeof(float));
+
         for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, nek1 - ic);
 
             // skip the tile entirely if all the masks are -inf
             if (mask) {
                 bool can_skip = true;
                 for (int tq = 0; tq < tile_rows; tq++) {
                     const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    for (int tk = 0; tk < kv_tile; tk++) {
                         mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
                         if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
                             can_skip = false;
                         }
                     }
+                    // Pad remaining mask entries with -inf
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
 
                 if (can_skip) {
@@ -8505,13 +8512,32 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 }
             }
 
-            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                const void * q_row = (const char *)Q_q + tq * DK * kv_type_size;
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3);
-                    float s;
-                    kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1);
-                    KQ[tq * KV_TILE_SZ + tk] = s * scale;
+            // Pack K tile transposed: K_f32[dk][kv] so KV_TILE is contiguous (SIMD dim)
+            // Zero-pad the last tile so the GEMM always operates on KV_TILE_SZ columns
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * k_data = (const char *)k->data + (ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
+                if (kv_type == GGML_TYPE_F16) {
+                    const ggml_fp16_t * k_f16 = (const ggml_fp16_t *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = GGML_CPU_FP16_TO_FP32(k_f16[dk]);
+                    }
+                } else {
+                    const float * k_f32_src = (const float *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = k_f32_src[dk];
+                    }
+                }
+            }
+            memset(KQ, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+            simd_gemm(KQ, (const float *)Q_q, K_f32, Q_TILE_SZ, DK, KV_TILE_SZ);
+            ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, scale);
+
+            // Set padded KQ entries to -inf so softmax gives them zero weight
+            if (kv_tile < KV_TILE_SZ) {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        KQ[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
             }
 
@@ -8551,33 +8577,22 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
             }
 
-            // Convert V tile to F32 first (if F16), then do MAD
-            // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster.
-            // TODO: on ARM, native f16 should be faster
-            if (kv_type == GGML_TYPE_F16) {
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                    ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV);
-                }
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p);
-                    }
-                }
-            } else {
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                        ggml_vec_mad_f32(DV, vkq_row, v_row, p);
-                    }
+            // V accumulation: VKQ32 += softmax(KQ) * V
+            // Pack V tile to contiguous F32, zero-padded
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
+                if (kv_type == GGML_TYPE_F16) {
+                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                } else {
+                    memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
                 }
             }
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                if (skip[tq]) {
+                    memset(KQ + tq * KV_TILE_SZ, 0, KV_TILE_SZ * sizeof(float));
+                }
+            }
+            simd_gemm(VKQ32, KQ, V32, Q_TILE_SZ, KV_TILE_SZ, DV);
         }
 
         // sinks (apply only to valid rows in the tile)
@@ -8794,15 +8809,15 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
         const int64_t dr = (nr + nchunk - 1) / nchunk;
 
-        static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
         static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
-        const bool use_tiled = !use_ref &&
+        bool use_tiled = !use_ref &&
                                (q->type == GGML_TYPE_F32 &&
                                 kv_is_f32_or_f16 &&
                                 k->type == v->type &&
-                                nek1 % KV_TILE_SZ == 0 &&
                                 neq1 >= Q_TILE_SZ);
-
+#ifdef GGML_SIMD
+        use_tiled &= (DV % GGML_F32_EPR == 0);
+#endif
         int current_chunk = ith;
 
         while (current_chunk < nchunk) {

ggml/src/ggml-cpu/simd-gemm.h

@@ -0,0 +1,136 @@
+#pragma once
+
+// Computes C[M x N] += A[M x K] * B[K x N]
+
+#include "simd-mappings.h"
+
+// TODO: add support for sizeless vector types
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+
+// TODO: untested on avx512
+// These are in units of GGML_F32_EPR
+#if defined(__AVX512F__) || defined (__ARM_NEON__)
+    static constexpr int GEMM_RM = 4;
+    static constexpr int GEMM_RN = 4; // 16+4+1 = 25/32
+#elif defined(__AVX2__) || defined(__AVX__)
+    static constexpr int GEMM_RM = 6;
+    static constexpr int GEMM_RN = 2; // 12+2+1 = 15/16
+#else
+    static constexpr int GEMM_RM = 2;
+    static constexpr int GEMM_RN = 2;
+#endif
+
+template <int RM, int RN>
+static inline void simd_gemm_ukernel(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    GGML_F32_VEC acc[RM][RN];
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            acc[i][r] = GGML_F32_VEC_LOAD(C + i * N + r * KN);
+        }
+    }
+
+    for (int64_t kk = 0; kk < K; kk++) {
+        GGML_F32_VEC Bv[RN];
+        for (int r = 0; r < RN; r++) {
+            Bv[r] = GGML_F32_VEC_LOAD(B + kk * N + r * KN);
+        }
+        for (int64_t i = 0; i < RM; i++) {
+            GGML_F32_VEC p = GGML_F32_VEC_SET1(A[i * K + kk]);
+            for (int r = 0; r < RN; r++) {
+                acc[i][r] = GGML_F32_VEC_FMA(acc[i][r], Bv[r], p);
+            }
+        }
+    }
+
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            GGML_F32_VEC_STORE(C + i * N + r * KN, acc[i][r]);
+        }
+    }
+}
+
+// C[M x N] += A[M x K] * B[K x N]
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    int64_t ii = 0;
+    for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<GEMM_RM, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<GEMM_RM, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            for (int64_t i = 0; i < GEMM_RM; i++) {
+                float a = C[i * N + jj];
+                for (int64_t kk = 0; kk < K; kk++) {
+                    a += A[i + kk] * B[kk * N + jj];
+                }
+                C[i * N + jj] = a;
+            }
+        }
+
+        A += GEMM_RM * K;
+        C += GEMM_RM * N;
+    }
+
+    // Tail rows: one at a time
+    for (; ii < M; ii++) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<1, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<1, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            float a = C[jj];
+            for (int64_t kk = 0; kk < K; kk++) {
+                a += A[kk] * B[kk * N + jj];
+            }
+            C[jj] = a;
+        }
+
+        A += K;
+        C += N;
+    }
+}
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
+#else // scalar path
+
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    for (int64_t i = 0; i < M; i++) {
+        for (int64_t j = 0; j < N; j++) {
+            float sum = C[i * N + j];
+            for (int64_t kk = 0; kk < K; kk++) {
+                sum += A[i * K + kk] * B[kk * N + j];
+            }
+            C[i * N + j] = sum;
+        }
+    }
+}
+
+#endif // GGML_SIMD

ggml/src/ggml-cpu/simd-mappings.h

@@ -1160,6 +1160,14 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     float32x4_t tmp = x[0] + vec_reve(x[0]);        \
     res = tmp[0] + tmp[1];                          \
 }
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
+{                                                \
+    float32x4_t v = vec_add(vec_add(s0, s1),     \
+                            vec_add(s2, s3));    \
+    v = vec_add(v, vec_sld(v, v, 8));            \
+    v = vec_add(v, vec_sld(v, v, 4));            \
+    res += (ggml_float)vec_extract(v, 0);        \
+}
 
 #define GGML_F32_VEC        GGML_F32x4
 #define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
@@ -1209,6 +1217,24 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_F16_VEC_MUL            GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
 
+// BF16 s390x
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR  8
+
+#define GGML_BF16x8         __vector unsigned short
+#define GGML_BF16x8_ZERO    vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC      GGML_BF16x8
+#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
+#define GGML_BF16_TO_F32_LO(v) ((float32x4_t) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((float32x4_t) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_FMA_LO(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
 #elif defined(__riscv_v_intrinsic)
 
 // compatible with vlen >= 128

ggml/src/ggml-cpu/vec.cpp

@@ -236,8 +236,7 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
     vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
     sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
 
-#endif
-#if defined(__POWER9_VECTOR__)
+#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
     const int np = (n & ~(GGML_BF16_STEP - 1));
     if (np > 0) {
         GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -63,7 +63,7 @@ static __global__ void flash_attn_ext_f16(
     constexpr int frag_m = ncols == 8 ? 32 : 16;
     constexpr int frag_n = ncols == 8 ?  8 : 16;
     static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
-#if defined(GGML_USE_HIP)
+#if defined(GGML_USE_HIP) && HIP_VERSION >= 60500000
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::row_major> frag_a_K;
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_a_V;
     typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_b;
@@ -135,7 +135,7 @@ static __global__ void flash_attn_ext_f16(
     __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
     half2 * VKQ2 = (half2 *) VKQ;
 
-#if defined(GGML_USE_HIP)
+#if defined(GGML_USE_HIP) && HIP_VERSION >= 60500000
     const _Float16 * K_h_f16  = reinterpret_cast<const _Float16 *>(K_h);
     const _Float16 * V_h_f16  = reinterpret_cast<const _Float16 *>(V_h);
     _Float16       * KQ_f16   = reinterpret_cast<_Float16 *>(KQ);

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

@@ -2872,6 +2872,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
     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];
@@ -2902,7 +2903,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
             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, 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
@@ -4544,6 +4546,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
                 case GGML_UNARY_OP_TRUNC:
+                    // TODO: should become:
+                    //return ggml_is_contiguous_rows(op->src[0]);
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;

ggml/src/ggml-cuda/mmq.cuh

@@ -2715,14 +2715,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
 #pragma unroll
         for (int l = 0; l < QR2_XXS; ++l) {
-            const int * grid_pos = (const int *) (iq2xxs_grid + aux8[l]);
-            const int signs_packed = ksigns_iq2xs[(aux32 >> (7*l)) & 0x7F];
+            const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]];
+            const uint32_t signs = unpack_ksigns(aux32 >> (7 * l));
 
-            const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-            const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-            const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
@@ -2733,12 +2733,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
-        const int ls = aux32 >> 28;
+        const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
         const float d = bxi->d;
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
-        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = (ls*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #else
-        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4;
+        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 }
@@ -2776,11 +2776,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
     #pragma unroll
         for (int l = 0; l < QR2_XS; ++l) {
-            const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l] & 0x000001FF));
-            const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l] >> 9));
+            const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF];
+            const uint32_t signs = unpack_ksigns(q2[l] >> 9);
 
-            const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
@@ -2904,11 +2907,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #pragma unroll
         for (int l = 0; l < QR3_XXS; ++l) {
             const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
+            const uint32_t signs = unpack_ksigns(aux32 >> (7*l));
 
-            const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l)) & 0x7F));
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;

ggml/src/ggml-cuda/vecdotq.cuh

@@ -94,6 +94,15 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
 #endif
 }
 
+static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
+    // v is a 7 bit int, with the 8th sign being encodable as popcnt
+    // with xor we can "correct" the bit instead of having to mask
+    const uint32_t p = __popc(v) & 1;
+    const uint32_t s = v ^ p << 7;
+    // broadcast over uint to allow for 0x08040201 / 0x80402010 as selectors
+    return s * 0x01010101;
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
@@ -905,22 +914,22 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     int sumi = 0;
 #pragma unroll
     for (int k0 = 0; k0 < 8; k0 += 2) {
-        const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]);
-        const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F];
+        const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[k0/2]];
+        const uint32_t signs = unpack_ksigns(aux32 >> (7 * k0 / 2));
 
-        const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-        const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0);
         sumi = ggml_cuda_dp4a(grid0, u0, sumi);
 
-        const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-        const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1);
         sumi = ggml_cuda_dp4a(grid1, u1, sumi);
     }
 
-    const int ls = aux32 >> 28;
-    sumi = (ls*sumi + sumi/2)/4;
+    const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
+    sumi = sumi * ls / 8;           // (sumi * scale + sumi / 2) / 4
     const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
     return d * sumi;
 }
@@ -942,13 +951,15 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     int sumi1 = 0;
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
-        const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF));
-        const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l0/2] >> 9));
-
-        const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+        const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l0/2] & 0x1FF];
+        const uint32_t signs = unpack_ksigns(q2[l0/2] >> 9);
 
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         if (l0 < 4) {
@@ -1028,13 +1039,16 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
         const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]);
+        const uint32_t signs = unpack_ksigns(aux32 >> (7*l0/2));
 
-        const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F));
-
-        const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         sumi = ggml_cuda_dp4a(grid_l, u0, sumi);

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-metal/ggml-metal-common.cpp

@@ -273,6 +273,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             case GGML_OP_DIAG:
             case GGML_OP_MUL:
             case GGML_OP_ADD:
+            case GGML_OP_SUB:
             case GGML_OP_DIV:
             case GGML_OP_GLU:
             case GGML_OP_SCALE:

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");

gguf-py/gguf/constants.py

@@ -3830,6 +3830,7 @@ class VisionProjectorType:
     MUSIC_FLAMINGO = "musicflamingo" # audio
     GLM4V = "glm4v"
     YOUTUVL = "youtuvl"
+    NEMOTRON_V2_VL = "nemotron_v2_vl"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -1346,6 +1346,7 @@ class TensorNameMap:
             "model.vision_tower.embeddings.cls_token", # Intern-S1
             "vision_model.class_embedding", # llama 4
             "model.vision.patch_embedding.cls_embedding", # cogvlm
+            "vision_model.radio_model.model.patch_generator.cls_token.token", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_PATCH: (
@@ -1360,6 +1361,7 @@ class TensorNameMap:
             "vision_tower.patch_embed.proj", # kimi-vl
             "model.vision.patch_embedding.proj", # cogvlm
             "siglip2.vision_model.embeddings.patch_embedding",
+            "vision_model.radio_model.model.patch_generator.embedder", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_NORM: (
@@ -1376,12 +1378,14 @@ class TensorNameMap:
             "visual.pos_embed", # qwen3vl
             "model.vision.patch_embedding.position_embedding", # cogvlm
             "visual.embeddings.position_embedding", # glm4v
+            "vision_model.radio_model.model.patch_generator.pos_embed", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
             "visual.blocks.{bid}.attn.qkv", # qwen3vl
             "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm
-            "vision_tower.encoder.blocks.{bid}.wqkv" # Kimi-K2.5
+            "vision_tower.encoder.blocks.{bid}.wqkv", # Kimi-K2.5
+            "vision_model.radio_model.model.blocks.{bid}.attn.qkv", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q: (
@@ -1446,6 +1450,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm1",
+            "vision_model.radio_model.model.blocks.{bid}.norm1", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1462,6 +1467,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.wo", # kimi-vl
             "model.vision.transformer.layers.{bid}.attention.dense", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl
+            "vision_model.radio_model.model.blocks.{bid}.attn.proj", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
@@ -1477,6 +1483,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm2",
+            "vision_model.radio_model.model.blocks.{bid}.norm2", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1493,6 +1500,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.mlp.fc1",
+            "vision_model.radio_model.model.blocks.{bid}.mlp.fc1", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1515,6 +1523,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.mlp.fc2",
+            "vision_model.radio_model.model.blocks.{bid}.mlp.fc2", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_LAYER_SCALE_1: (

scripts/sync-ggml.last

@@ -1 +1 @@
-a8db410a252c8c8f2d120c6f2e7133ebe032f35d
+d6754f3d0e6d0acd21c12442353c9fd2f94188e7

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "f80864ca031932351abef49b74097c67f14719c6"
+HTTPLIB_VERSION = "d4180e923f846b44a3d30acd938438d6e64fc9f6"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

src/CMakeLists.txt

@@ -57,13 +57,14 @@ add_library(llama
             models/deci.cpp
             models/deepseek.cpp
             models/deepseek2.cpp
+            models/delta-net-base.cpp
             models/dots1.cpp
             models/dream.cpp
             models/ernie4-5-moe.cpp
             models/ernie4-5.cpp
+            models/exaone-moe.cpp
             models/exaone.cpp
             models/exaone4.cpp
-            models/exaone-moe.cpp
             models/falcon-h1.cpp
             models/falcon.cpp
             models/gemma-embedding.cpp
@@ -91,10 +92,12 @@ add_library(llama
             models/llama-iswa.cpp
             models/llama.cpp
             models/maincoder.cpp
+            models/mamba-base.cpp
             models/mamba.cpp
             models/mimo2-iswa.cpp
             models/minicpm3.cpp
             models/minimax-m2.cpp
+            models/mistral3.cpp
             models/modern-bert.cpp
             models/mpt.cpp
             models/nemotron-h.cpp
@@ -118,12 +121,12 @@ add_library(llama
             models/qwen2moe.cpp
             models/qwen2vl.cpp
             models/qwen3.cpp
-            models/qwen3vl.cpp
-            models/qwen3vl-moe.cpp
-            models/qwen3moe.cpp
-            models/qwen3next.cpp
             models/qwen35.cpp
             models/qwen35moe.cpp
+            models/qwen3moe.cpp
+            models/qwen3next.cpp
+            models/qwen3vl-moe.cpp
+            models/qwen3vl.cpp
             models/refact.cpp
             models/rnd1.cpp
             models/rwkv6-base.cpp
@@ -142,8 +145,6 @@ add_library(llama
             models/t5-enc.cpp
             models/wavtokenizer-dec.cpp
             models/xverse.cpp
-            models/mistral3.cpp
-            models/graph-context-mamba.cpp
             )
 
 set_target_properties(llama PROPERTIES

src/llama-adapter.h

@@ -39,6 +39,8 @@ private:
     std::vector<ggml_tensor *> tensors; // per layer
 };
 
+using llama_adapter_cvec_ptr = std::shared_ptr<llama_adapter_cvec>;
+
 //
 // llama_adapter_lora
 //
@@ -84,3 +86,4 @@ struct llama_adapter_lora {
 };
 
 using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;
+using llama_adapter_loras_ptr = std::unique_ptr<llama_adapter_loras>;

src/llama-context.cpp

@@ -22,6 +22,8 @@ llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
     model(model),
+    cvec(std::make_unique<llama_adapter_cvec>()),
+    loras(std::make_unique<llama_adapter_loras>()),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
     // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
     //     may need to be backend-dependent
@@ -878,6 +880,7 @@ const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) {
         }
     } catch (const std::exception & err) {
         // fallback to full vocab list
+        GGML_UNUSED(err);
     }
 
     return sampling.token_ids_full_vocab.data();
@@ -1064,11 +1067,11 @@ void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_a
         return;
     }
 
-    loras.clear();
+    loras.reset(new llama_adapter_loras());
 
     for (size_t i = 0; i < n_adapters; i ++) {
         if (scales[i] != 0.0f) {
-            loras[adapters[i]] = scales[i];
+            loras->insert({adapters[i], scales[i]});
         }
     }
 
@@ -1078,14 +1081,14 @@ void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_a
 bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) {
     LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
 
-    if (n_adapters != loras.size()) {
+    if (n_adapters != loras->size()) {
         return false;
     }
 
     for (size_t i = 0; i < n_adapters; i ++) {
-        auto it = loras.find(adapters[i]);
+        auto it = loras->find(adapters[i]);
 
-        if (it == loras.end() || it->second != scales[i]) {
+        if (it == loras->end() || it->second != scales[i]) {
             return false;
         }
     }
@@ -1103,7 +1106,7 @@ bool llama_context::set_adapter_cvec(
 
     // TODO: should we reserve?
 
-    return cvec.apply(model, data, len, n_embd, il_start, il_end);
+    return cvec->apply(model, data, len, n_embd, il_start, il_end);
 }
 
 llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) {
@@ -1809,7 +1812,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
 //
 
 uint32_t llama_context::output_reserve(int32_t n_outputs) {
-
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1893,11 +1895,6 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     embd = has_embd ? buffer_view<float>{(float *) (base + offset), embd.size} : buffer_view<float>{nullptr, 0};
     offset += embd.size * sizeof(float);
 
-    sampling.logits     = {nullptr, 0};
-    sampling.probs      = {nullptr, 0};
-    sampling.sampled    = {nullptr, 0};
-    sampling.candidates = {nullptr, 0};
-
     if (has_sampling) {
         sampling.logits = {(float *) (base + offset), (size_t)(n_vocab*n_outputs_max)};
         offset += sampling.logits.size * sizeof(float);
@@ -1923,6 +1920,15 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         std::fill(sampling.candidates_count.begin(), sampling.candidates_count.end(), 0);
 
         std::fill_n(sampling.sampled.data, sampling.sampled.size, LLAMA_TOKEN_NULL);
+    } else {
+        sampling.logits     = {nullptr, 0};
+        sampling.probs      = {nullptr, 0};
+        sampling.sampled    = {nullptr, 0};
+        sampling.candidates = {nullptr, 0};
+
+        sampling.logits_count.clear();
+        sampling.probs_count.clear();
+        sampling.candidates_count.clear();
     }
 
     // set all ids as invalid (negative)
@@ -1953,37 +1959,30 @@ void llama_context::output_reorder() {
             }
         }
 
-        if (sampling.logits.has_data()) {
+        if (!sampling.samplers.empty()) {
+            assert(sampling.logits.size > 0);
+            assert(sampling.probs.size > 0);
+            assert(sampling.candidates.size > 0);
+            assert(sampling.sampled.size > 0);
+            assert(sampling.logits_count.size() > 0);
+            assert(sampling.probs_count.size() > 0);
+            assert(sampling.candidates_count.size() > 0);
+
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.logits.data[i0*n_vocab + k], sampling.logits.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.probs.has_data()) {
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.probs.data[i0*n_vocab + k], sampling.probs.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.candidates.has_data()) {
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.candidates.data[i0*n_vocab + k], sampling.candidates.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.sampled.has_data()) {
-            std::swap(sampling.sampled.data[i0], sampling.sampled.data[i1]);
-        }
-
-        if (!sampling.logits_count.empty()) {
-            std::swap(sampling.logits_count[i0], sampling.logits_count[i1]);
-        }
-
-        if (!sampling.probs_count.empty()) {
-            std::swap(sampling.probs_count[i0], sampling.probs_count[i1]);
-        }
-
-        if (!sampling.candidates_count.empty()) {
+            std::swap(sampling.sampled.data[i0],     sampling.sampled.data[i1]);
+            std::swap(sampling.logits_count[i0],     sampling.logits_count[i1]);
+            std::swap(sampling.probs_count[i0],      sampling.probs_count[i1]);
             std::swap(sampling.candidates_count[i0], sampling.candidates_count[i1]);
         }
     }
@@ -2084,8 +2083,8 @@ llm_graph_params llama_context::graph_params(
         /*.gtype       =*/ gtype,
         /*.sched       =*/ sched.get(),
         /*.backend_cpu =*/ backend_cpu,
-        /*.cvec        =*/ &cvec,
-        /*.loras       =*/ &loras,
+        /*.cvec        =*/ cvec.get(),
+        /*.loras       =*/ loras.get(),
         /*.mctx        =*/ mctx,
         /*.cross       =*/ &cross,
         /*.samplers    =*/ sampling.samplers,

src/llama-context.h

@@ -256,33 +256,36 @@ private:
 
     const llama_model & model;
 
-    llama_cparams       cparams;
-    llama_adapter_cvec  cvec;
-    llama_adapter_loras loras;
+    llama_cparams cparams;
+
+    llama_adapter_cvec_ptr  cvec;
+    llama_adapter_loras_ptr loras;
 
     llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
 
     std::unique_ptr<llama_memory_i> memory;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
-    struct buffer_view<float>  logits = {nullptr, 0};
+    buffer_view<float> logits = {nullptr, 0};
 
     // embeddings output (2-dimensional array: [n_outputs][n_embd])
     // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
-    struct buffer_view<float>  embd = {nullptr, 0};
+    buffer_view<float> embd = {nullptr, 0};
 
     struct sampling_info {
+        // !samplers.empty() to check if any samplers are active
         std::map<llama_seq_id, llama_sampler *> samplers;
 
-        struct buffer_view<float>       logits     = {nullptr, 0};
-        struct buffer_view<llama_token> sampled    = {nullptr, 0};
-        struct buffer_view<float>       probs      = {nullptr, 0};
-        struct buffer_view<llama_token> candidates = {nullptr, 0};
+        buffer_view<float>       logits     = {nullptr, 0};
+        buffer_view<llama_token> sampled    = {nullptr, 0};
+        buffer_view<float>       probs      = {nullptr, 0};
+        buffer_view<llama_token> candidates = {nullptr, 0};
 
         std::vector<uint32_t> logits_count;
         std::vector<uint32_t> probs_count;
         std::vector<uint32_t> candidates_count;
 
+        // optimization
         std::vector<llama_token> token_ids_full_vocab;
     };
 

src/llama-graph.cpp

@@ -17,6 +17,41 @@
 #include <sstream>
 #include <unordered_set>
 
+// dedup helpers
+
+static ggml_tensor * build_kq_mask(
+        ggml_context * ctx,
+        const llama_kv_cache_context * mctx,
+        const llama_ubatch & ubatch,
+        const llama_cparams & cparams) {
+    const auto n_kv     = mctx->get_n_kv();
+    const auto n_tokens = ubatch.n_tokens;
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+    return ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+}
+
+static bool can_reuse_kq_mask(
+        ggml_tensor * kq_mask,
+        const llama_kv_cache_context * mctx,
+        const llama_ubatch & ubatch,
+        const llama_cparams & cparams) {
+    const auto n_kv     = mctx->get_n_kv();
+    const auto n_tokens = ubatch.n_tokens;
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+    bool res = true;
+
+    res &= (kq_mask->ne[0] == n_kv);
+    res &= (kq_mask->ne[1] == n_tokens/n_stream);
+    res &= (kq_mask->ne[2] == 1);
+    res &= (kq_mask->ne[3] == n_stream);
+
+    return res;
+}
+
+// impl
+
 void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     if (ubatch->token) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -403,8 +438,7 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
     res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
   //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
-    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
-    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    res &= can_reuse_kq_mask(self_kq_mask, mctx, params.ubatch, params.cparams);
 
     return res;
 }
@@ -424,8 +458,7 @@ bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
 
     res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
 
-    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
-    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    res &= can_reuse_kq_mask(self_kq_mask, mctx, params.ubatch, params.cparams);
 
     return res;
 }
@@ -455,11 +488,8 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
     res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
   //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
-    res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
-    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
-
-    res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
-    res &= self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
+    res &= can_reuse_kq_mask(self_kq_mask,     mctx->get_base(), params.ubatch, params.cparams);
+    res &= can_reuse_kq_mask(self_kq_mask_swa, mctx->get_swa(),  params.ubatch, params.cparams);
 
     return res;
 }
@@ -521,8 +551,7 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
     res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
   //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
-    res &= inp_attn->self_kq_mask->ne[0] == mctx->get_attn()->get_n_kv();
-    res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    res &= can_reuse_kq_mask(inp_attn->self_kq_mask, mctx->get_attn(), params.ubatch, params.cparams);
 
     res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
 
@@ -565,8 +594,7 @@ bool llm_graph_input_mem_hybrid_k::can_reuse(const llm_graph_params & params) {
 
     res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
 
-    res &= inp_attn->self_kq_mask->ne[0] == mctx->get_attn()->get_n_kv();
-    res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    res &= can_reuse_kq_mask(inp_attn->self_kq_mask, mctx->get_attn(), params.ubatch, params.cparams);
 
     res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
 
@@ -625,8 +653,7 @@ bool llm_graph_input_mem_hybrid_iswa::can_reuse(const llm_graph_params & params)
         res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
       //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
-        res &= inp_attn->self_kq_mask->ne[0] == attn_ctx->get_base()->get_n_kv();
-        res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+        res &= can_reuse_kq_mask(inp_attn->self_kq_mask, attn_ctx->get_base(), params.ubatch, params.cparams);
     }
 
     // swa tensors may not be allocated if there are no SWA attention layers
@@ -634,8 +661,7 @@ bool llm_graph_input_mem_hybrid_iswa::can_reuse(const llm_graph_params & params)
         res &= inp_attn->self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
       //res &= inp_attn->self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
-        res &= inp_attn->self_kq_mask_swa->ne[0] == attn_ctx->get_swa()->get_n_kv();
-        res &= inp_attn->self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
+        res &= can_reuse_kq_mask(inp_attn->self_kq_mask_swa, attn_ctx->get_swa(), params.ubatch, params.cparams);
     }
 
     res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
@@ -1891,14 +1917,11 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
 
-        const auto n_kv     = mctx_cur->get_n_kv();
-        const auto n_tokens = ubatch.n_tokens;
-        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
-
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp->self_kq_mask = build_kq_mask(ctx0, mctx_cur, ubatch, cparams);
+
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1983,13 +2006,9 @@ static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
 
-        const auto n_kv     = mctx_cur->get_n_kv();
-        const auto n_tokens = ubatch.n_tokens;
-        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
-
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp->self_kq_mask = build_kq_mask(ctx0, mctx_cur, ubatch, cparams);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -2188,15 +2207,11 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
 
     auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);
 
-    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
-
     {
-        const auto n_kv = mctx_cur->get_base()->get_n_kv();
-
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp->self_kq_mask = build_kq_mask(ctx0, mctx_cur->get_base(), ubatch, cparams);
         ggml_set_input(inp->self_kq_mask);
         ggml_set_name(inp->self_kq_mask, "self_kq_mask");
 
@@ -2207,12 +2222,10 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
     {
         GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA");
 
-        const auto n_kv = mctx_cur->get_swa()->get_n_kv();
-
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp->self_kq_mask_swa = build_kq_mask(ctx0, mctx_cur->get_swa(), ubatch, cparams);
         ggml_set_input(inp->self_kq_mask_swa);
         ggml_set_name(inp->self_kq_mask_swa, "self_kq_mask_swa");
 
@@ -2374,27 +2387,21 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()
 
     auto inp_attn = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, attn_ctx);
 
-    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
-
     {
-        const auto n_kv = attn_ctx->get_base()->get_n_kv();
-
         inp_attn->self_k_idxs = attn_ctx->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp_attn->self_v_idxs = attn_ctx->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp_attn->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp_attn->self_kq_mask = build_kq_mask(ctx0, attn_ctx->get_base(), ubatch, cparams);
         ggml_set_input(inp_attn->self_kq_mask);
 
         inp_attn->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask, GGML_TYPE_F16) : inp_attn->self_kq_mask;
     }
 
     {
-        const auto n_kv = attn_ctx->get_swa()->get_n_kv();
-
         inp_attn->self_k_idxs_swa = attn_ctx->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp_attn->self_v_idxs_swa = attn_ctx->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp_attn->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        inp_attn->self_kq_mask_swa = build_kq_mask(ctx0, attn_ctx->get_swa(), ubatch, cparams);
         ggml_set_input(inp_attn->self_kq_mask_swa);
 
         inp_attn->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask_swa, GGML_TYPE_F16) : inp_attn->self_kq_mask_swa;

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/models/delta-net-base.cpp

@@ -0,0 +1,333 @@
+#include "models.h"
+
+#define CHUNK_SIZE 64
+
+// utility to get one slice from the third dimension
+// input dim:  [x, y, c, b]
+// output dim: [x, y, 1, b]
+static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
+    return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
+        t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
+}
+
+llm_build_delta_net_base::llm_build_delta_net_base(const llm_graph_params & params) : llm_graph_context(params) {}
+
+std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_chunking(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * b,
+        ggml_tensor * s,
+        int           il) {
+    const int64_t S_k      = q->ne[0];
+    const int64_t H_k      = q->ne[1];
+    const int64_t n_tokens = q->ne[2];
+    const int64_t n_seqs   = q->ne[3];
+
+    const int64_t S_v = v->ne[0];
+    const int64_t H_v = v->ne[1];
+
+    GGML_ASSERT(S_k == S_v);
+    GGML_ASSERT(H_v % H_k == 0);
+
+    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
+
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+
+    const float scale = 1.0f / sqrtf(S_k);
+
+    q = ggml_scale(ctx0, q, scale);
+
+    cb(q, "q_in", il);
+    cb(k, "k_in", il);
+    cb(v, "v_in", il);
+    cb(b, "b_in", il);
+    cb(g, "g_in", il);
+
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
+    g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [  1, n_tokens, H_v, n_seqs]
+    b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [  1, n_tokens, H_v, n_seqs]
+
+    const int CS = CHUNK_SIZE;
+
+    const int pad = (CS - n_tokens % CS) % CS;
+    const int n_chunks = (n_tokens + pad) / CS;
+
+    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
+    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
+    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
+    g = ggml_pad(ctx0, g, 0, pad, 0, 0);
+    b = ggml_pad(ctx0, b, 0, pad, 0, 0);
+
+    ggml_tensor * v_b = ggml_mul(ctx0, v, b);
+    ggml_tensor * k_b = ggml_mul(ctx0, k, b);
+
+    cb(v_b, "v_b", il);
+    cb(k_b, "k_b", il);
+
+    q   = ggml_reshape_4d(ctx0, q,   S_k, CS, n_chunks, H_k * n_seqs);
+    k   = ggml_reshape_4d(ctx0, k,   S_k, CS, n_chunks, H_k * n_seqs);
+    k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs);
+    v   = ggml_reshape_4d(ctx0, v,   S_v, CS, n_chunks, H_v * n_seqs);
+    v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
+
+    g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
+
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
+    cb(g_cs, "g_cs", il);
+
+    ggml_tensor * g_cs_i = g_cs;
+    ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
+
+    g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
+
+    // [CS, CS, n_chunks, H_v * n_seqs]
+    ggml_tensor * decay_mask;
+    decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+    decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
+    decay_mask = ggml_exp(ctx0, decay_mask);
+    cb(decay_mask, "decay_mask", il);
+
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * kb;
+    kb = ggml_mul_mat(ctx0, k,  k_b);
+    kb = ggml_mul    (ctx0, kb, decay_mask);
+
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * attn;
+    attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
+
+    ggml_tensor * identity;
+    identity = ggml_view_1d(ctx0, attn, CS, 0);
+    identity = ggml_fill   (ctx0, identity, 1.0f);
+    identity = ggml_diag   (ctx0, identity);
+
+    ggml_tensor * lhs = ggml_add(ctx0, attn, identity);
+    cb(lhs, "dnet_add_ch_lhs", il);
+
+    attn = ggml_neg(ctx0, attn);
+
+    ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
+    attn = ggml_add(ctx0, lin_solve, identity);
+    cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
+
+    // [S_v, CS, n_chunks, H_v * n_seqs]
+    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
+
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
+
+    k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
+
+    // [CS, S_k, n_chunks, H_k * n_seqs]
+    ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
+    cb(kbg, "k_beta_g_exp", il);
+
+    // [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
+    cb(k_cd, "k_cumdecay", il);
+
+    // [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp);
+    ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
+
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+    kq = ggml_mul(ctx0, kq, decay_mask);
+    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
+    cb(kq, "kq", il);
+
+    // vectorized calculation of key_gdiff
+    // improved from the chunked version:
+    //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
+    //   g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
+    //   key_gdiff = key * g_diff.unsqueeze(-1)
+    //   kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
+    //   last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
+
+    // get last element in g_cumsum along CS dimension (ne0)
+    // example: [[x, y, z, ..., last], ...] -> [[last], ...]
+    // [1, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
+            g_cs->nb[1],
+            g_cs->nb[2],
+            g_cs->nb[3],
+            ggml_row_size(g_cs->type, g_cs->ne[0] - 1));
+    cb(g_last, "g_last", il);
+
+    // TODO: remove this cont when CUDA supports non-cont unary ops
+    g_last = ggml_cont(ctx0, g_last);
+
+    // [1, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
+    cb(g_last_exp, "g_last_exp", il);
+
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
+    cb(g_diff, "g_diff", il);
+
+    ggml_tensor * g_diff_exp   = ggml_exp(ctx0, g_diff);
+    ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp);
+
+    // [S_k, CS, n_chunks, H_v * n_seqs]
+    ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
+    cb(kg, "key_gdiff", il);
+
+    // [CS, S_k, n_chunks, H_v * n_seqs]
+    ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg));
+    cb(kg_t, "key_gdiff_t", il);
+
+    ggml_tensor * s_t = ggml_transpose(ctx0, s);
+    s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs);
+    cb(s_t, "dnet_add_ch_state", il);
+
+    // [CS, S_v, n_chunks, H_v * n_seqs]
+    ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
+
+    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
+        ggml_tensor * ch_k_cd    = get_slice_2d(ctx0, k_cd,    chunk); // [S_k,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_v_t     = get_slice_2d(ctx0, v_t,     chunk); // [ CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * ch_kq      = get_slice_2d(ctx0, kq,      chunk); // [ CS,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_kg_t    = get_slice_2d(ctx0, kg_t,    chunk); // [ CS, S_k, 1, H_v * n_seqs]
+
+        // [CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t);
+        cb(v_t_p, "v_prime", il);
+
+        // [CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
+        cb(v_t_new, "v_t_new", il);
+
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
+        cb(v_attn, "v_attn", il);
+
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp);
+        cb(attn_inter, "attn_inter", il);
+
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn);
+        cb(o_ch, "dnet_add_ch_attn_out", il);
+
+        v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
+
+        // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
+        // TODO: head broadcast might not work here - probably will need a transpose
+        ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
+
+        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
+        ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
+        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
+        s_t = ggml_add(ctx0, s_t, kgv);
+        cb(s_t, "dnet_add_ch_state", il);
+    }
+
+    s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs);
+
+    // truncate padded tokens
+    ggml_tensor * o = ggml_view_4d(ctx0, v,
+            S_v, n_tokens, H_v, n_seqs,
+            ggml_row_size(v->type, S_v),
+            ggml_row_size(v->type, S_v * CS * n_chunks),
+            ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
+
+    o = ggml_permute  (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
+    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
+
+    return {o, s};
+}
+
+std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_autoregressive(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * b, // beta
+        ggml_tensor * s, // state
+        int           il) {
+    const int64_t S_k      = q->ne[0];
+    const int64_t H_k      = q->ne[1];
+    const int64_t n_tokens = q->ne[2];
+    const int64_t n_seqs   = q->ne[3];
+
+    const int64_t S_v = v->ne[0];
+    const int64_t H_v = v->ne[1];
+
+    GGML_ASSERT(n_tokens == 1);
+
+    GGML_ASSERT(S_k == S_v);
+    GGML_ASSERT(H_v % H_k == 0);
+
+    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
+
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+
+    const float scale = 1.0f / sqrtf(S_k);
+
+    q = ggml_scale(ctx0, q, scale);
+
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
+
+    cb(q, "q_in", il);
+    cb(k, "k_in", il);
+    cb(v, "v_in", il);
+    cb(b, "b_in", il);
+    cb(g, "g_in", il);
+
+    g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
+
+    // [S_v, S_v, H_v, n_seqs]
+    g = ggml_exp(ctx0, g);
+    s = ggml_mul(ctx0, s, g);
+
+    ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s));
+
+    // [1, S_v, H_v, n_seqs]
+    ggml_tensor * sk;
+    sk = ggml_mul     (ctx0, s_t, k);
+    sk = ggml_sum_rows(ctx0, sk);
+
+    // [S_v, 1, H_v, n_seqs]
+    ggml_tensor * d;
+    d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
+    d = ggml_mul(ctx0, d, b);
+
+    // [1, S_v, H_v, n_seqs]
+    ggml_tensor * d_t;
+    d_t = ggml_transpose(ctx0, d);
+
+    // [S_v, S_v, H_v, n_seqs]
+    ggml_tensor * kd;
+    k  = ggml_repeat(ctx0, k, s);
+    kd = ggml_mul   (ctx0, k, d_t);
+
+    s_t = ggml_add(ctx0, s_t, kd);
+
+    cb(s_t, "dnet_add_ar_state", il);
+
+    ggml_tensor * s_q = ggml_mul     (ctx0, s_t, q);
+    ggml_tensor * o   = ggml_sum_rows(ctx0, s_q);
+
+    o = ggml_permute  (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
+    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
+
+    return {o, s};
+}

src/models/falcon-h1.cpp

@@ -1,9 +1,7 @@
 #include "models.h"
 
-
-
 llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params) {
+    llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     ggml_tensor * cur;

src/models/granite-hybrid.cpp

@@ -2,7 +2,7 @@
 
 
 llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params) {
+    llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 

src/models/jamba.cpp

@@ -1,6 +1,6 @@
 #include "models.h"
 
-llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
+llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     ggml_tensor * cur;

src/models/kimi-linear.cpp

@@ -1,6 +1,8 @@
 #include "models.h"
 #include "ggml.h"
 
+#include "llama-memory-recurrent.h"
+
 #define CHUNK_SIZE 64
 
 // Causal Conv1d function for Q,K,V
@@ -65,7 +67,7 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
 }
 
 llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params), model(model) {
+    llm_build_mamba_base(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 

src/models/mamba-base.cpp

@@ -1,8 +1,10 @@
 #include "models.h"
 
-llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
+#include "llama-memory-recurrent.h"
 
-ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * inp,
+llm_build_mamba_base::llm_build_mamba_base(const llm_graph_params & params) : llm_graph_context(params) {}
+
+ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
                                                          ggml_tensor *        cur,
                                                          const llama_model &  model,
                                                          const llama_ubatch & ubatch,
@@ -143,7 +145,7 @@ ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * in
     return cur;
 }
 
-ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp,
+ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
                                                           ggml_tensor *        cur,
                                                           const llama_model &  model,
                                                           const llama_ubatch & ubatch,

src/models/mamba.cpp

@@ -1,7 +1,6 @@
 #include "models.h"
 
-
-llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
+llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 

src/models/models.h

@@ -1,23 +1,51 @@
 #pragma once
 
-#include "../llama-model.h"
-#include "../llama-graph.h"
+#include "llama-model.h"
+#include "llama-graph.h"
 
-// TODO: remove in follow-up PR - move to .cpp files
-#include "../llama-memory-recurrent.h"
+// note: almost all graphs require atleast sqrtf, so include cmath globally
 #include <cmath>
 
-struct llm_graph_context_mamba : public llm_graph_context {
-    llm_graph_context_mamba(const llm_graph_params & params);
+//
+// base classes
+//
 
-    virtual ~llm_graph_context_mamba() = default;
+struct llm_build_mamba_base : public llm_graph_context {
+    llm_build_mamba_base(const llm_graph_params & params);
+
+    virtual ~llm_build_mamba_base() = default;
 
     ggml_tensor * build_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
     ggml_tensor * build_mamba2_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il) const;
 
 };
 
-// Base class for RWKV-related models
+struct llm_build_delta_net_base : public llm_graph_context {
+    llm_build_delta_net_base(const llm_graph_params & params);
+
+    virtual ~llm_build_delta_net_base() = default;
+
+    // returns pair of output and new state
+    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * b,
+                ggml_tensor * s,
+                        int   il);
+
+    // returns pair of output and new state
+    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * b,
+                ggml_tensor * s,
+                int           il);
+};
+
 struct llm_build_rwkv6_base : public llm_graph_context {
     const llama_model & model;
 
@@ -58,6 +86,10 @@ struct llm_build_rwkv7_base : public llm_graph_context {
                                        int                  il) const;
 };
 
+//
+// models
+//
+
 struct llm_build_afmoe : public llm_graph_context {
     llm_build_afmoe(const llama_model & model, const llm_graph_params & params);
 };
@@ -175,7 +207,7 @@ struct llm_build_falcon : public llm_graph_context {
     llm_build_falcon(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_falcon_h1 : public llm_graph_context_mamba {
+struct llm_build_falcon_h1 : public llm_build_mamba_base {
     llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params);
 };
 
@@ -253,7 +285,7 @@ private:
         const int                 il);
 };
 
-struct llm_build_granite_hybrid : public llm_graph_context_mamba {
+struct llm_build_granite_hybrid : public llm_build_mamba_base {
     llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params);
     ggml_tensor * build_layer_ffn(ggml_tensor * cur, ggml_tensor * inpSA, const llama_model & model, const int il);
     ggml_tensor * build_attention_layer(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn,
@@ -284,11 +316,12 @@ struct llm_build_jais : public llm_graph_context {
     llm_build_jais(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_jamba : public llm_graph_context_mamba {
+struct llm_build_jamba : public llm_build_mamba_base {
     llm_build_jamba(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_kimi_linear : public llm_graph_context_mamba {
+// TODO: derive llm_build_delta_net_base instead
+struct llm_build_kimi_linear : public llm_build_mamba_base {
     llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params);
 
     std::pair<ggml_tensor *, ggml_tensor *> build_kda_autoregressive(
@@ -347,7 +380,7 @@ struct llm_build_maincoder : public llm_graph_context {
     llm_build_maincoder(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_mamba : public llm_graph_context_mamba {
+struct llm_build_mamba : public llm_build_mamba_base {
     llm_build_mamba(const llama_model & model, const llm_graph_params & params);
 };
 
@@ -379,11 +412,11 @@ struct llm_build_nemotron : public llm_graph_context {
     llm_build_nemotron(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_nemotron_h : public llm_graph_context_mamba {
+struct llm_build_nemotron_h : public llm_build_mamba_base {
     llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params);
-    ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il);
+    ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il);
     ggml_tensor * build_attention_layer(ggml_tensor * cur, llm_graph_input_attn_kv * inp_attn,
-        const llama_model & model, const int64_t n_embd_head, const int il);
+        const llama_model & model, int64_t n_embd_head, int il);
 };
 
 struct llm_build_neo_bert : public llm_graph_context {
@@ -428,7 +461,7 @@ struct llm_build_phi3 : public llm_graph_context {
     llm_build_phi3(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_plamo2 : public llm_graph_context_mamba {
+struct llm_build_plamo2 : public llm_build_mamba_base {
     llm_build_plamo2(const llama_model & model, const llm_graph_params & params);
     private:
         ggml_tensor * build_plamo2_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
@@ -477,7 +510,7 @@ struct llm_build_qwen3vlmoe : public llm_graph_context {
     llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_qwen3next : public llm_graph_context_mamba {
+struct llm_build_qwen3next : public llm_build_delta_net_base {
     llm_build_qwen3next(const llama_model & model, const llm_graph_params & params);
 private:
     ggml_tensor * build_layer_attn(
@@ -489,38 +522,12 @@ private:
     ggml_tensor * build_layer_attn_linear(
          llm_graph_input_rs * inp,
                 ggml_tensor * cur,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
-                        int   il);
-
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                int           il);
-
     ggml_tensor * build_norm_gated(
                 ggml_tensor * input,
                 ggml_tensor * weights,
@@ -535,7 +542,8 @@ private:
     const llama_model & model;
 };
 
-struct llm_build_qwen35 : public llm_graph_context_mamba {
+// TODO: derive llm_build_delta_net_base instead
+struct llm_build_qwen35 : public llm_graph_context {
     llm_build_qwen35(const llama_model & model, const llm_graph_params & params);
 private:
     ggml_tensor * build_layer_attn(
@@ -553,6 +561,7 @@ private:
                 ggml_tensor * diag_mask,
                         int   il);
 
+
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
@@ -594,7 +603,8 @@ private:
     const llama_model & model;
 };
 
-struct llm_build_qwen35moe : public llm_graph_context_mamba {
+// TODO: derive llm_build_delta_net_base instead
+struct llm_build_qwen35moe : public llm_graph_context {
     llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params);
 private:
     ggml_tensor * build_layer_attn(

src/models/nemotron-h.cpp

@@ -1,9 +1,7 @@
 #include "models.h"
 
-
-
 llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params) {
+    llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
@@ -65,8 +63,8 @@ llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_
 ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *             cur,
                                                           llm_graph_input_attn_kv * inp_attn,
                                                           const llama_model &       model,
-                                                          const int64_t             n_embd_head,
-                                                          const int                 il) {
+                                                                int64_t             n_embd_head,
+                                                                int                 il) {
     // compute Q and K
     ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
     cb(Qcur, "Qcur", il);
@@ -106,7 +104,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
     return cur;
 }
 
-ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
+ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {
     if (model.layers[il].ffn_gate_inp == nullptr) {
         cur = build_ffn(cur,
                 model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,

src/models/plamo2.cpp

@@ -1,7 +1,9 @@
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params) {
+    llm_build_mamba_base(params) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 

src/models/qwen35.cpp

@@ -1,10 +1,11 @@
-#include "ggml.h"
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 #define CHUNK_SIZE 64
 
 llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params), model(model) {
+    llm_graph_context(params), model(model) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);

src/models/qwen35moe.cpp

@@ -1,10 +1,11 @@
-#include "ggml.h"
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 #define CHUNK_SIZE 64
 
 llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params), model(model) {
+    llm_graph_context(params), model(model) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);

src/models/qwen3next.cpp

@@ -1,10 +1,9 @@
-#include "ggml.h"
 #include "models.h"
 
-#define CHUNK_SIZE 64
+#include "llama-memory-recurrent.h"
 
 llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context_mamba(params), model(model) {
+    llm_build_delta_net_base(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
@@ -16,17 +15,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
-                    GGML_TRI_TYPE_LOWER);
-
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
-    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
-
-    ggml_build_forward_expand(gf, causal_mask);
-    ggml_build_forward_expand(gf, identity);
-    ggml_build_forward_expand(gf, diag_mask);
-
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -36,7 +24,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
@@ -94,354 +82,6 @@ static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t
         t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
 }
 
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chunking(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
-        ggml_tensor * diag_mask,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q = ggml_scale(ctx0, q, scale);
-
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
-
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
-
-    // Do padding
-    const int64_t chunk_size = CHUNK_SIZE;
-
-    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
-    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
-
-    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
-    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
-    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
-    g = ggml_pad(ctx0, g, pad, 0, 0, 0);
-    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
-
-    cb(q, "q_pad", il);
-    cb(k, "k_pad", il);
-    cb(v, "v_pad", il);
-    cb(beta, "beta_pad", il);
-    cb(g, "g_pad", il);
-
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
-
-    cb(v_beta, "v_beta", il);
-    cb(k_beta, "k_beta", il);
-
-    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
-
-    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
-
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
-    cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * gcs_j_broadcast =
-        ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-    cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
-
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-    cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-    cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
-
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
-
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-    cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-    cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
-    attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
-    attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
-    cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-
-    // vectorized calculation of key_gdiff
-    // improved from the chunked version:
-    //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
-    //   g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
-    //   key_gdiff = key * g_diff.unsqueeze(-1)
-    //   kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-    //   last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-
-    // get last element in g_cumsum along chunk_size dimension (ne0)
-    // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
-                                        g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
-                                        (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
-    g_last = ggml_cont(ctx0, g_last);
-    cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
-    cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
-                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
-
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
-    cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
-    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
-
-
-    // state to be updated per chunk
-    ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
-    cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
-
-    // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
-    ggml_tensor * core_attn_out = nullptr;
-
-    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
-        // shape: (S_k, chunk_size, 1, H_k * n_seqs)
-        ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
-
-        // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-        ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
-
-        // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-        // replaced by precomputed attn_kq
-        ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
-        cb(attn_chunk, "attn_chunk", il);
-
-        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
-
-        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
-        cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
-
-        // v_new = v_i - v_prime
-        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
-        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-        cb(v_new, "v_new_chunk", il);
-
-        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-        ggml_tensor * q_g_exp    = ggml_mul(ctx0, q_chunk, gexp_chunk);
-        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-        cb(attn_inter, "attn_inter_chunk", il);
-
-        // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
-        cb(v_attn, "v_attn_chunk", il);
-
-        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
-        cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-
-        core_attn_out = core_attn_out == nullptr
-            ? core_attn_out_chunk
-            : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
-
-        // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
-        //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
-
-        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
-        new_state = ggml_add(ctx0,
-            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
-            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
-    }
-
-    // truncate padded tokens
-    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
-            S_v, n_tokens, H_v, n_seqs,
-            ggml_row_size(core_attn_out->type, S_v),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-    cb(output_tokens, "output_tokens", il);
-
-    // permute back to (S_v, H_v, n_tokens, n_seqs)
-    output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-
-    return {output_tokens, new_state};
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q    = ggml_scale(ctx0, q, scale);
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
-    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
-
-    // Apply exponential to g_t
-    g_t = ggml_exp(ctx0, g_t);
-
-    // Apply the gated delta rule for the single timestep
-    // last_recurrent_state = last_recurrent_state * g_t
-    state = ggml_mul(ctx0, state, g_t);
-
-    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
-    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
-    // we need to sum over dim=-2, so we transpose, sum, then transpose again
-    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
-
-    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
-    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    // delta = (v_t - kv_mem) * beta_t
-    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
-    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
-
-    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
-    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
-    state                   = ggml_add(ctx0, state, k_t_delta);
-
-    // Compute the attention output
-    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
-    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
-    // again, since it's over dim = -2, transpose, sum, transpose back
-    ggml_tensor * core_attn_out =
-        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
-
-    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
-    cb(core_attn_out, "output_tokens", il);
-    cb(state, "new_state", il);
-
-    return {core_attn_out, state};
-}
-
 ggml_tensor * llm_build_qwen3next::build_norm_gated(
         ggml_tensor * input,
         ggml_tensor * weights,
@@ -472,39 +112,29 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
     // Split Q projection into query and gate
     // The split should be along dimension 0 (the feature dimension)
     ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
-                                             Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
+                                            Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
+    cb(Qcur, "Qcur_view", il);
+
     ggml_tensor * gate =
         ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
                      Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
-    cb(Qcur, "Qcur", il);
     cb(gate, "gate", il);
 
-    // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention
-    Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-    cb(Qcur, "Qcur_reshaped", il);
-
-    // Apply Q normalization
-    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
-    cb(Qcur, "Qcur_normed", il);
-
     ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
     cb(Kcur, "Kcur", il);
 
     ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
     cb(Vcur, "Vcur", il);
 
-    // Apply K normalization
     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Qcur, "Qcur_normed", il);
+
     Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
     cb(Kcur, "Kcur_normed", il);
 
-    // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads)
-    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
-    cb(gate, "gate_reshaped", il);
-
-    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-    // Apply RoPE
     Qcur = ggml_rope_ext(
             ctx0, Qcur, inp_pos, nullptr,
             n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -519,7 +149,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
     cb(Kcur, "Kcur", il);
     cb(Vcur, "Vcur", il);
 
-    // Attention computation
     const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
     cur = build_attn(inp,
@@ -527,10 +156,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
                 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
     cb(cur, "attn_pregate", il);
 
-    ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
-    cb(gate_sigmoid, "gate_sigmoid", il);
+    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
+    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
 
-    cur = ggml_mul(ctx0, cur, gate_sigmoid);
+    gate = ggml_sigmoid(ctx0, gate);
+    cb(gate, "gate_sigmoid", il);
+
+    gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
+
+    cur = ggml_mul(ctx0, cur, gate);
     cb(cur, "attn_gated", il);
 
     cur = build_lora_mm(model.layers[il].wo, cur);
@@ -560,7 +194,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
         cb(z, "z", il);
 
         return { qkv_mixed, z };
-
     } else {
         // legacy (slower) path
         ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);
@@ -624,9 +257,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
 ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
         llm_graph_input_rs * inp,
         ggml_tensor *        cur,
-        ggml_tensor *        causal_mask,
-        ggml_tensor *        identity,
-        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -671,7 +301,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
                                    split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));
     cb(a, "a", il);
 
-    ggml_tensor * beta  = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs);
+    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
+    b = ggml_cont(ctx0, b);
+
+    ggml_tensor * beta = ggml_sigmoid(ctx0, b);
+
+    beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
 
     // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
     ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
@@ -679,6 +314,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
+
     ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
     cb(gate, "gate", il);
 
@@ -686,8 +322,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
     ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
 
-    // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
-
     // Build the convolution states tensor
     ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     cb(conv_states, "conv_states", il);
@@ -696,11 +330,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;
     const int64_t conv_kernel_size = conv_kernel->ne[0];
     const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
-    conv_states                    = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
+
+    conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
     cb(conv_states, "conv_states_reshaped", il);
 
-    qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
-    cb(qkv_mixed, "qkv_mixed_permuted", il);
+    qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
+    cb(qkv_mixed, "qkv_mixed_transposed", il);
 
     ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
     cb(conv_input, "conv_input", il);
@@ -720,7 +355,10 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
     cb(conv_states_all, "conv_states_updated", il);
 
-    // Apply SSM convolution
+    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
+    cb(state, "state_predelta", il);
+
     ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
     cb(conv_output_proper, "conv_output_raw", il);
 
@@ -734,26 +372,36 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
 
     // Extract the convolved Q, K, V from conv_output
-    ggml_tensor * q_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
+    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            0);
+
+    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+
+    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_v_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
+
     cb(q_conv, "q_conv", il);
-    ggml_tensor * k_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(k_conv, "k_conv", il);
-    ggml_tensor * v_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(v_conv, "v_conv", il);
 
-    // Unsqueeze them
-    q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
-    state               = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
-    cb(state, "state_predelta", il);
+    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
+    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
+
+    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // if head keys and value keys are different, repeat to force tensors into matching shapes
     if (num_k_heads != num_v_heads) {
@@ -786,7 +434,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
     }
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
@@ -795,19 +443,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
 
     // Update the recurrent states
     ggml_build_forward_expand(gf,
-                              ggml_cpy(ctx0, new_state,
-                                       ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
-                                                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
-
-    // Reshape both attn_out_final and z to 2D tensors for normalization
-    // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+            ggml_cpy(ctx0, new_state,
+                ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
 
     // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // Apply gated normalization: self.norm(core_attn_out, z)
-    ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
+    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
 
     // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
     ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -818,7 +462,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(cur, "linear_attn_out", il);
 
     // Reshape back to original dimensions
-    cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+
     return cur;
 }
 
@@ -839,7 +484,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
         if (model.layers[il].ffn_up_shexp != nullptr) {
             ggml_tensor * ffn_shexp =
                 build_ffn(cur,
-                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
                     model.layers[il].ffn_gate_shexp, NULL, NULL,
                     model.layers[il].ffn_down_shexp, NULL, NULL,
                     NULL,
@@ -852,11 +497,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
             ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
             cb(shared_gate, "shared_expert_gate", il);
 
-            // Apply sigmoid to the gate
             shared_gate = ggml_sigmoid(ctx0, shared_gate);
             cb(shared_gate, "shared_expert_gate_sigmoid", il);
 
-            // Apply the gate to the shared expert output
             ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
             cb(ffn_shexp, "ffn_shexp_gated", il);
 

src/models/rwkv6-base.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model) {}

src/models/rwkv7-base.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 llm_build_rwkv7_base::llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model) {}

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;

tests/test-backend-ops.cpp

@@ -8301,7 +8301,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                         //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
                                         for (int kv : { 113, 512, 1024, }) {
                                             if (nr2 != 1 && kv != 512) continue;
-                                            for (int nb : { 1, 3, 32, 35, }) {
+                                            for (int nb : { 1, 3, 32, 75, }) {
                                                 for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
                                                     if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
                                                     for (ggml_type type_KV : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {

tools/mtmd/CMakeLists.txt

@@ -20,6 +20,7 @@ add_library(mtmd
             models/internvl.cpp
             models/kimivl.cpp
             models/kimik25.cpp
+            models/nemotron-v2-vl.cpp
             models/llama4.cpp
             models/llava.cpp
             models/minicpmv.cpp

tools/mtmd/clip-impl.h

@@ -236,6 +236,7 @@ enum projector_type {
     PROJECTOR_TYPE_GLM4V,
     PROJECTOR_TYPE_YOUTUVL,
     PROJECTOR_TYPE_KIMIK25,
+    PROJECTOR_TYPE_NEMOTRON_V2_VL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -270,6 +271,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GLM4V,     "glm4v"},
     { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
     { PROJECTOR_TYPE_KIMIK25,   "kimik25"},
+    { PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip-model.h

@@ -15,6 +15,7 @@ enum ffn_op_type {
     FFN_GELU_ERF,
     FFN_SILU,
     FFN_GELU_QUICK,
+    FFN_RELU_SQR,
 };
 
 enum norm_type {

tools/mtmd/clip.cpp

@@ -559,6 +559,12 @@ ggml_tensor * clip_graph::build_ffn(
                 cur = ggml_gelu_quick(ctx0, cur);
                 cb(cur, "ffn_gelu_quick", il);
             } break;
+        case FFN_RELU_SQR:
+            {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_relu_sqr", il);
+            } break;
     }
 
     if (down) {
@@ -810,6 +816,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_internvl>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
+            {
+                builder = std::make_unique<clip_graph_nemotron_v2_vl>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_LLAMA4:
             {
                 builder = std::make_unique<clip_graph_llama4>(ctx, img);
@@ -1110,6 +1120,7 @@ struct clip_model_loader {
                         }
                     } break;
                 case PROJECTOR_TYPE_INTERNVL:
+                case PROJECTOR_TYPE_NEMOTRON_V2_VL:
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
@@ -1767,6 +1778,12 @@ struct clip_model_loader {
                     model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
                     model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
                 } break;
+            case PROJECTOR_TYPE_NEMOTRON_V2_VL:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
+                    model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
+                } break;
             case PROJECTOR_TYPE_GLMA:
                 {
                     model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -3088,6 +3105,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         case PROJECTOR_TYPE_GLM_EDGE:
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
             {
                 clip_image_u8 resized_image;
                 int sz = params.image_size;
@@ -3397,6 +3415,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
         case PROJECTOR_TYPE_LLAMA4:
             {
                 // both X and Y are downscaled by the scale factor
@@ -3805,6 +3824,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_GEMMA3NV:
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_GLMA:
         case PROJECTOR_TYPE_ULTRAVOX:
@@ -3968,6 +3988,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
             return ctx->model.mm_3_w->ne[1];
         case PROJECTOR_TYPE_LLAMA4:
             return ctx->model.mm_model_proj->ne[1];

tools/mtmd/models/models.h

@@ -42,6 +42,11 @@ struct clip_graph_internvl : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_nemotron_v2_vl : clip_graph {
+    clip_graph_nemotron_v2_vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_llama4 : clip_graph {
     clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;

tools/mtmd/models/nemotron-v2-vl.cpp

@@ -0,0 +1,35 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_nemotron_v2_vl::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_registers = model.class_embedding->ne[1];
+    const int n_pos = n_patches + n_registers;
+
+    ggml_tensor * inp = build_inp();
+
+    // add position embeddings (pre-downsampled during GGUF conversion for fixed 512x512 input)
+    inp = ggml_add(ctx0, inp, model.position_embeddings);
+    cb(inp, "inp_pos", -1);
+
+    inp = ggml_concat(ctx0, model.class_embedding, inp, 1);
+
+    ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, nullptr);
+
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd),
+        n_registers * ggml_row_size(cur->type, n_embd));
+
+    cur = build_patch_merge_permute(cur, model.hparams.n_merge);
+
+    {
+        cur = build_norm(cur, model.mm_0_w, nullptr, NORM_TYPE_RMS, 1e-6, -1);
+        cur = build_ffn(cur, model.mm_1_w, nullptr, nullptr, nullptr, model.mm_3_w, nullptr, FFN_RELU_SQR, -1);
+    }
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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/rpc/rpc-server.cpp

@@ -132,7 +132,8 @@ static std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
+    defined(__OpenBSD__) || defined(__NetBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else if (std::getenv("HOME")) {

tools/server/CMakeLists.txt

@@ -28,10 +28,6 @@ target_link_libraries(${TARGET} PUBLIC common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
 set(TARGET llama-server)
 
-if (NOT LLAMA_HTTPLIB)
-    message(FATAL_ERROR "LLAMA_HTTPLIB is OFF, cannot build llama-server. Hint: to skip building server, set -DLLAMA_BUILD_SERVER=OFF")
-endif()
-
 set(TARGET_SRCS
     server.cpp
     server-http.cpp

tools/server/webui/.storybook/main.ts

@@ -1,17 +1,24 @@
 import type { StorybookConfig } from '@storybook/sveltekit';
+import { dirname, resolve } from 'path';
+import { fileURLToPath } from 'url';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
 
 const config: StorybookConfig = {
 	stories: ['../tests/stories/**/*.mdx', '../tests/stories/**/*.stories.@(js|ts|svelte)'],
 	addons: [
 		'@storybook/addon-svelte-csf',
 		'@chromatic-com/storybook',
-		'@storybook/addon-docs',
+		'@storybook/addon-vitest',
 		'@storybook/addon-a11y',
-		'@storybook/addon-vitest'
+		'@storybook/addon-docs'
 	],
-	framework: {
-		name: '@storybook/sveltekit',
-		options: {}
+	framework: '@storybook/sveltekit',
+	viteFinal: async (config) => {
+		config.server = config.server || {};
+		config.server.fs = config.server.fs || {};
+		config.server.fs.allow = [...(config.server.fs.allow || []), resolve(__dirname, '../tests')];
+		return config;
 	}
 };
 export default config;

tools/server/webui/.storybook/preview.ts

@@ -13,7 +13,7 @@ const preview: Preview = {
 		},
 
 		backgrounds: {
-			disable: true
+			disabled: true
 		},
 
 		a11y: {

tools/server/webui/docs/flows/settings-flow.md

@@ -49,14 +49,20 @@ sequenceDiagram
     settingsStore->>serverStore: defaultParams
     serverStore-->>settingsStore: {temperature, top_p, top_k, ...}
 
-    settingsStore->>ParamSvc: extractServerDefaults(defaultParams)
-    ParamSvc-->>settingsStore: Record<string, value>
+    loop each SYNCABLE_PARAMETER
+        alt key NOT in userOverrides
+            settingsStore->>settingsStore: config[key] = serverDefault[key]
+            Note right of settingsStore: Non-overridden params adopt server default
+        else key in userOverrides
+            Note right of settingsStore: Keep user value, skip server default
+        end
+    end
 
-    settingsStore->>ParamSvc: mergeWithServerDefaults(config, serverDefaults)
-    Note right of ParamSvc: For each syncable parameter:<br/>- If NOT in userOverrides → use server default<br/>- If in userOverrides → keep user value
-    ParamSvc-->>settingsStore: mergedConfig
+    alt serverStore.props has webuiSettings
+        settingsStore->>settingsStore: Apply webuiSettings from server
+        Note right of settingsStore: Server-provided UI settings<br/>(e.g. showRawOutputSwitch)
+    end
 
-    settingsStore->>settingsStore: config = mergedConfig
     settingsStore->>settingsStore: saveConfig()
     deactivate settingsStore
 
@@ -67,11 +73,18 @@ sequenceDiagram
     UI->>settingsStore: updateConfig(key, value)
     activate settingsStore
     settingsStore->>settingsStore: config[key] = value
-    settingsStore->>settingsStore: userOverrides.add(key)
-    Note right of settingsStore: Mark as user-modified (won't be overwritten by server)
+
+    alt value matches server default for key
+        settingsStore->>settingsStore: userOverrides.delete(key)
+        Note right of settingsStore: Matches server default, remove override
+    else value differs from server default
+        settingsStore->>settingsStore: userOverrides.add(key)
+        Note right of settingsStore: Mark as user-modified (won't be overwritten)
+    end
+
     settingsStore->>settingsStore: saveConfig()
-    settingsStore->>LS: set("llama-config", config)
-    settingsStore->>LS: set("llama-userOverrides", [...userOverrides])
+    settingsStore->>LS: set(CONFIG_LOCALSTORAGE_KEY, config)
+    settingsStore->>LS: set(USER_OVERRIDES_LOCALSTORAGE_KEY, [...userOverrides])
     deactivate settingsStore
 
     UI->>settingsStore: updateMultipleConfig({key1: val1, key2: val2})
@@ -88,10 +101,9 @@ sequenceDiagram
 
     UI->>settingsStore: resetConfig()
     activate settingsStore
-    settingsStore->>settingsStore: config = SETTING_CONFIG_DEFAULT
+    settingsStore->>settingsStore: config = {...SETTING_CONFIG_DEFAULT}
     settingsStore->>settingsStore: userOverrides.clear()
-    settingsStore->>settingsStore: syncWithServerDefaults()
-    Note right of settingsStore: Apply server defaults for syncable params
+    Note right of settingsStore: All params reset to defaults<br/>Next syncWithServerDefaults will adopt server values
     settingsStore->>settingsStore: saveConfig()
     deactivate settingsStore
 

tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte

@@ -1,6 +1,6 @@
 <script lang="ts">
 	import { Eye } from '@lucide/svelte';
-	import ActionIconCopyToClipboard from '$lib/components/app/actions/ActionIconCopyToClipboard.svelte';
+	import { ActionIconCopyToClipboard } from '$lib/components/app';
 	import { FileTypeText } from '$lib/enums';
 
 	interface Props {

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatForm.svelte

@@ -57,13 +57,13 @@
 	let currentConfig = $derived(config());
 	let fileInputRef: ChatFormFileInputInvisible | undefined = $state(undefined);
 	let isRecording = $state(false);
-	let message = $state(initialMessage);
+	let message = $derived(initialMessage);
 	let pasteLongTextToFileLength = $derived.by(() => {
 		const n = Number(currentConfig.pasteLongTextToFileLen);
 		return Number.isNaN(n) ? Number(SETTING_CONFIG_DEFAULT.pasteLongTextToFileLen) : n;
 	});
-	let previousIsLoading = $state(isLoading);
-	let previousInitialMessage = $state(initialMessage);
+	let previousIsLoading = $derived(isLoading);
+	let previousInitialMessage = $derived(initialMessage);
 	let recordingSupported = $state(false);
 	let textareaRef: ChatFormTextarea | undefined = $state(undefined);
 
@@ -289,7 +289,7 @@
 
 <form
 	onsubmit={handleSubmit}
-	class="{INPUT_CLASSES} border-radius-bottom-none mx-auto max-w-[48rem] overflow-hidden rounded-3xl backdrop-blur-md {disabled
+	class="relative {INPUT_CLASSES} border-radius-bottom-none mx-auto max-w-[48rem] overflow-hidden rounded-3xl backdrop-blur-md {disabled
 		? 'cursor-not-allowed opacity-60'
 		: ''} {className}"
 	data-slot="chat-form"
@@ -304,10 +304,11 @@
 	/>
 
 	<div
-		class="flex-column relative min-h-[48px] items-center rounded-3xl px-5 py-3 shadow-sm transition-all focus-within:shadow-md"
+		class="flex-column relative min-h-[48px] items-center rounded-3xl py-2 pb-2.25 shadow-sm transition-all focus-within:shadow-md md:!py-3"
 		onpaste={handlePaste}
 	>
 		<ChatFormTextarea
+			class="px-5 py-1.5 md:pt-0"
 			bind:this={textareaRef}
 			bind:value={message}
 			onKeydown={handleKeydown}
@@ -315,6 +316,7 @@
 		/>
 
 		<ChatFormActions
+			class="px-3"
 			bind:this={chatFormActionsRef}
 			canSend={message.trim().length > 0 || uploadedFiles.length > 0}
 			hasText={message.trim().length > 0}

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAttachmentsDropdown.svelte

@@ -0,0 +1,189 @@
+<script lang="ts">
+	import { page } from '$app/state';
+	import { MessageSquare, Plus } from '@lucide/svelte';
+	import { Button } from '$lib/components/ui/button';
+	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { FILE_TYPE_ICONS } from '$lib/constants/icons';
+	import { TOOLTIP_DELAY_DURATION } from '$lib/constants/tooltip-config';
+
+	interface Props {
+		class?: string;
+		disabled?: boolean;
+		hasAudioModality?: boolean;
+		hasVisionModality?: boolean;
+		onFileUpload?: () => void;
+		onSystemPromptClick?: () => void;
+	}
+
+	type AttachmentActionId = 'images' | 'audio' | 'text' | 'pdf' | 'system';
+
+	interface AttachmentAction {
+		id: AttachmentActionId;
+		label: string;
+		disabled?: boolean;
+		disabledReason?: string;
+		tooltip?: string;
+	}
+
+	let {
+		class: className = '',
+		disabled = false,
+		hasAudioModality = false,
+		hasVisionModality = false,
+		onFileUpload,
+		onSystemPromptClick
+	}: Props = $props();
+
+	let isNewChat = $derived(!page.params.id);
+	let systemMessageTooltip = $derived(
+		isNewChat
+			? 'Add custom system message for a new conversation'
+			: 'Inject custom system message at the beginning of the conversation'
+	);
+
+	let actions = $derived.by<AttachmentAction[]>(() => [
+		{
+			id: 'images',
+			label: 'Images',
+			disabled: !hasVisionModality,
+			disabledReason: !hasVisionModality
+				? 'Images require vision models to be processed'
+				: undefined
+		},
+		{
+			id: 'audio',
+			label: 'Audio Files',
+			disabled: !hasAudioModality,
+			disabledReason: !hasAudioModality
+				? 'Audio files require audio models to be processed'
+				: undefined
+		},
+		{
+			id: 'text',
+			label: 'Text Files'
+		},
+		{
+			id: 'pdf',
+			label: 'PDF Files',
+			tooltip: !hasVisionModality
+				? 'PDFs will be converted to text. Image-based PDFs may not work properly.'
+				: undefined
+		},
+		{
+			id: 'system',
+			label: 'System Message',
+			tooltip: systemMessageTooltip
+		}
+	]);
+
+	function handleActionClick(id: AttachmentActionId) {
+		if (id === 'system') {
+			onSystemPromptClick?.();
+			return;
+		}
+
+		onFileUpload?.();
+	}
+
+	const triggerTooltipText = 'Add files or system message';
+	const itemClass = 'flex cursor-pointer items-center gap-2';
+</script>
+
+<div class="flex items-center gap-1 {className}">
+	<DropdownMenu.Root>
+		<DropdownMenu.Trigger name="Attach files" {disabled}>
+			<Tooltip.Root>
+				<Tooltip.Trigger class="w-full">
+					<Button
+						class="file-upload-button h-8 w-8 rounded-full p-0"
+						{disabled}
+						variant="secondary"
+						type="button"
+					>
+						<span class="sr-only">{triggerTooltipText}</span>
+
+						<Plus class="h-4 w-4" />
+					</Button>
+				</Tooltip.Trigger>
+
+				<Tooltip.Content>
+					<p>{triggerTooltipText}</p>
+				</Tooltip.Content>
+			</Tooltip.Root>
+		</DropdownMenu.Trigger>
+
+		<DropdownMenu.Content align="start" class="w-56">
+			{#each actions as item (item.id)}
+				{@const hasDisabledTooltip = !!item.disabled && !!item.disabledReason}
+				{@const hasEnabledTooltip = !item.disabled && !!item.tooltip}
+
+				{#if hasDisabledTooltip}
+					<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+						<Tooltip.Trigger class="w-full">
+							<DropdownMenu.Item class={itemClass} disabled>
+								{#if item.id === 'images'}
+									<FILE_TYPE_ICONS.image class="h-4 w-4" />
+								{:else if item.id === 'audio'}
+									<FILE_TYPE_ICONS.audio class="h-4 w-4" />
+								{:else if item.id === 'text'}
+									<FILE_TYPE_ICONS.text class="h-4 w-4" />
+								{:else if item.id === 'pdf'}
+									<FILE_TYPE_ICONS.pdf class="h-4 w-4" />
+								{:else}
+									<MessageSquare class="h-4 w-4" />
+								{/if}
+
+								<span>{item.label}</span>
+							</DropdownMenu.Item>
+						</Tooltip.Trigger>
+
+						<Tooltip.Content side="right">
+							<p>{item.disabledReason}</p>
+						</Tooltip.Content>
+					</Tooltip.Root>
+				{:else if hasEnabledTooltip}
+					<Tooltip.Root delayDuration={TOOLTIP_DELAY_DURATION}>
+						<Tooltip.Trigger class="w-full">
+							<DropdownMenu.Item class={itemClass} onclick={() => handleActionClick(item.id)}>
+								{#if item.id === 'images'}
+									<FILE_TYPE_ICONS.image class="h-4 w-4" />
+								{:else if item.id === 'audio'}
+									<FILE_TYPE_ICONS.audio class="h-4 w-4" />
+								{:else if item.id === 'text'}
+									<FILE_TYPE_ICONS.text class="h-4 w-4" />
+								{:else if item.id === 'pdf'}
+									<FILE_TYPE_ICONS.pdf class="h-4 w-4" />
+								{:else}
+									<MessageSquare class="h-4 w-4" />
+								{/if}
+
+								<span>{item.label}</span>
+							</DropdownMenu.Item>
+						</Tooltip.Trigger>
+
+						<Tooltip.Content side="right">
+							<p>{item.tooltip}</p>
+						</Tooltip.Content>
+					</Tooltip.Root>
+				{:else}
+					<DropdownMenu.Item class={itemClass} onclick={() => handleActionClick(item.id)}>
+						{#if item.id === 'images'}
+							<FILE_TYPE_ICONS.image class="h-4 w-4" />
+						{:else if item.id === 'audio'}
+							<FILE_TYPE_ICONS.audio class="h-4 w-4" />
+						{:else if item.id === 'text'}
+							<FILE_TYPE_ICONS.text class="h-4 w-4" />
+						{:else if item.id === 'pdf'}
+							<FILE_TYPE_ICONS.pdf class="h-4 w-4" />
+						{:else}
+							<MessageSquare class="h-4 w-4" />
+						{/if}
+
+						<span>{item.label}</span>
+					</DropdownMenu.Item>
+				{/if}
+			{/each}
+		</DropdownMenu.Content>
+	</DropdownMenu.Root>
+</div>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte

@@ -2,7 +2,7 @@
 	import { Square } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import {
-		ChatFormActionFileAttachments,
+		ChatFormActionAttachmentsDropdown,
 		ChatFormActionRecord,
 		ChatFormActionSubmit,
 		ModelsSelector
@@ -157,7 +157,7 @@
 
 	const { handleModelChange } = useModelChangeValidation({
 		getRequiredModalities: () => usedModalities(),
-		onValidationFailure: async (previousModelId) => {
+		onValidationFailure: async (previousModelId: string | null) => {
 			if (previousModelId) {
 				await modelsStore.selectModelById(previousModelId);
 			}
@@ -166,32 +166,39 @@
 </script>
 
 <div class="flex w-full items-center gap-3 {className}" style="container-type: inline-size">
-	<ChatFormActionFileAttachments
-		class="mr-auto"
-		{disabled}
-		{hasAudioModality}
-		{hasVisionModality}
-		{onFileUpload}
-		{onSystemPromptClick}
-	/>
+	<div class="mr-auto flex items-center gap-2">
+		<ChatFormActionAttachmentsDropdown
+			{disabled}
+			{hasAudioModality}
+			{hasVisionModality}
+			{onFileUpload}
+			{onSystemPromptClick}
+		/>
+	</div>
 
-	<ModelsSelector
-		{disabled}
-		bind:this={selectorModelRef}
-		currentModel={conversationModel}
-		forceForegroundText={true}
-		useGlobalSelection={true}
-		onModelChange={handleModelChange}
-	/>
+	<div class="ml-auto flex items-center gap-1.5">
+		<ModelsSelector
+			{disabled}
+			bind:this={selectorModelRef}
+			currentModel={conversationModel}
+			forceForegroundText={true}
+			useGlobalSelection={true}
+			onModelChange={handleModelChange}
+		/>
+	</div>
 
 	{#if isLoading}
 		<Button
 			type="button"
+			variant="secondary"
 			onclick={onStop}
-			class="h-8 w-8 bg-transparent p-0 hover:bg-destructive/20"
+			class="group h-8 w-8 rounded-full p-0 hover:bg-destructive/10!"
 		>
 			<span class="sr-only">Stop</span>
-			<Square class="h-8 w-8 fill-destructive stroke-destructive" />
+
+			<Square
+				class="h-8 w-8 fill-muted-foreground stroke-muted-foreground group-hover:fill-destructive group-hover:stroke-destructive hover:fill-destructive hover:stroke-destructive"
+			/>
 		</Button>
 	{:else if shouldShowRecordButton}
 		<ChatFormActionRecord {disabled} {hasAudioModality} {isLoading} {isRecording} {onMicClick} />

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte

@@ -62,8 +62,8 @@
 		assistantMessages: number;
 		messageTypes: string[];
 	} | null>(null);
-	let editedContent = $state(message.content);
-	let editedExtras = $state<DatabaseMessageExtra[]>(message.extra ? [...message.extra] : []);
+	let editedContent = $derived(message.content);
+	let editedExtras = $derived<DatabaseMessageExtra[]>(message.extra ? [...message.extra] : []);
 	let editedUploadedFiles = $state<ChatUploadedFile[]>([]);
 	let isEditing = $state(false);
 	let showDeleteDialog = $state(false);

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte

@@ -105,7 +105,7 @@
 
 	const { handleModelChange } = useModelChangeValidation({
 		getRequiredModalities: () => conversationsStore.getModalitiesUpToMessage(message.id),
-		onSuccess: (modelName) => onRegenerate(modelName)
+		onSuccess: (modelName: string) => onRegenerate(modelName)
 	});
 
 	function handleCopyModel() {

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte

@@ -133,7 +133,7 @@
 
 	const { handleModelChange } = useModelChangeValidation({
 		getRequiredModalities,
-		onValidationFailure: async (previousModelId) => {
+		onValidationFailure: async (previousModelId: string | null) => {
 			if (previousModelId) {
 				await modelsStore.selectModelById(previousModelId);
 			}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte

@@ -28,7 +28,7 @@
 		initialView = ChatMessageStatsView.GENERATION
 	}: Props = $props();
 
-	let activeView: ChatMessageStatsView = $state(initialView);
+	let activeView: ChatMessageStatsView = $derived(initialView);
 	let hasAutoSwitchedToGeneration = $state(false);
 
 	// In live mode: auto-switch to GENERATION tab when prompt processing completes

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

@@ -35,6 +35,7 @@
 	import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte';
 	import { isFileTypeSupported, filterFilesByModalities } from '$lib/utils';
 	import { parseFilesToMessageExtras, processFilesToChatUploaded } from '$lib/utils/browser-only';
+	import { ErrorDialogType } from '$lib/enums';
 	import { onMount } from 'svelte';
 	import { fade, fly, slide } from 'svelte/transition';
 	import { Trash2, AlertTriangle, RefreshCw } from '@lucide/svelte';
@@ -616,7 +617,7 @@
 	contextInfo={activeErrorDialog?.contextInfo}
 	onOpenChange={handleErrorDialogOpenChange}
 	open={Boolean(activeErrorDialog)}
-	type={activeErrorDialog?.type ?? 'server'}
+	type={(activeErrorDialog?.type as ErrorDialogType) ?? ErrorDialogType.SERVER}
 />
 
 <style>

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte

@@ -0,0 +1,47 @@
+<script lang="ts">
+	import ChatForm from '$lib/components/app/chat/ChatForm/ChatForm.svelte';
+
+	interface Props {
+		class?: string;
+		disabled?: boolean;
+		initialMessage?: string;
+		isLoading?: boolean;
+		onFileRemove?: (fileId: string) => void;
+		onFileUpload?: (files: File[]) => void;
+		onSend?: (message: string, files?: ChatUploadedFile[]) => Promise<boolean>;
+		onStop?: () => void;
+		onSystemPromptAdd?: (draft: { message: string; files: ChatUploadedFile[] }) => void;
+		showHelperText?: boolean;
+		uploadedFiles?: ChatUploadedFile[];
+	}
+
+	let {
+		class: className,
+		disabled = false,
+		initialMessage = '',
+		isLoading = false,
+		onFileRemove,
+		onFileUpload,
+		onSend,
+		onStop,
+		onSystemPromptAdd,
+		showHelperText = true,
+		uploadedFiles = $bindable([])
+	}: Props = $props();
+</script>
+
+<div class="relative mx-auto max-w-[48rem]">
+	<ChatForm
+		class={className}
+		{disabled}
+		{initialMessage}
+		{isLoading}
+		{onFileRemove}
+		{onFileUpload}
+		{onSend}
+		{onStop}
+		{onSystemPromptAdd}
+		{showHelperText}
+		bind:uploadedFiles
+	/>
+</div>

tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettings.svelte

@@ -18,19 +18,24 @@
 	} from '$lib/components/app';
 	import { ScrollArea } from '$lib/components/ui/scroll-area';
 	import { config, settingsStore } from '$lib/stores/settings.svelte';
+	import {
+		SETTINGS_SECTION_TITLES,
+		type SettingsSectionTitle
+	} from '$lib/constants/settings-sections';
 	import { setMode } from 'mode-watcher';
 	import type { Component } from 'svelte';
 
 	interface Props {
 		onSave?: () => void;
+		initialSection?: SettingsSectionTitle;
 	}
 
-	let { onSave }: Props = $props();
+	let { onSave, initialSection }: Props = $props();
 
 	const settingSections: Array<{
 		fields: SettingsFieldConfig[];
 		icon: Component;
-		title: string;
+		title: SettingsSectionTitle;
 	}> = [
 		{
 			title: 'General',
@@ -285,7 +290,9 @@
 		// }
 	];
 
-	let activeSection = $state('General');
+	let activeSection = $derived<SettingsSectionTitle>(
+		initialSection ?? SETTINGS_SECTION_TITLES.GENERAL
+	);
 	let currentSection = $derived(
 		settingSections.find((section) => section.title === activeSection) || settingSections[0]
 	);
@@ -295,6 +302,16 @@
 	let canScrollRight = $state(false);
 	let scrollContainer: HTMLDivElement | undefined = $state();
 
+	$effect(() => {
+		if (!initialSection) {
+			return;
+		}
+
+		if (settingSections.some((section) => section.title === initialSection)) {
+			activeSection = initialSection;
+		}
+	});
+
 	function handleThemeChange(newTheme: string) {
 		localConfig.theme = newTheme;
 

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebarConversationItem.svelte

@@ -142,7 +142,7 @@
 					{
 						icon: Download,
 						label: 'Export',
-						onclick: (e) => {
+						onclick: (e: Event) => {
 							e.stopPropagation();
 							conversationsStore.downloadConversation(conversation.id);
 						},

tools/server/webui/src/lib/components/app/content/MarkdownContent.svelte

@@ -15,6 +15,7 @@
 	import { rehypeRestoreTableHtml } from '$lib/markdown/table-html-restorer';
 	import { rehypeEnhanceLinks } from '$lib/markdown/enhance-links';
 	import { rehypeEnhanceCodeBlocks } from '$lib/markdown/enhance-code-blocks';
+	import { rehypeResolveAttachmentImages } from '$lib/markdown/resolve-attachment-images';
 	import { remarkLiteralHtml } from '$lib/markdown/literal-html';
 	import { copyCodeToClipboard, preprocessLaTeX, getImageErrorFallbackHtml } from '$lib/utils';
 	import {
@@ -23,6 +24,7 @@
 		DATA_ERROR_HANDLED_ATTR,
 		BOOL_TRUE_STRING
 	} from '$lib/constants/markdown';
+	import { UrlPrefix } from '$lib/enums';
 	import { FileTypeText } from '$lib/enums/files';
 	import {
 		highlightCode,
@@ -33,8 +35,7 @@
 	import githubDarkCss from 'highlight.js/styles/github-dark.css?inline';
 	import githubLightCss from 'highlight.js/styles/github.css?inline';
 	import { mode } from 'mode-watcher';
-	import ActionIconsCodeBlock from '$lib/components/app/actions/ActionIconsCodeBlock.svelte';
-	import DialogCodePreview from '$lib/components/app/misc/CodePreviewDialog.svelte';
+	import { ActionIconsCodeBlock, DialogCodePreview } from '$lib/components/app';
 	import { createAutoScrollController } from '$lib/hooks/use-auto-scroll.svelte';
 	import type { DatabaseMessageExtra } from '$lib/types/database';
 
@@ -100,6 +101,7 @@
 			.use(rehypeRestoreTableHtml) // Restore limited HTML (e.g., <br>, <ul>) inside Markdown tables
 			.use(rehypeEnhanceLinks) // Add target="_blank" to links
 			.use(rehypeEnhanceCodeBlocks) // Wrap code blocks with header and actions
+			.use(rehypeResolveAttachmentImages, { attachments })
 			.use(rehypeStringify, { allowDangerousHtml: true }); // Convert to HTML string
 	});
 
@@ -500,7 +502,10 @@
 		if (!img || !img.src) return;
 
 		// Don't handle data URLs or already-handled images
-		if (img.src.startsWith('data:') || img.dataset[DATA_ERROR_HANDLED_ATTR] === BOOL_TRUE_STRING)
+		if (
+			img.src.startsWith(UrlPrefix.DATA) ||
+			img.dataset[DATA_ERROR_HANDLED_ATTR] === BOOL_TRUE_STRING
+		)
 			return;
 		img.dataset[DATA_ERROR_HANDLED_ATTR] = BOOL_TRUE_STRING;
 

tools/server/webui/src/lib/components/app/dialogs/DialogChatError.svelte

@@ -1,10 +1,11 @@
 <script lang="ts">
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
 	import { AlertTriangle, TimerOff } from '@lucide/svelte';
+	import { ErrorDialogType } from '$lib/enums';
 
 	interface Props {
 		open: boolean;
-		type: 'timeout' | 'server';
+		type: ErrorDialogType;
 		message: string;
 		contextInfo?: { n_prompt_tokens: number; n_ctx: number };
 		onOpenChange?: (open: boolean) => void;
@@ -12,7 +13,7 @@
 
 	let { open = $bindable(), type, message, contextInfo, onOpenChange }: Props = $props();
 
-	const isTimeout = $derived(type === 'timeout');
+	const isTimeout = $derived(type === ErrorDialogType.TIMEOUT);
 	const title = $derived(isTimeout ? 'TCP Timeout' : 'Server Error');
 	const description = $derived(
 		isTimeout
@@ -58,7 +59,12 @@
 						<span class="font-medium">Prompt tokens:</span>
 						{contextInfo.n_prompt_tokens.toLocaleString()}
 					</p>
-					<p><span class="font-medium">Context size:</span> {contextInfo.n_ctx.toLocaleString()}</p>
+					{#if contextInfo.n_ctx}
+						<p>
+							<span class="font-medium">Context size:</span>
+							{contextInfo.n_ctx.toLocaleString()}
+						</p>
+					{/if}
 				</div>
 			{/if}
 		</div>

tools/server/webui/src/lib/components/app/dialogs/DialogChatSettings.svelte

@@ -1,13 +1,15 @@
 <script lang="ts">
 	import * as Dialog from '$lib/components/ui/dialog';
 	import { ChatSettings } from '$lib/components/app';
+	import type { SettingsSectionTitle } from '$lib/constants/settings-sections';
 
 	interface Props {
 		onOpenChange?: (open: boolean) => void;
 		open?: boolean;
+		initialSection?: SettingsSectionTitle;
 	}
 
-	let { onOpenChange, open = false }: Props = $props();
+	let { onOpenChange, open = false, initialSection }: Props = $props();
 
 	let chatSettingsRef: ChatSettings | undefined = $state();
 
@@ -28,10 +30,9 @@
 
 <Dialog.Root {open} onOpenChange={handleClose}>
 	<Dialog.Content
-		class="z-999999 flex h-[100dvh] max-h-[100dvh] min-h-[100dvh] flex-col gap-0 rounded-none p-0
-			md:h-[64vh] md:max-h-[64vh] md:min-h-0 md:rounded-lg"
-		style="max-width: 48rem;"
+		class="z-999999 flex h-[100dvh] max-h-[100dvh] min-h-[100dvh] max-w-4xl! flex-col gap-0 rounded-none
+			p-0 md:h-[64vh] md:max-h-[64vh] md:min-h-0 md:rounded-lg"
 	>
-		<ChatSettings bind:this={chatSettingsRef} onSave={handleSave} />
+		<ChatSettings bind:this={chatSettingsRef} onSave={handleSave} {initialSection} />
 	</Dialog.Content>
 </Dialog.Root>

tools/server/webui/src/lib/components/app/dialogs/DialogCodePreview.svelte

@@ -37,7 +37,7 @@
 			<iframe
 				bind:this={iframeRef}
 				title="Preview {language}"
-				sandbox="allow-scripts"
+				sandbox="allow-scripts allow-same-origin"
 				class="code-preview-iframe"
 			></iframe>
 

tools/server/webui/src/lib/components/app/dialogs/DialogConfirmation.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
 	import type { Component } from 'svelte';
+	import { KeyboardKey } from '$lib/enums';
 
 	interface Props {
 		open: boolean;
@@ -29,7 +30,7 @@
 	}: Props = $props();
 
 	function handleKeydown(event: KeyboardEvent) {
-		if (event.key === 'Enter') {
+		if (event.key === KeyboardKey.ENTER) {
 			event.preventDefault();
 			onConfirm();
 		}

tools/server/webui/src/lib/components/app/dialogs/DialogModelInformation.svelte

@@ -1,7 +1,7 @@
 <script lang="ts">
 	import * as Dialog from '$lib/components/ui/dialog';
 	import * as Table from '$lib/components/ui/table';
-	import { BadgeModality, CopyToClipboardIcon } from '$lib/components/app';
+	import { BadgeModality, ActionIconCopyToClipboard } from '$lib/components/app';
 	import { serverStore } from '$lib/stores/server.svelte';
 	import { modelsStore, modelOptions, modelsLoading } from '$lib/stores/models.svelte';
 	import { formatFileSize, formatParameters, formatNumber } from '$lib/utils';
@@ -47,6 +47,7 @@
 
 		<Dialog.Header>
 			<Dialog.Title>Model Information</Dialog.Title>
+
 			<Dialog.Description>Current model details and capabilities</Dialog.Description>
 		</Dialog.Header>
 
@@ -73,7 +74,7 @@
 											{modelName}
 										</span>
 
-										<CopyToClipboardIcon
+										<ActionIconCopyToClipboard
 											text={modelName || ''}
 											canCopy={!!modelName}
 											ariaLabel="Copy model name to clipboard"
@@ -97,7 +98,7 @@
 										{serverProps.model_path}
 									</span>
 
-									<CopyToClipboardIcon
+									<ActionIconCopyToClipboard
 										text={serverProps.model_path}
 										ariaLabel="Copy model path to clipboard"
 									/>
@@ -105,17 +106,29 @@
 							</Table.Row>
 
 							<!-- Context Size -->
-							<Table.Row>
-								<Table.Cell class="h-10 align-middle font-medium">Context Size</Table.Cell>
-								<Table.Cell
-									>{formatNumber(serverProps.default_generation_settings.n_ctx)} tokens</Table.Cell
-								>
-							</Table.Row>
+							{#if serverProps?.default_generation_settings?.n_ctx}
+								<Table.Row>
+									<Table.Cell class="h-10 align-middle font-medium">Context Size</Table.Cell>
+
+									<Table.Cell
+										>{formatNumber(serverProps.default_generation_settings.n_ctx)} tokens</Table.Cell
+									>
+								</Table.Row>
+							{:else}
+								<Table.Row>
+									<Table.Cell class="h-10 align-middle font-medium text-red-500"
+										>Context Size</Table.Cell
+									>
+
+									<Table.Cell class="text-red-500">Not available</Table.Cell>
+								</Table.Row>
+							{/if}
 
 							<!-- Training Context -->
 							{#if modelMeta?.n_ctx_train}
 								<Table.Row>
 									<Table.Cell class="h-10 align-middle font-medium">Training Context</Table.Cell>
+
 									<Table.Cell>{formatNumber(modelMeta.n_ctx_train)} tokens</Table.Cell>
 								</Table.Row>
 							{/if}
@@ -124,6 +137,7 @@
 							{#if modelMeta?.size}
 								<Table.Row>
 									<Table.Cell class="h-10 align-middle font-medium">Model Size</Table.Cell>
+
 									<Table.Cell>{formatFileSize(modelMeta.size)}</Table.Cell>
 								</Table.Row>
 							{/if}
@@ -132,6 +146,7 @@
 							{#if modelMeta?.n_params}
 								<Table.Row>
 									<Table.Cell class="h-10 align-middle font-medium">Parameters</Table.Cell>
+
 									<Table.Cell>{formatParameters(modelMeta.n_params)}</Table.Cell>
 								</Table.Row>
 							{/if}
@@ -140,6 +155,7 @@
 							{#if modelMeta?.n_embd}
 								<Table.Row>
 									<Table.Cell class="align-middle font-medium">Embedding Size</Table.Cell>
+
 									<Table.Cell>{formatNumber(modelMeta.n_embd)}</Table.Cell>
 								</Table.Row>
 							{/if}
@@ -148,6 +164,7 @@
 							{#if modelMeta?.n_vocab}
 								<Table.Row>
 									<Table.Cell class="align-middle font-medium">Vocabulary Size</Table.Cell>
+
 									<Table.Cell>{formatNumber(modelMeta.n_vocab)} tokens</Table.Cell>
 								</Table.Row>
 							{/if}
@@ -163,6 +180,7 @@
 							<!-- Total Slots -->
 							<Table.Row>
 								<Table.Cell class="align-middle font-medium">Parallel Slots</Table.Cell>
+
 								<Table.Cell>{serverProps.total_slots}</Table.Cell>
 							</Table.Row>
 
@@ -170,6 +188,7 @@
 							{#if modalities.length > 0}
 								<Table.Row>
 									<Table.Cell class="align-middle font-medium">Modalities</Table.Cell>
+
 									<Table.Cell>
 										<div class="flex flex-wrap gap-1">
 											<BadgeModality {modalities} />
@@ -181,6 +200,7 @@
 							<!-- Build Info -->
 							<Table.Row>
 								<Table.Cell class="align-middle font-medium">Build Info</Table.Cell>
+
 								<Table.Cell class="align-middle font-mono text-xs"
 									>{serverProps.build_info}</Table.Cell
 								>
@@ -190,6 +210,7 @@
 							{#if serverProps.chat_template}
 								<Table.Row>
 									<Table.Cell class="align-middle font-medium">Chat Template</Table.Cell>
+
 									<Table.Cell class="py-10">
 										<div class="max-h-120 overflow-y-auto rounded-md bg-muted p-4">
 											<pre

tools/server/webui/src/lib/components/app/forms/KeyValuePairs.svelte

@@ -0,0 +1,110 @@
+<script lang="ts">
+	import { Plus, Trash2 } from '@lucide/svelte';
+	import { Input } from '$lib/components/ui/input';
+	import { autoResizeTextarea } from '$lib/utils';
+	import type { KeyValuePair } from '$lib/types';
+
+	interface Props {
+		class?: string;
+		pairs: KeyValuePair[];
+		onPairsChange: (pairs: KeyValuePair[]) => void;
+		keyPlaceholder?: string;
+		valuePlaceholder?: string;
+		addButtonLabel?: string;
+		emptyMessage?: string;
+		sectionLabel?: string;
+		sectionLabelOptional?: boolean;
+	}
+
+	let {
+		class: className = '',
+		pairs,
+		onPairsChange,
+		keyPlaceholder = 'Key',
+		valuePlaceholder = 'Value',
+		addButtonLabel = 'Add',
+		emptyMessage = 'No items configured.',
+		sectionLabel,
+		sectionLabelOptional = true
+	}: Props = $props();
+
+	function addPair() {
+		onPairsChange([...pairs, { key: '', value: '' }]);
+	}
+
+	function removePair(index: number) {
+		onPairsChange(pairs.filter((_, i) => i !== index));
+	}
+
+	function updatePairKey(index: number, key: string) {
+		const newPairs = [...pairs];
+		newPairs[index] = { ...newPairs[index], key };
+		onPairsChange(newPairs);
+	}
+
+	function updatePairValue(index: number, value: string) {
+		const newPairs = [...pairs];
+		newPairs[index] = { ...newPairs[index], value };
+		onPairsChange(newPairs);
+	}
+</script>
+
+<div class={className}>
+	<div class="mb-2 flex items-center justify-between">
+		{#if sectionLabel}
+			<span class="text-xs font-medium">
+				{sectionLabel}
+				{#if sectionLabelOptional}
+					<span class="text-muted-foreground">(optional)</span>
+				{/if}
+			</span>
+		{/if}
+
+		<button
+			type="button"
+			class="inline-flex cursor-pointer items-center gap-1 rounded-md px-1.5 py-1 text-xs text-muted-foreground hover:bg-muted hover:text-foreground"
+			onclick={addPair}
+		>
+			<Plus class="h-3 w-3" />
+			{addButtonLabel}
+		</button>
+	</div>
+	{#if pairs.length > 0}
+		<div class="space-y-3">
+			{#each pairs as pair, index (index)}
+				<div class="flex items-start gap-2">
+					<Input
+						type="text"
+						placeholder={keyPlaceholder}
+						value={pair.key}
+						oninput={(e) => updatePairKey(index, e.currentTarget.value)}
+						class="flex-1"
+					/>
+
+					<textarea
+						use:autoResizeTextarea
+						placeholder={valuePlaceholder}
+						value={pair.value}
+						oninput={(e) => {
+							updatePairValue(index, e.currentTarget.value);
+							autoResizeTextarea(e.currentTarget);
+						}}
+						class="flex-1 resize-none rounded-md border border-input bg-transparent px-3 py-2 text-sm leading-5 placeholder:text-muted-foreground focus-visible:ring-1 focus-visible:ring-ring focus-visible:outline-none"
+						rows="1"
+					></textarea>
+
+					<button
+						type="button"
+						class="mt-1.5 shrink-0 cursor-pointer rounded-md p-1 text-muted-foreground hover:bg-destructive/10 hover:text-destructive"
+						onclick={() => removePair(index)}
+						aria-label="Remove item"
+					>
+						<Trash2 class="h-3.5 w-3.5" />
+					</button>
+				</div>
+			{/each}
+		</div>
+	{:else}
+		<p class="text-xs text-muted-foreground">{emptyMessage}</p>
+	{/if}
+</div>

tools/server/webui/src/lib/components/app/forms/SearchInput.svelte

@@ -46,7 +46,7 @@
 
 <div class="relative {className}">
 	<Search
-		class="absolute top-1/2 left-3 h-4 w-4 -translate-y-1/2 transform text-muted-foreground"
+		class="absolute top-1/2 left-3 z-10 h-4 w-4 -translate-y-1/2 transform text-muted-foreground"
 	/>
 
 	<Input

tools/server/webui/src/lib/components/app/forms/index.ts

@@ -0,0 +1,30 @@
+/**
+ *
+ * FORMS & INPUTS
+ *
+ * Form-related utility components.
+ *
+ */
+
+/**
+ * **SearchInput** - Search field with clear button
+ *
+ * Input field optimized for search with clear button and keyboard handling.
+ * Supports placeholder, autofocus, and change callbacks.
+ */
+export { default as SearchInput } from './SearchInput.svelte';
+
+/**
+ * **KeyValuePairs** - Editable key-value list
+ *
+ * Dynamic list of key-value pairs with add/remove functionality.
+ * Used for HTTP headers, metadata, and configuration.
+ *
+ * **Features:**
+ * - Add new pairs with button
+ * - Remove individual pairs
+ * - Customizable placeholders and labels
+ * - Empty state message
+ * - Auto-resize value textarea
+ */
+export { default as KeyValuePairs } from './KeyValuePairs.svelte';

tools/server/webui/src/lib/components/app/index.ts

@@ -1,12 +1,20 @@
-// Chat
+export * from './actions';
+export * from './badges';
+export * from './content';
+export * from './forms';
+export * from './misc';
+export * from './models';
+export * from './navigation';
+export * from './server';
 
+// Chat
 export { default as ChatAttachmentPreview } from './chat/ChatAttachments/ChatAttachmentPreview.svelte';
 export { default as ChatAttachmentThumbnailFile } from './chat/ChatAttachments/ChatAttachmentThumbnailFile.svelte';
 export { default as ChatAttachmentThumbnailImage } from './chat/ChatAttachments/ChatAttachmentThumbnailImage.svelte';
 export { default as ChatAttachmentsList } from './chat/ChatAttachments/ChatAttachmentsList.svelte';
 export { default as ChatAttachmentsViewAll } from './chat/ChatAttachments/ChatAttachmentsViewAll.svelte';
-
 export { default as ChatForm } from './chat/ChatForm/ChatForm.svelte';
+export { default as ChatFormActionAttachmentsDropdown } from './chat/ChatForm/ChatFormActions/ChatFormActionAttachmentsDropdown.svelte';
 export { default as ChatFormActionFileAttachments } from './chat/ChatForm/ChatFormActions/ChatFormActionFileAttachments.svelte';
 export { default as ChatFormActionRecord } from './chat/ChatForm/ChatFormActions/ChatFormActionRecord.svelte';
 export { default as ChatFormActions } from './chat/ChatForm/ChatFormActions/ChatFormActions.svelte';
@@ -14,36 +22,38 @@ export { default as ChatFormActionSubmit } from './chat/ChatForm/ChatFormActions
 export { default as ChatFormFileInputInvisible } from './chat/ChatForm/ChatFormFileInputInvisible.svelte';
 export { default as ChatFormHelperText } from './chat/ChatForm/ChatFormHelperText.svelte';
 export { default as ChatFormTextarea } from './chat/ChatForm/ChatFormTextarea.svelte';
-
 export { default as ChatMessage } from './chat/ChatMessages/ChatMessage.svelte';
 export { default as ChatMessageActions } from './chat/ChatMessages/ChatMessageActions.svelte';
+export { default as ChatMessageAssistant } from './chat/ChatMessages/ChatMessageAssistant.svelte';
 export { default as ChatMessageBranchingControls } from './chat/ChatMessages/ChatMessageBranchingControls.svelte';
+export { default as ChatMessageEditForm } from './chat/ChatMessages/ChatMessageEditForm.svelte';
 export { default as ChatMessageStatistics } from './chat/ChatMessages/ChatMessageStatistics.svelte';
 export { default as ChatMessageSystem } from './chat/ChatMessages/ChatMessageSystem.svelte';
 export { default as ChatMessageThinkingBlock } from './chat/ChatMessages/ChatMessageThinkingBlock.svelte';
+export { default as ChatMessageUser } from './chat/ChatMessages/ChatMessageUser.svelte';
 export { default as ChatMessages } from './chat/ChatMessages/ChatMessages.svelte';
 export { default as MessageBranchingControls } from './chat/ChatMessages/ChatMessageBranchingControls.svelte';
-
 export { default as ChatScreen } from './chat/ChatScreen/ChatScreen.svelte';
+export { default as ChatScreenDragOverlay } from './chat/ChatScreen/ChatScreenDragOverlay.svelte';
+export { default as ChatScreenForm } from './chat/ChatScreen/ChatScreenForm.svelte';
 export { default as ChatScreenHeader } from './chat/ChatScreen/ChatScreenHeader.svelte';
 export { default as ChatScreenProcessingInfo } from './chat/ChatScreen/ChatScreenProcessingInfo.svelte';
-
 export { default as ChatSettings } from './chat/ChatSettings/ChatSettings.svelte';
 export { default as ChatSettingsFooter } from './chat/ChatSettings/ChatSettingsFooter.svelte';
 export { default as ChatSettingsFields } from './chat/ChatSettings/ChatSettingsFields.svelte';
 export { default as ChatSettingsImportExportTab } from './chat/ChatSettings/ChatSettingsImportExportTab.svelte';
 export { default as ChatSettingsParameterSourceIndicator } from './chat/ChatSettings/ChatSettingsParameterSourceIndicator.svelte';
-
 export { default as ChatSidebar } from './chat/ChatSidebar/ChatSidebar.svelte';
+export { default as ChatSidebarActions } from './chat/ChatSidebar/ChatSidebarActions.svelte';
 export { default as ChatSidebarConversationItem } from './chat/ChatSidebar/ChatSidebarConversationItem.svelte';
 export { default as ChatSidebarSearch } from './chat/ChatSidebar/ChatSidebarSearch.svelte';
 
 // Dialogs
-
 export { default as DialogChatAttachmentPreview } from './dialogs/DialogChatAttachmentPreview.svelte';
 export { default as DialogChatAttachmentsViewAll } from './dialogs/DialogChatAttachmentsViewAll.svelte';
 export { default as DialogChatError } from './dialogs/DialogChatError.svelte';
 export { default as DialogChatSettings } from './dialogs/DialogChatSettings.svelte';
+export { default as DialogCodePreview } from './dialogs/DialogCodePreview.svelte';
 export { default as DialogConfirmation } from './dialogs/DialogConfirmation.svelte';
 export { default as DialogConversationSelection } from './dialogs/DialogConversationSelection.svelte';
 export { default as DialogConversationTitleUpdate } from './dialogs/DialogConversationTitleUpdate.svelte';
@@ -51,25 +61,8 @@ export { default as DialogEmptyFileAlert } from './dialogs/DialogEmptyFileAlert.
 export { default as DialogModelInformation } from './dialogs/DialogModelInformation.svelte';
 export { default as DialogModelNotAvailable } from './dialogs/DialogModelNotAvailable.svelte';
 
-// Miscellanous
-
-export { default as ActionButton } from './misc/ActionButton.svelte';
-export { default as ActionDropdown } from './misc/ActionDropdown.svelte';
-export { default as BadgeChatStatistic } from './misc/BadgeChatStatistic.svelte';
-export { default as BadgeInfo } from './misc/BadgeInfo.svelte';
-export { default as ModelBadge } from './models/ModelBadge.svelte';
-export { default as BadgeModality } from './misc/BadgeModality.svelte';
-export { default as ConversationSelection } from './misc/ConversationSelection.svelte';
-export { default as CopyToClipboardIcon } from './misc/CopyToClipboardIcon.svelte';
-export { default as KeyboardShortcutInfo } from './misc/KeyboardShortcutInfo.svelte';
-export { default as MarkdownContent } from './misc/MarkdownContent.svelte';
-export { default as RemoveButton } from './misc/RemoveButton.svelte';
-export { default as SearchInput } from './misc/SearchInput.svelte';
-export { default as SyntaxHighlightedCode } from './misc/SyntaxHighlightedCode.svelte';
-export { default as ModelsSelector } from './models/ModelsSelector.svelte';
-
-// Server
-
-export { default as ServerStatus } from './server/ServerStatus.svelte';
-export { default as ServerErrorSplash } from './server/ServerErrorSplash.svelte';
-export { default as ServerLoadingSplash } from './server/ServerLoadingSplash.svelte';
+// Compatibility aliases
+export { default as ActionButton } from './actions/ActionIcon.svelte';
+export { default as ActionDropdown } from './navigation/DropdownMenuActions.svelte';
+export { default as CopyToClipboardIcon } from './actions/ActionIconCopyToClipboard.svelte';
+export { default as RemoveButton } from './actions/ActionIconRemove.svelte';

tools/server/webui/src/lib/components/app/misc/ActionButton.svelte

@@ -1,47 +0,0 @@
-<script lang="ts">
-	import { Button } from '$lib/components/ui/button';
-	import * as Tooltip from '$lib/components/ui/tooltip';
-	import type { Component } from 'svelte';
-
-	interface Props {
-		icon: Component;
-		tooltip: string;
-		variant?: 'default' | 'destructive' | 'outline' | 'secondary' | 'ghost' | 'link';
-		size?: 'default' | 'sm' | 'lg' | 'icon';
-		class?: string;
-		disabled?: boolean;
-		onclick: () => void;
-		'aria-label'?: string;
-	}
-
-	let {
-		icon,
-		tooltip,
-		variant = 'ghost',
-		size = 'sm',
-		class: className = '',
-		disabled = false,
-		onclick,
-		'aria-label': ariaLabel
-	}: Props = $props();
-</script>
-
-<Tooltip.Root>
-	<Tooltip.Trigger>
-		<Button
-			{variant}
-			{size}
-			{disabled}
-			{onclick}
-			class="h-6 w-6 p-0 {className} flex"
-			aria-label={ariaLabel || tooltip}
-		>
-			{@const IconComponent = icon}
-			<IconComponent class="h-3 w-3" />
-		</Button>
-	</Tooltip.Trigger>
-
-	<Tooltip.Content>
-		<p>{tooltip}</p>
-	</Tooltip.Content>
-</Tooltip.Root>

tools/server/webui/src/lib/components/app/misc/ActionDropdown.svelte

@@ -1,86 +0,0 @@
-<script lang="ts">
-	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
-	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { KeyboardShortcutInfo } from '$lib/components/app';
-	import type { Component } from 'svelte';
-
-	interface ActionItem {
-		icon: Component;
-		label: string;
-		onclick: (event: Event) => void;
-		variant?: 'default' | 'destructive';
-		disabled?: boolean;
-		shortcut?: string[];
-		separator?: boolean;
-	}
-
-	interface Props {
-		triggerIcon: Component;
-		triggerTooltip?: string;
-		triggerClass?: string;
-		actions: ActionItem[];
-		align?: 'start' | 'center' | 'end';
-		open?: boolean;
-	}
-
-	let {
-		triggerIcon,
-		triggerTooltip,
-		triggerClass = '',
-		actions,
-		align = 'end',
-		open = $bindable(false)
-	}: Props = $props();
-</script>
-
-<DropdownMenu.Root bind:open>
-	<DropdownMenu.Trigger
-		class="flex h-6 w-6 cursor-pointer items-center justify-center rounded-md p-0 text-sm font-medium transition-colors hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground focus:outline-none disabled:pointer-events-none disabled:opacity-50 data-[state=open]:bg-accent data-[state=open]:text-accent-foreground {triggerClass}"
-		onclick={(e) => e.stopPropagation()}
-	>
-		{#if triggerTooltip}
-			<Tooltip.Root>
-				<Tooltip.Trigger>
-					{@render iconComponent(triggerIcon, 'h-3 w-3')}
-					<span class="sr-only">{triggerTooltip}</span>
-				</Tooltip.Trigger>
-				<Tooltip.Content>
-					<p>{triggerTooltip}</p>
-				</Tooltip.Content>
-			</Tooltip.Root>
-		{:else}
-			{@render iconComponent(triggerIcon, 'h-3 w-3')}
-		{/if}
-	</DropdownMenu.Trigger>
-
-	<DropdownMenu.Content {align} class="z-[999999] w-48">
-		{#each actions as action, index (action.label)}
-			{#if action.separator && index > 0}
-				<DropdownMenu.Separator />
-			{/if}
-
-			<DropdownMenu.Item
-				onclick={action.onclick}
-				variant={action.variant}
-				disabled={action.disabled}
-				class="flex items-center justify-between hover:[&>kbd]:opacity-100"
-			>
-				<div class="flex items-center gap-2">
-					{@render iconComponent(
-						action.icon,
-						`h-4 w-4 ${action.variant === 'destructive' ? 'text-destructive' : ''}`
-					)}
-					{action.label}
-				</div>
-
-				{#if action.shortcut}
-					<KeyboardShortcutInfo keys={action.shortcut} variant={action.variant} />
-				{/if}
-			</DropdownMenu.Item>
-		{/each}
-	</DropdownMenu.Content>
-</DropdownMenu.Root>
-
-{#snippet iconComponent(IconComponent: Component, className: string)}
-	<IconComponent class={className} />
-{/snippet}

tools/server/webui/src/lib/components/app/misc/BadgeChatStatistic.svelte

@@ -1,44 +0,0 @@
-<script lang="ts">
-	import { BadgeInfo } from '$lib/components/app';
-	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { copyToClipboard } from '$lib/utils';
-	import type { Component } from 'svelte';
-
-	interface Props {
-		class?: string;
-		icon: Component;
-		value: string | number;
-		tooltipLabel?: string;
-	}
-
-	let { class: className = '', icon: Icon, value, tooltipLabel }: Props = $props();
-
-	function handleClick() {
-		void copyToClipboard(String(value));
-	}
-</script>
-
-{#if tooltipLabel}
-	<Tooltip.Root>
-		<Tooltip.Trigger>
-			<BadgeInfo class={className} onclick={handleClick}>
-				{#snippet icon()}
-					<Icon class="h-3 w-3" />
-				{/snippet}
-
-				{value}
-			</BadgeInfo>
-		</Tooltip.Trigger>
-		<Tooltip.Content>
-			<p>{tooltipLabel}</p>
-		</Tooltip.Content>
-	</Tooltip.Root>
-{:else}
-	<BadgeInfo class={className} onclick={handleClick}>
-		{#snippet icon()}
-			<Icon class="h-3 w-3" />
-		{/snippet}
-
-		{value}
-	</BadgeInfo>
-{/if}

tools/server/webui/src/lib/components/app/misc/BadgeInfo.svelte

@@ -1,27 +0,0 @@
-<script lang="ts">
-	import { cn } from '$lib/components/ui/utils';
-	import type { Snippet } from 'svelte';
-
-	interface Props {
-		children: Snippet;
-		class?: string;
-		icon?: Snippet;
-		onclick?: () => void;
-	}
-
-	let { children, class: className = '', icon, onclick }: Props = $props();
-</script>
-
-<button
-	class={cn(
-		'inline-flex cursor-pointer items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75',
-		className
-	)}
-	{onclick}
->
-	{#if icon}
-		{@render icon()}
-	{/if}
-
-	{@render children()}
-</button>

tools/server/webui/src/lib/components/app/misc/BadgeModality.svelte

@@ -1,39 +0,0 @@
-<script lang="ts">
-	import { ModelModality } from '$lib/enums';
-	import { MODALITY_ICONS, MODALITY_LABELS } from '$lib/constants/icons';
-	import { cn } from '$lib/components/ui/utils';
-
-	type DisplayableModality = ModelModality.VISION | ModelModality.AUDIO;
-
-	interface Props {
-		modalities: ModelModality[];
-		class?: string;
-	}
-
-	let { modalities, class: className = '' }: Props = $props();
-
-	// Filter to only modalities that have icons (VISION, AUDIO)
-	const displayableModalities = $derived(
-		modalities.filter(
-			(m): m is DisplayableModality => m === ModelModality.VISION || m === ModelModality.AUDIO
-		)
-	);
-</script>
-
-{#each displayableModalities as modality, index (index)}
-	{@const IconComponent = MODALITY_ICONS[modality]}
-	{@const label = MODALITY_LABELS[modality]}
-
-	<span
-		class={cn(
-			'inline-flex items-center gap-1 rounded-md bg-muted px-2 py-1 text-xs font-medium',
-			className
-		)}
-	>
-		{#if IconComponent}
-			<IconComponent class="h-3 w-3" />
-		{/if}
-
-		{label}
-	</span>
-{/each}

tools/server/webui/src/lib/components/app/misc/CopyToClipboardIcon.svelte

@@ -1,18 +0,0 @@
-<script lang="ts">
-	import { Copy } from '@lucide/svelte';
-	import { copyToClipboard } from '$lib/utils';
-
-	interface Props {
-		ariaLabel?: string;
-		canCopy?: boolean;
-		text: string;
-	}
-
-	let { ariaLabel = 'Copy to clipboard', canCopy = true, text }: Props = $props();
-</script>
-
-<Copy
-	class="h-3 w-3 flex-shrink-0 cursor-{canCopy ? 'pointer' : 'not-allowed'}"
-	aria-label={ariaLabel}
-	onclick={() => canCopy && copyToClipboard(text)}
-/>

tools/server/webui/src/lib/components/app/misc/DropdownMenuSearchable.svelte

@@ -1,88 +0,0 @@
-<script lang="ts">
-	import type { Snippet } from 'svelte';
-	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
-	import { cn } from '$lib/components/ui/utils';
-	import { SearchInput } from '$lib/components/app';
-
-	interface Props {
-		open?: boolean;
-		onOpenChange?: (open: boolean) => void;
-		placeholder?: string;
-		searchValue?: string;
-		onSearchChange?: (value: string) => void;
-		onSearchKeyDown?: (event: KeyboardEvent) => void;
-		align?: 'start' | 'center' | 'end';
-		contentClass?: string;
-		emptyMessage?: string;
-		isEmpty?: boolean;
-		disabled?: boolean;
-		trigger: Snippet;
-		children: Snippet;
-		footer?: Snippet;
-	}
-
-	let {
-		open = $bindable(false),
-		onOpenChange,
-		placeholder = 'Search...',
-		searchValue = $bindable(''),
-		onSearchChange,
-		onSearchKeyDown,
-		align = 'start',
-		contentClass = 'w-72',
-		emptyMessage = 'No items found',
-		isEmpty = false,
-		disabled = false,
-		trigger,
-		children,
-		footer
-	}: Props = $props();
-
-	function handleOpenChange(newOpen: boolean) {
-		open = newOpen;
-
-		if (!newOpen) {
-			searchValue = '';
-			onSearchChange?.('');
-		}
-
-		onOpenChange?.(newOpen);
-	}
-</script>
-
-<DropdownMenu.Root bind:open onOpenChange={handleOpenChange}>
-	<DropdownMenu.Trigger
-		{disabled}
-		onclick={(e) => {
-			e.preventDefault();
-			e.stopPropagation();
-		}}
-	>
-		{@render trigger()}
-	</DropdownMenu.Trigger>
-
-	<DropdownMenu.Content {align} class={cn(contentClass, 'pt-0')}>
-		<div class="sticky top-0 z-10 mb-2 bg-popover p-1 pt-2">
-			<SearchInput
-				{placeholder}
-				bind:value={searchValue}
-				onInput={onSearchChange}
-				onKeyDown={onSearchKeyDown}
-			/>
-		</div>
-
-		<div class={cn('overflow-y-auto')}>
-			{@render children()}
-
-			{#if isEmpty}
-				<div class="px-2 py-3 text-center text-sm text-muted-foreground">{emptyMessage}</div>
-			{/if}
-		</div>
-
-		{#if footer}
-			<DropdownMenu.Separator />
-
-			{@render footer()}
-		{/if}
-	</DropdownMenu.Content>
-</DropdownMenu.Root>

tools/server/webui/src/lib/components/app/misc/MarkdownContent.svelte

@@ -1,872 +0,0 @@
-<script lang="ts">
-	import { remark } from 'remark';
-	import remarkBreaks from 'remark-breaks';
-	import remarkGfm from 'remark-gfm';
-	import remarkMath from 'remark-math';
-	import rehypeHighlight from 'rehype-highlight';
-	import remarkRehype from 'remark-rehype';
-	import rehypeKatex from 'rehype-katex';
-	import rehypeStringify from 'rehype-stringify';
-	import type { Root as HastRoot, RootContent as HastRootContent } from 'hast';
-	import type { Root as MdastRoot } from 'mdast';
-	import { browser } from '$app/environment';
-	import { onDestroy, tick } from 'svelte';
-	import { rehypeRestoreTableHtml } from '$lib/markdown/table-html-restorer';
-	import { rehypeEnhanceLinks } from '$lib/markdown/enhance-links';
-	import { rehypeEnhanceCodeBlocks } from '$lib/markdown/enhance-code-blocks';
-	import { remarkLiteralHtml } from '$lib/markdown/literal-html';
-	import { copyCodeToClipboard, preprocessLaTeX } from '$lib/utils';
-	import '$styles/katex-custom.scss';
-	import githubDarkCss from 'highlight.js/styles/github-dark.css?inline';
-	import githubLightCss from 'highlight.js/styles/github.css?inline';
-	import { mode } from 'mode-watcher';
-	import CodePreviewDialog from './CodePreviewDialog.svelte';
-
-	interface Props {
-		content: string;
-		class?: string;
-	}
-
-	interface MarkdownBlock {
-		id: string;
-		html: string;
-	}
-
-	let { content, class: className = '' }: Props = $props();
-
-	let containerRef = $state<HTMLDivElement>();
-	let renderedBlocks = $state<MarkdownBlock[]>([]);
-	let unstableBlockHtml = $state('');
-	let previewDialogOpen = $state(false);
-	let previewCode = $state('');
-	let previewLanguage = $state('text');
-
-	let pendingMarkdown: string | null = null;
-	let isProcessing = false;
-
-	const themeStyleId = `highlight-theme-${(window.idxThemeStyle = (window.idxThemeStyle ?? 0) + 1)}`;
-
-	let processor = $derived(() => {
-		return remark()
-			.use(remarkGfm) // GitHub Flavored Markdown
-			.use(remarkMath) // Parse $inline$ and $$block$$ math
-			.use(remarkBreaks) // Convert line breaks to <br>
-			.use(remarkLiteralHtml) // Treat raw HTML as literal text with preserved indentation
-			.use(remarkRehype) // Convert Markdown AST to rehype
-			.use(rehypeKatex) // Render math using KaTeX
-			.use(rehypeHighlight) // Add syntax highlighting
-			.use(rehypeRestoreTableHtml) // Restore limited HTML (e.g., <br>, <ul>) inside Markdown tables
-			.use(rehypeEnhanceLinks) // Add target="_blank" to links
-			.use(rehypeEnhanceCodeBlocks) // Wrap code blocks with header and actions
-			.use(rehypeStringify, { allowDangerousHtml: true }); // Convert to HTML string
-	});
-
-	/**
-	 * Removes click event listeners from copy and preview buttons.
-	 * Called on component destroy.
-	 */
-	function cleanupEventListeners() {
-		if (!containerRef) return;
-
-		const copyButtons = containerRef.querySelectorAll<HTMLButtonElement>('.copy-code-btn');
-		const previewButtons = containerRef.querySelectorAll<HTMLButtonElement>('.preview-code-btn');
-
-		for (const button of copyButtons) {
-			button.removeEventListener('click', handleCopyClick);
-		}
-
-		for (const button of previewButtons) {
-			button.removeEventListener('click', handlePreviewClick);
-		}
-	}
-
-	/**
-	 * Removes this component's highlight.js theme style from the document head.
-	 * Called on component destroy to clean up injected styles.
-	 */
-	function cleanupHighlightTheme() {
-		if (!browser) return;
-
-		const existingTheme = document.getElementById(themeStyleId);
-		existingTheme?.remove();
-	}
-
-	/**
-	 * Loads the appropriate highlight.js theme based on dark/light mode.
-	 * Injects a scoped style element into the document head.
-	 * @param isDark - Whether to load the dark theme (true) or light theme (false)
-	 */
-	function loadHighlightTheme(isDark: boolean) {
-		if (!browser) return;
-
-		const existingTheme = document.getElementById(themeStyleId);
-		existingTheme?.remove();
-
-		const style = document.createElement('style');
-		style.id = themeStyleId;
-		style.textContent = isDark ? githubDarkCss : githubLightCss;
-
-		document.head.appendChild(style);
-	}
-
-	/**
-	 * Extracts code information from a button click target within a code block.
-	 * @param target - The clicked button element
-	 * @returns Object with rawCode and language, or null if extraction fails
-	 */
-	function getCodeInfoFromTarget(target: HTMLElement) {
-		const wrapper = target.closest('.code-block-wrapper');
-
-		if (!wrapper) {
-			console.error('No wrapper found');
-			return null;
-		}
-
-		const codeElement = wrapper.querySelector<HTMLElement>('code[data-code-id]');
-
-		if (!codeElement) {
-			console.error('No code element found in wrapper');
-			return null;
-		}
-
-		const rawCode = codeElement.textContent ?? '';
-
-		const languageLabel = wrapper.querySelector<HTMLElement>('.code-language');
-		const language = languageLabel?.textContent?.trim() || 'text';
-
-		return { rawCode, language };
-	}
-
-	/**
-	 * Generates a unique identifier for a HAST node based on its position.
-	 * Used for stable block identification during incremental rendering.
-	 * @param node - The HAST root content node
-	 * @param indexFallback - Fallback index if position is unavailable
-	 * @returns Unique string identifier for the node
-	 */
-	function getHastNodeId(node: HastRootContent, indexFallback: number): string {
-		const position = node.position;
-
-		if (position?.start?.offset != null && position?.end?.offset != null) {
-			return `hast-${position.start.offset}-${position.end.offset}`;
-		}
-
-		return `${node.type}-${indexFallback}`;
-	}
-
-	/**
-	 * Handles click events on copy buttons within code blocks.
-	 * Copies the raw code content to the clipboard.
-	 * @param event - The click event from the copy button
-	 */
-	async function handleCopyClick(event: Event) {
-		event.preventDefault();
-		event.stopPropagation();
-
-		const target = event.currentTarget as HTMLButtonElement | null;
-
-		if (!target) {
-			return;
-		}
-
-		const info = getCodeInfoFromTarget(target);
-
-		if (!info) {
-			return;
-		}
-
-		try {
-			await copyCodeToClipboard(info.rawCode);
-		} catch (error) {
-			console.error('Failed to copy code:', error);
-		}
-	}
-
-	/**
-	 * Handles preview dialog open state changes.
-	 * Clears preview content when dialog is closed.
-	 * @param open - Whether the dialog is being opened or closed
-	 */
-	function handlePreviewDialogOpenChange(open: boolean) {
-		previewDialogOpen = open;
-
-		if (!open) {
-			previewCode = '';
-			previewLanguage = 'text';
-		}
-	}
-
-	/**
-	 * Handles click events on preview buttons within HTML code blocks.
-	 * Opens a preview dialog with the rendered HTML content.
-	 * @param event - The click event from the preview button
-	 */
-	function handlePreviewClick(event: Event) {
-		event.preventDefault();
-		event.stopPropagation();
-
-		const target = event.currentTarget as HTMLButtonElement | null;
-
-		if (!target) {
-			return;
-		}
-
-		const info = getCodeInfoFromTarget(target);
-
-		if (!info) {
-			return;
-		}
-
-		previewCode = info.rawCode;
-		previewLanguage = info.language;
-		previewDialogOpen = true;
-	}
-
-	/**
-	 * Processes markdown content into stable and unstable HTML blocks.
-	 * Uses incremental rendering: stable blocks are cached, unstable block is re-rendered.
-	 * @param markdown - The raw markdown string to process
-	 */
-	async function processMarkdown(markdown: string) {
-		if (!markdown) {
-			renderedBlocks = [];
-			unstableBlockHtml = '';
-			return;
-		}
-
-		const normalized = preprocessLaTeX(markdown);
-		const processorInstance = processor();
-		const ast = processorInstance.parse(normalized) as MdastRoot;
-		const processedRoot = (await processorInstance.run(ast)) as HastRoot;
-		const processedChildren = processedRoot.children ?? [];
-		const stableCount = Math.max(processedChildren.length - 1, 0);
-		const nextBlocks: MarkdownBlock[] = [];
-
-		for (let index = 0; index < stableCount; index++) {
-			const hastChild = processedChildren[index];
-			const id = getHastNodeId(hastChild, index);
-			const existing = renderedBlocks[index];
-
-			if (existing && existing.id === id) {
-				nextBlocks.push(existing);
-				continue;
-			}
-
-			const html = stringifyProcessedNode(
-				processorInstance,
-				processedRoot,
-				processedChildren[index]
-			);
-
-			nextBlocks.push({ id, html });
-		}
-
-		let unstableHtml = '';
-
-		if (processedChildren.length > stableCount) {
-			const unstableChild = processedChildren[stableCount];
-			unstableHtml = stringifyProcessedNode(processorInstance, processedRoot, unstableChild);
-		}
-
-		renderedBlocks = nextBlocks;
-		await tick(); // Force DOM sync before updating unstable HTML block
-		unstableBlockHtml = unstableHtml;
-	}
-
-	/**
-	 * Attaches click event listeners to copy and preview buttons in code blocks.
-	 * Uses data-listener-bound attribute to prevent duplicate bindings.
-	 */
-	function setupCodeBlockActions() {
-		if (!containerRef) return;
-
-		const wrappers = containerRef.querySelectorAll<HTMLElement>('.code-block-wrapper');
-
-		for (const wrapper of wrappers) {
-			const copyButton = wrapper.querySelector<HTMLButtonElement>('.copy-code-btn');
-			const previewButton = wrapper.querySelector<HTMLButtonElement>('.preview-code-btn');
-
-			if (copyButton && copyButton.dataset.listenerBound !== 'true') {
-				copyButton.dataset.listenerBound = 'true';
-				copyButton.addEventListener('click', handleCopyClick);
-			}
-
-			if (previewButton && previewButton.dataset.listenerBound !== 'true') {
-				previewButton.dataset.listenerBound = 'true';
-				previewButton.addEventListener('click', handlePreviewClick);
-			}
-		}
-	}
-
-	/**
-	 * Converts a single HAST node to an enhanced HTML string.
-	 * Applies link and code block enhancements to the output.
-	 * @param processorInstance - The remark/rehype processor instance
-	 * @param processedRoot - The full processed HAST root (for context)
-	 * @param child - The specific HAST child node to stringify
-	 * @returns Enhanced HTML string representation of the node
-	 */
-	function stringifyProcessedNode(
-		processorInstance: ReturnType<typeof processor>,
-		processedRoot: HastRoot,
-		child: unknown
-	) {
-		const root: HastRoot = {
-			...(processedRoot as HastRoot),
-			children: [child as never]
-		};
-
-		return processorInstance.stringify(root);
-	}
-
-	/**
-	 * Queues markdown for processing with coalescing support.
-	 * Only processes the latest markdown when multiple updates arrive quickly.
-	 * @param markdown - The markdown content to render
-	 */
-	async function updateRenderedBlocks(markdown: string) {
-		pendingMarkdown = markdown;
-
-		if (isProcessing) {
-			return;
-		}
-
-		isProcessing = true;
-
-		try {
-			while (pendingMarkdown !== null) {
-				const nextMarkdown = pendingMarkdown;
-				pendingMarkdown = null;
-
-				await processMarkdown(nextMarkdown);
-			}
-		} catch (error) {
-			console.error('Failed to process markdown:', error);
-			renderedBlocks = [];
-			unstableBlockHtml = markdown.replace(/\n/g, '<br>');
-		} finally {
-			isProcessing = false;
-		}
-	}
-
-	$effect(() => {
-		const currentMode = mode.current;
-		const isDark = currentMode === 'dark';
-
-		loadHighlightTheme(isDark);
-	});
-
-	$effect(() => {
-		updateRenderedBlocks(content);
-	});
-
-	$effect(() => {
-		const hasRenderedBlocks = renderedBlocks.length > 0;
-		const hasUnstableBlock = Boolean(unstableBlockHtml);
-
-		if ((hasRenderedBlocks || hasUnstableBlock) && containerRef) {
-			setupCodeBlockActions();
-		}
-	});
-
-	onDestroy(() => {
-		cleanupEventListeners();
-		cleanupHighlightTheme();
-	});
-</script>
-
-<div bind:this={containerRef} class={className}>
-	{#each renderedBlocks as block (block.id)}
-		<div class="markdown-block" data-block-id={block.id}>
-			<!-- eslint-disable-next-line no-at-html-tags -->
-			{@html block.html}
-		</div>
-	{/each}
-
-	{#if unstableBlockHtml}
-		<div class="markdown-block markdown-block--unstable" data-block-id="unstable">
-			<!-- eslint-disable-next-line no-at-html-tags -->
-			{@html unstableBlockHtml}
-		</div>
-	{/if}
-</div>
-
-<CodePreviewDialog
-	open={previewDialogOpen}
-	code={previewCode}
-	language={previewLanguage}
-	onOpenChange={handlePreviewDialogOpenChange}
-/>
-
-<style>
-	.markdown-block,
-	.markdown-block--unstable {
-		display: contents;
-	}
-
-	/* Base typography styles */
-	div :global(p:not(:last-child)) {
-		margin-bottom: 1rem;
-		line-height: 1.75;
-	}
-
-	div :global(:is(h1, h2, h3, h4, h5, h6):first-child) {
-		margin-top: 0;
-	}
-
-	/* Headers with consistent spacing */
-	div :global(h1) {
-		font-size: 1.875rem;
-		font-weight: 700;
-		line-height: 1.2;
-		margin: 1.5rem 0 0.75rem 0;
-	}
-
-	div :global(h2) {
-		font-size: 1.5rem;
-		font-weight: 600;
-		line-height: 1.3;
-		margin: 1.25rem 0 0.5rem 0;
-	}
-
-	div :global(h3) {
-		font-size: 1.25rem;
-		font-weight: 600;
-		margin: 1.5rem 0 0.5rem 0;
-		line-height: 1.4;
-	}
-
-	div :global(h4) {
-		font-size: 1.125rem;
-		font-weight: 600;
-		margin: 0.75rem 0 0.25rem 0;
-	}
-
-	div :global(h5) {
-		font-size: 1rem;
-		font-weight: 600;
-		margin: 0.5rem 0 0.25rem 0;
-	}
-
-	div :global(h6) {
-		font-size: 0.875rem;
-		font-weight: 600;
-		margin: 0.5rem 0 0.25rem 0;
-	}
-
-	/* Text formatting */
-	div :global(strong) {
-		font-weight: 600;
-	}
-
-	div :global(em) {
-		font-style: italic;
-	}
-
-	div :global(del) {
-		text-decoration: line-through;
-		opacity: 0.7;
-	}
-
-	/* Inline code */
-	div :global(code:not(pre code)) {
-		background: var(--muted);
-		color: var(--muted-foreground);
-		padding: 0.125rem 0.375rem;
-		border-radius: 0.375rem;
-		font-size: 0.875rem;
-		font-family:
-			ui-monospace, SFMono-Regular, 'SF Mono', Monaco, 'Cascadia Code', 'Roboto Mono', Consolas,
-			'Liberation Mono', Menlo, monospace;
-	}
-
-	/* Links */
-	div :global(a) {
-		color: var(--primary);
-		text-decoration: underline;
-		text-underline-offset: 2px;
-		transition: color 0.2s ease;
-		overflow-wrap: anywhere;
-		word-break: break-all;
-	}
-
-	div :global(a:hover) {
-		color: var(--primary);
-	}
-
-	/* Lists */
-	div :global(ul) {
-		list-style-type: disc;
-		margin-left: 1.5rem;
-		margin-bottom: 1rem;
-	}
-
-	div :global(ol) {
-		list-style-type: decimal;
-		margin-left: 1.5rem;
-		margin-bottom: 1rem;
-	}
-
-	div :global(li) {
-		margin-bottom: 0.25rem;
-		padding-left: 0.5rem;
-	}
-
-	div :global(li::marker) {
-		color: var(--muted-foreground);
-	}
-
-	/* Nested lists */
-	div :global(ul ul) {
-		list-style-type: circle;
-		margin-top: 0.25rem;
-		margin-bottom: 0.25rem;
-	}
-
-	div :global(ol ol) {
-		list-style-type: lower-alpha;
-		margin-top: 0.25rem;
-		margin-bottom: 0.25rem;
-	}
-
-	/* Task lists */
-	div :global(.task-list-item) {
-		list-style: none;
-		margin-left: 0;
-		padding-left: 0;
-	}
-
-	div :global(.task-list-item-checkbox) {
-		margin-right: 0.5rem;
-		margin-top: 0.125rem;
-	}
-
-	/* Blockquotes */
-	div :global(blockquote) {
-		border-left: 4px solid var(--border);
-		padding: 0.5rem 1rem;
-		margin: 1.5rem 0;
-		font-style: italic;
-		color: var(--muted-foreground);
-		background: var(--muted);
-		border-radius: 0 0.375rem 0.375rem 0;
-	}
-
-	/* Tables */
-	div :global(table) {
-		width: 100%;
-		margin: 1.5rem 0;
-		border-collapse: collapse;
-		border: 1px solid var(--border);
-		border-radius: 0.375rem;
-		overflow: hidden;
-	}
-
-	div :global(th) {
-		background: hsl(var(--muted) / 0.3);
-		border: 1px solid var(--border);
-		padding: 0.5rem 0.75rem;
-		text-align: left;
-		font-weight: 600;
-	}
-
-	div :global(td) {
-		border: 1px solid var(--border);
-		padding: 0.5rem 0.75rem;
-	}
-
-	div :global(tr:nth-child(even)) {
-		background: hsl(var(--muted) / 0.1);
-	}
-
-	/* User message markdown should keep table borders visible on light primary backgrounds */
-	div.markdown-user-content :global(table),
-	div.markdown-user-content :global(th),
-	div.markdown-user-content :global(td),
-	div.markdown-user-content :global(.table-wrapper) {
-		border-color: currentColor;
-	}
-
-	/* Horizontal rules */
-	div :global(hr) {
-		border: none;
-		border-top: 1px solid var(--border);
-		margin: 1.5rem 0;
-	}
-
-	/* Images */
-	div :global(img) {
-		border-radius: 0.5rem;
-		box-shadow:
-			0 1px 3px 0 rgb(0 0 0 / 0.1),
-			0 1px 2px -1px rgb(0 0 0 / 0.1);
-		margin: 1.5rem 0;
-		max-width: 100%;
-		height: auto;
-	}
-
-	/* Code blocks */
-
-	div :global(.code-block-wrapper) {
-		margin: 1.5rem 0;
-		border-radius: 0.75rem;
-		overflow: hidden;
-		border: 1px solid var(--border);
-		background: var(--code-background);
-	}
-
-	div :global(.code-block-header) {
-		display: flex;
-		justify-content: space-between;
-		align-items: center;
-		padding: 0.5rem 1rem;
-		background: hsl(var(--muted) / 0.5);
-		border-bottom: 1px solid var(--border);
-		font-size: 0.875rem;
-	}
-
-	div :global(.code-language) {
-		color: var(--code-foreground);
-		font-weight: 500;
-		font-family:
-			ui-monospace, SFMono-Regular, 'SF Mono', Monaco, 'Cascadia Code', 'Roboto Mono', Consolas,
-			'Liberation Mono', Menlo, monospace;
-		text-transform: uppercase;
-		font-size: 0.75rem;
-		letter-spacing: 0.05em;
-	}
-
-	div :global(.code-block-actions) {
-		display: flex;
-		align-items: center;
-		gap: 0.5rem;
-	}
-
-	div :global(.copy-code-btn),
-	div :global(.preview-code-btn) {
-		display: flex;
-		align-items: center;
-		justify-content: center;
-		padding: 0;
-		background: transparent;
-		color: var(--code-foreground);
-		cursor: pointer;
-		transition: all 0.2s ease;
-	}
-
-	div :global(.copy-code-btn:hover),
-	div :global(.preview-code-btn:hover) {
-		transform: scale(1.05);
-	}
-
-	div :global(.copy-code-btn:active),
-	div :global(.preview-code-btn:active) {
-		transform: scale(0.95);
-	}
-
-	div :global(.code-block-wrapper pre) {
-		background: transparent;
-		padding: 1rem;
-		margin: 0;
-		overflow-x: auto;
-		border-radius: 0;
-		border: none;
-		font-size: 0.875rem;
-		line-height: 1.5;
-	}
-
-	div :global(pre) {
-		background: var(--muted);
-		margin: 1.5rem 0;
-		overflow-x: auto;
-		border-radius: 1rem;
-		border: none;
-	}
-
-	div :global(code) {
-		background: transparent;
-		color: var(--code-foreground);
-	}
-
-	/* Mentions and hashtags */
-	div :global(.mention) {
-		color: hsl(var(--primary));
-		font-weight: 500;
-		text-decoration: none;
-	}
-
-	div :global(.mention:hover) {
-		text-decoration: underline;
-	}
-
-	div :global(.hashtag) {
-		color: hsl(var(--primary));
-		font-weight: 500;
-		text-decoration: none;
-	}
-
-	div :global(.hashtag:hover) {
-		text-decoration: underline;
-	}
-
-	/* Advanced table enhancements */
-	div :global(table) {
-		transition: all 0.2s ease;
-	}
-
-	div :global(table:hover) {
-		box-shadow:
-			0 4px 6px -1px rgb(0 0 0 / 0.1),
-			0 2px 4px -2px rgb(0 0 0 / 0.1);
-	}
-
-	div :global(th:hover),
-	div :global(td:hover) {
-		background: var(--muted);
-	}
-
-	/* Disable hover effects when rendering user messages */
-	.markdown-user-content :global(a),
-	.markdown-user-content :global(a:hover) {
-		color: var(--primary-foreground);
-	}
-
-	.markdown-user-content :global(table:hover) {
-		box-shadow: none;
-	}
-
-	.markdown-user-content :global(th:hover),
-	.markdown-user-content :global(td:hover) {
-		background: inherit;
-	}
-
-	/* Enhanced blockquotes */
-	div :global(blockquote) {
-		transition: all 0.2s ease;
-		position: relative;
-	}
-
-	div :global(blockquote:hover) {
-		border-left-width: 6px;
-		background: var(--muted);
-		transform: translateX(2px);
-	}
-
-	div :global(blockquote::before) {
-		content: '"';
-		position: absolute;
-		top: -0.5rem;
-		left: 0.5rem;
-		font-size: 3rem;
-		color: var(--muted-foreground);
-		font-family: serif;
-		line-height: 1;
-	}
-
-	/* Enhanced images */
-	div :global(img) {
-		transition: all 0.3s ease;
-		cursor: pointer;
-	}
-
-	div :global(img:hover) {
-		transform: scale(1.02);
-		box-shadow:
-			0 10px 15px -3px rgb(0 0 0 / 0.1),
-			0 4px 6px -4px rgb(0 0 0 / 0.1);
-	}
-
-	/* Image zoom overlay */
-	div :global(.image-zoom-overlay) {
-		position: fixed;
-		top: 0;
-		left: 0;
-		right: 0;
-		bottom: 0;
-		background: rgba(0, 0, 0, 0.8);
-		display: flex;
-		align-items: center;
-		justify-content: center;
-		z-index: 1000;
-		cursor: pointer;
-	}
-
-	div :global(.image-zoom-overlay img) {
-		max-width: 90vw;
-		max-height: 90vh;
-		border-radius: 0.5rem;
-		box-shadow: 0 25px 50px -12px rgb(0 0 0 / 0.25);
-	}
-
-	/* Enhanced horizontal rules */
-	div :global(hr) {
-		border: none;
-		height: 2px;
-		background: linear-gradient(to right, transparent, var(--border), transparent);
-		margin: 2rem 0;
-		position: relative;
-	}
-
-	div :global(hr::after) {
-		content: '';
-		position: absolute;
-		top: 50%;
-		left: 50%;
-		transform: translate(-50%, -50%);
-		width: 1rem;
-		height: 1rem;
-		background: var(--border);
-		border-radius: 50%;
-	}
-
-	/* Scrollable tables */
-	div :global(.table-wrapper) {
-		overflow-x: auto;
-		margin: 1.5rem 0;
-		border-radius: 0.5rem;
-		border: 1px solid var(--border);
-	}
-
-	div :global(.table-wrapper table) {
-		margin: 0;
-		border: none;
-	}
-
-	/* Responsive adjustments */
-	@media (max-width: 640px) {
-		div :global(h1) {
-			font-size: 1.5rem;
-		}
-
-		div :global(h2) {
-			font-size: 1.25rem;
-		}
-
-		div :global(h3) {
-			font-size: 1.125rem;
-		}
-
-		div :global(table) {
-			font-size: 0.875rem;
-		}
-
-		div :global(th),
-		div :global(td) {
-			padding: 0.375rem 0.5rem;
-		}
-
-		div :global(.table-wrapper) {
-			margin: 0.5rem -1rem;
-			border-radius: 0;
-			border-left: none;
-			border-right: none;
-		}
-	}
-
-	/* Dark mode adjustments */
-	@media (prefers-color-scheme: dark) {
-		div :global(blockquote:hover) {
-			background: var(--muted);
-		}
-	}
-</style>

tools/server/webui/src/lib/components/app/misc/RemoveButton.svelte

@@ -1,26 +0,0 @@
-<script lang="ts">
-	import { X } from '@lucide/svelte';
-	import { Button } from '$lib/components/ui/button';
-
-	interface Props {
-		id: string;
-		onRemove?: (id: string) => void;
-		class?: string;
-	}
-
-	let { id, onRemove, class: className = '' }: Props = $props();
-</script>
-
-<Button
-	type="button"
-	variant="ghost"
-	size="sm"
-	class="h-6 w-6 bg-white/20 p-0 hover:bg-white/30 {className}"
-	onclick={(e) => {
-		e.stopPropagation();
-		onRemove?.(id);
-	}}
-	aria-label="Remove file"
->
-	<X class="h-3 w-3" />
-</Button>

tools/server/webui/src/lib/components/app/misc/SyntaxHighlightedCode.svelte

@@ -1,97 +0,0 @@
-<script lang="ts">
-	import hljs from 'highlight.js';
-	import { browser } from '$app/environment';
-	import { mode } from 'mode-watcher';
-
-	import githubDarkCss from 'highlight.js/styles/github-dark.css?inline';
-	import githubLightCss from 'highlight.js/styles/github.css?inline';
-
-	interface Props {
-		code: string;
-		language?: string;
-		class?: string;
-		maxHeight?: string;
-		maxWidth?: string;
-	}
-
-	let {
-		code,
-		language = 'text',
-		class: className = '',
-		maxHeight = '60vh',
-		maxWidth = ''
-	}: Props = $props();
-
-	let highlightedHtml = $state('');
-
-	function loadHighlightTheme(isDark: boolean) {
-		if (!browser) return;
-
-		const existingThemes = document.querySelectorAll('style[data-highlight-theme-preview]');
-		existingThemes.forEach((style) => style.remove());
-
-		const style = document.createElement('style');
-		style.setAttribute('data-highlight-theme-preview', 'true');
-		style.textContent = isDark ? githubDarkCss : githubLightCss;
-
-		document.head.appendChild(style);
-	}
-
-	$effect(() => {
-		const currentMode = mode.current;
-		const isDark = currentMode === 'dark';
-
-		loadHighlightTheme(isDark);
-	});
-
-	$effect(() => {
-		if (!code) {
-			highlightedHtml = '';
-			return;
-		}
-
-		try {
-			// Check if the language is supported
-			const lang = language.toLowerCase();
-			const isSupported = hljs.getLanguage(lang);
-
-			if (isSupported) {
-				const result = hljs.highlight(code, { language: lang });
-				highlightedHtml = result.value;
-			} else {
-				// Try auto-detection or fallback to plain text
-				const result = hljs.highlightAuto(code);
-				highlightedHtml = result.value;
-			}
-		} catch {
-			// Fallback to escaped plain text
-			highlightedHtml = code.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;');
-		}
-	});
-</script>
-
-<div
-	class="code-preview-wrapper overflow-auto rounded-lg border border-border bg-muted {className}"
-	style="max-height: {maxHeight}; max-width: {maxWidth};"
->
-	<!-- Needs to be formatted as single line for proper rendering -->
-	<pre class="m-0 overflow-x-auto p-4"><code class="hljs text-sm leading-relaxed"
-			>{@html highlightedHtml}</code
-		></pre>
-</div>
-
-<style>
-	.code-preview-wrapper {
-		font-family:
-			ui-monospace, SFMono-Regular, 'SF Mono', Monaco, 'Cascadia Code', 'Roboto Mono', Consolas,
-			'Liberation Mono', Menlo, monospace;
-	}
-
-	.code-preview-wrapper pre {
-		background: transparent;
-	}
-
-	.code-preview-wrapper code {
-		background: transparent;
-	}
-</style>