metal : per-device tuned (Q, NE) for flash-attn vec (#25750 )

* rebase Q-generic FA vec body from 01dc93607 (#23114) * add 53 f16 (Q,NE) flash-attn vec instantiations (vec 80 -> 133) * add FA vec (Q,NE) tuning table + dispatch wiring + SMEM cap fallback * add FA vec (Q,NE) perf sweep * fill tuning result * fold family table into a per-family representative SKU * refactor tuning result format * extend FA vec tuning to quantized KV caches * sync fa vec tuner bucketing with runtime, use pointwise tuning regret * update tuned table * format and cleanup * prefix fa_vec tuning procs with ggml_backend_metal_tuning_, drop unused fa_vec_override_active
ggml-metal: FWHT kernel for metal backend (#25924 )
2026-07-29 15:55:56 +02:00 · 2026-07-29 21:46:01 +08:00 · 2026-07-28 19:05:42 +03:00 · 2026-07-28 19:02:22 +03:00 · 2026-07-28 19:00:04 +03:00 · 2026-07-28 18:48:08 +03:00
58 changed files with 14774 additions and 11631 deletions
@@ -39,28 +39,48 @@ class NemotronNanoV2VLModel(MmprojModel):
        }
        return vision_config

+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("sound_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]

+        if self.hparams_audio is not None:
+            self.has_vision_encoder = True
+            self.has_audio_encoder = True
+            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["num_mel_bins"])
+            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+            self.gguf_writer.add_audio_subsampling_factor(self.hparams_audio["subsampling_factor"])
+            self.gguf_writer.add_audio_conv_kernel_size(self.hparams_audio["conv_kernel_size"])
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.PARAKEET)
+            self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+        else:
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+
        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
+        if "sound_encoder" in name or new_name.startswith("mm.a."):
+            if "bias" in new_name or "norm" in new_name:
+                return gguf.GGMLQuantizationType.F32
+            if "conv" in new_name and "weight" in new_name:
+                return gguf.GGMLQuantizationType.F32
+
        return super().tensor_force_quant(name, new_name, bid, n_dims)

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem

        if "input_conditioner" in name:
            return None
@@ -69,14 +89,18 @@ class NemotronNanoV2VLModel(MmprojModel):
        if "radio_model.model.patch_generator.video_embedder" in name:
            return None

-        if not name.startswith("vision_model.radio_model.model.") and not name.startswith("mlp1."):
+        if not name.startswith(("vision_model.radio_model.model.", "mlp1.", "sound_encoder.", "sound_projection.")):
            return None

        if "patch_generator.pos_embed" in name:
            if not name.endswith(".weight"):
                name += ".weight"

-        return super().filter_tensors((name, gen))
+        # num_batches is only used for training not inference.
+        if "conv.norm" in name and "num_batches" in name:
+            return None
+
+        return name, gen

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
@@ -104,7 +128,26 @@ class NemotronNanoV2VLModel(MmprojModel):
            n_embd = self.hparams["hidden_size"]
            data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)

-        yield from super().modify_tensors(data_torch, name, bid)
+        if "depthwise_conv.weight" in name:
+            data_torch = data_torch.unsqueeze(-1)
+            data_torch = data_torch.permute(3, 1, 0, 2).contiguous()
+
+        if "pointwise_conv" in name and name.endswith(".weight"):
+            if len(data_torch.shape) == 3 and data_torch.shape[2] == 1:
+                data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
+
+        if "subsampling.layers" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        if "pointwise_conv" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        for mapped_name, tensor in super().modify_tensors(data_torch, name, bid):
+            if name.startswith("sound_projection.") and mapped_name.startswith("mm.model.mlp."):
+                mapped_name = mapped_name.replace("mm.model.mlp.", "mm.a.mlp.")
+            yield mapped_name, tensor


@ModelBase.register("NemotronForCausalLM")
@@ -16,22 +16,22 @@ conda-forge provides builds for:
 - Apple Metal (macOS)

 ```sh
-conda install -c conda-forge llama-cpp
+conda install -c conda-forge llama.cpp
 ```

 ```sh
-mamba install -c conda-forge llama-cpp
+mamba install -c conda-forge llama.cpp
 ```

 ```sh
 # Project-local installation
-pixi add llama-cpp
+pixi add llama.cpp

 # Global installation
-pixi global install llama-cpp
+pixi global install llama.cpp
 ```

-This distribution is managed on [`conda-forge/llama-cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).
+This distribution is managed on [`conda-forge/llama.cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).

 Shall you have any problems, please open an issue on [its issue tracker](https://github.com/conda-forge/llama.cpp-feedstock/issues).

@@ -341,9 +341,6 @@ set(GGML_PUBLIC_HEADERS
    include/gguf.h)

 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)

@@ -11,6 +11,7 @@ ggml_add_backend_library(ggml-metal
                         ggml-metal-common.cpp
                         ggml-metal-context.m
                         ggml-metal-ops.cpp
+                         ggml-metal-tuning.cpp
                        )

 target_link_libraries(ggml-metal PRIVATE
@@ -24,62 +25,119 @@ if (GGML_METAL_NDEBUG)
 endif()

 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
    enable_language(ASM)

    add_compile_definitions(GGML_METAL_EMBED_LIBRARY)

-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")

    file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")

-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})

-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")

-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
    configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
    configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)

+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
    if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
        #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
        # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
        set(XC_FLAGS -fno-fast-math -fno-inline)
    else()
        set(XC_FLAGS -O3)
    endif()

-    # Append macOS metal versioning flags
    if (GGML_METAL_MACOSX_VERSION_MIN)
        message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
        list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +148,46 @@ else()
        list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
    endif()

+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
    add_custom_command(
        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )

-    # FIXME: only add to the ggml-metal target?
    add_custom_target(
        ggml-metal-lib ALL
        DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY

 if (NOT GGML_METAL_EMBED_LIBRARY)
    install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )

-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()
@@ -1,6 +1,7 @@
 #include "ggml-metal-device.h"

 #include "ggml-metal-impl.h"
+#include "ggml-metal-tuning.h"

 #include "ggml-impl.h"

@@ -1480,6 +1481,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
        int32_t nsg,
        int32_t nwg) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
@@ -1493,11 +1496,17 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];

-    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
+    char qne_suffix[16] = {0};
+    if (!(nqpsg == 1 && ne == ggml_metal_tuning::fa_vec_baseline_ne(dk, dv))) {
+        snprintf(qne_suffix, sizeof(qne_suffix), "_q%d_ne%d", nqpsg, ne);
+    }
+
+    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d%s",
            "flash_attn_ext_vec",
            ggml_type_name(op->src[1]->type),
            dk,
-            dv);
+            dv,
+            qne_suffix);

    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
            base,
@@ -197,6 +197,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
        int32_t nsg,
        int32_t nwg);

@@ -264,6 +266,7 @@ struct ggml_metal_device_props {
    bool supports_gpu_family_apple7;

    enum ggml_metal_device_id device_id;
+    int gpu_family;

    int op_offload_min_batch_size;
 };
@@ -94,8 +94,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
    return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }

+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;

    ggml_metal_device_t dev;
    ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -103,160 +158,376 @@ struct ggml_metal_library {
    NSLock * lock;
 };

-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}

-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];

-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }

-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }

-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}

-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];

-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }

-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }

-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
            }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
        }

-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }

-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}

-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];

-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];

-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
            }

-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;

-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
            @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                MTLCompileOptions * options = [MTLCompileOptions new];
                options.preprocessorMacros = prep;

-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];

-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
            }
        }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
    }

-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);

-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
    res->dev       = dev;
    res->pipelines = ggml_metal_pipelines_init();
    res->lock      = [NSLock new];

+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
    return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }

 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -318,10 +589,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
        return NULL;
    }

-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];

    return res;
 }
@@ -331,8 +603,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
        return;
    }

-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
    }

    ggml_metal_pipelines_free(lib->pipelines);
@@ -393,11 +671,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_

        GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);

+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
        id<MTLFunction> mtl_function;
        if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
        } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
        }
        if (!mtl_function) {
            [lib->lock unlock];
@@ -878,7 +1173,8 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
            {
                for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
                    if ([dev->mtl_device supportsFamily:i]) {
-                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+                        dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1;
+                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, dev->props.gpu_family, i);
                        break;
                    }
                }
@@ -7,6 +7,7 @@
 #include "ggml-metal-impl.h"
 #include "ggml-metal-common.h"
 #include "ggml-metal-device.h"
+#include "ggml-metal-tuning.h"

 #include <cassert>
 #include <algorithm>
@@ -2942,12 +2943,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
    } else {
        // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
+        auto cfg = ggml_metal_tuning::fa_vec_pick(
+                props_dev->device_id,
+                props_dev->gpu_family,
+                (int) op->src[1]->type,
+                (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
+                ne11, ne01);
+        int nqptg = cfg.Q;                             // queries per threadgroup
        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
        const int nhptg = 1;                           // heads per threadgroup

        GGML_ASSERT(nqptg <= 32);
-        GGML_ASSERT(nqptg  % 1  == 0);
+        GGML_ASSERT(nqptg == 1 || nqptg == 2 || nqptg == 4);  // only instantiated Q values
        GGML_ASSERT(ncpsg  % 32 == 0);

        bool need_sync = false;
@@ -3006,7 +3013,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        // ne20*(nsg)
        // each simdgroup has a full f32 head vector in shared mem to accumulate results
        //
-#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg)*nqptg)*(sizeof(float)/2), 16))

        int64_t nsg = 1;

@@ -3026,6 +3033,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            }
        }

+        // fall back to baseline (Q=1) if the tuned config exceeds threadgroup memory
+        if ((size_t) FATTN_SMEM(nsg) > props_dev->max_theadgroup_memory_size) {
+            cfg   = ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20);
+            nqptg = cfg.Q;  // = 1
+        }
+
        ggml_metal_kargs_flash_attn_ext_vec args = {
            /*.ne01          =*/ ne01,
            /*.ne02          =*/ ne02,
@@ -3061,7 +3074,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.logit_softcap =*/ logit_softcap,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nqptg, cfg.NE, nsg, nwg);

        GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

@@ -0,0 +1,347 @@
+#include "ggml-metal-tuning.h"
+
+#include <cstddef>
+#include <cstring>
+#include <iterator>
+
+namespace ggml_metal_tuning {
+
+int fa_vec_ne11_bucket(int64_t ne11) {
+    for (int i = 0; i < (int) std::size(FA_VEC_NE11_BUCKETS); ++i) {
+        if (ne11 < FA_VEC_NE11_BUCKETS[i]) {
+            return i;
+        }
+    }
+    return (int) std::size(FA_VEC_NE11_BUCKETS);
+}
+
+int fa_vec_ne01_bucket(int64_t ne01) {
+    for (int i = 0; i < (int) std::size(FA_VEC_NE01_BUCKETS); ++i) {
+        if (ne01 < FA_VEC_NE01_BUCKETS[i]) {
+            return i;
+        }
+    }
+    return (int) std::size(FA_VEC_NE01_BUCKETS);
+}
+
+int fa_vec_baseline_ne(int dk, int dv) {
+    if (dk == 32 && dv == 32) {
+        return 4;
+    }
+    if (dk == 64 && dv == 64) {
+        return 2;
+    }
+    if (dk == 96 && dv == 96) {
+        return 4;
+    }
+    if (dk == 128 && dv == 128) {
+        return 1;
+    }
+    if (dk == 192 && dv == 192) {
+        return 2;
+    }
+    if (dk == 192 && dv == 128) {
+        return 2;
+    }
+    if (dk == 256 && dv == 256) {
+        return 1;
+    }
+    if (dk == 320 && dv == 256) {
+        return 2;
+    }
+    if (dk == 512 && dv == 512) {
+        return 1;
+    }
+    if (dk == 576 && dv == 512) {
+        return 2;
+    }
+    return 4;  // template default
+}
+
+fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv) {
+    return { 1, (int8_t) fa_vec_baseline_ne(dk, dv) };
+}
+
+// Generated by `test-backend-ops tune --tune-perf`; do not hand-edit.
+// One row per kept bucket, plus per-(dtype,dk,dv) ne11-collapsed domain defaults
+// (ne11_b = FA_VEC_NE11_DEFAULT, ne01_b = domain). To retune or add a device, re-run the
+// sweep and paste its block. See ggml-metal-tuning.h for the row/lookup semantics.
+constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
+    // ---- f16: 13 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 32, 32, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 128, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 192, 192, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 192, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_F16, 320, 256, -1, 1 }, { 2, 2 } },
+    // ---- q4_0: 29 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 4 }, { 1, 2 } },
+    // ---- q4_1: 28 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    // ---- q5_0: 45 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 4 }, { 1, 4 } },
+    // ---- q5_1: 49 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 4 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 512, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 4 }, { 1, 4 } },
+    // ---- q8_0: 29 rows ----
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M4_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+};
+
+static enum ggml_metal_device_id fa_vec_family_representative(int gpu_family) {
+    switch (gpu_family) {
+        case 9:  return GGML_METAL_DEVICE_M4_MAX;
+        default: return GGML_METAL_DEVICE_GENERIC;
+    }
+}
+
+static bool         g_override_set = false;
+static fa_vec_cfg_t g_override_cfg = { 1, 4 };
+
+void fa_vec_set_override(fa_vec_cfg_t cfg) {
+    g_override_cfg = cfg;
+    g_override_set = true;
+}
+
+void fa_vec_clear_override() {
+    g_override_set = false;
+}
+
+static const fa_vec_cfg_t * find_cfg(const fa_vec_entry_t * tbl, size_t n, const fa_vec_key_t & k) {
+    for (size_t i = 0; i < n; ++i) {
+        if (memcmp(&tbl[i].key, &k, sizeof(k)) == 0) {
+            return &tbl[i].cfg;
+        }
+    }
+    return nullptr;
+}
+
+fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01) {
+    if (g_override_set) {
+        return g_override_cfg;
+    }
+
+    const fa_vec_cfg_t baseline = fa_vec_baseline_cfg(dk, dv);
+
+    const int ne11_b = fa_vec_ne11_bucket(ne11);
+    if (ne11_b == 0) {
+        return baseline;  // short KV: attention is a small slice of the step, left to baseline
+    }
+    const int ne01_b = fa_vec_ne01_bucket(ne01);
+
+    fa_vec_key_t k{};
+    k.dtype = (int8_t) dtype;
+    k.dk    = (int16_t) dk;
+    k.dv    = (int16_t) dv;
+
+    // exact bucket, then the ne01 domain default (ne11 collapsed); tried under each device tier
+    auto lookup = [&](enum ggml_metal_device_id dev) -> const fa_vec_cfg_t * {
+        k.device_id = (int8_t) dev;
+        k.ne11_b    = (int8_t) ne11_b;
+        k.ne01_b    = (int8_t) ne01_b;
+        if (auto * c = find_cfg(fa_vec_tuned_table, std::size(fa_vec_tuned_table), k)) {
+            return c;
+        }
+        k.ne11_b = FA_VEC_NE11_DEFAULT;
+        k.ne01_b = (ne01_b == 0) ? FA_VEC_DOMAIN_DECODE : FA_VEC_DOMAIN_BATCH;
+        return find_cfg(fa_vec_tuned_table, std::size(fa_vec_tuned_table), k);
+    };
+
+    if (auto * c = lookup(device_id)) {
+        return *c;
+    }
+
+    // family fallback: retry under the family's representative SKU; none -> baseline
+    if (gpu_family > 0) {
+        const enum ggml_metal_device_id rep = fa_vec_family_representative(gpu_family);
+        if (rep != GGML_METAL_DEVICE_GENERIC) {
+            if (auto * c = lookup(rep)) {
+                return *c;
+            }
+        }
+    }
+
+    return baseline;
+}
+
+}  // namespace ggml_metal_tuning
@@ -0,0 +1,63 @@
+#pragma once
+
+#include "ggml-metal-device.h"  // enum ggml_metal_device_id
+#include "ggml.h"
+
+#include <cstdint>
+
+namespace ggml_metal_tuning {
+
+// FA vec selection buckets. ne01 (query rows) splits decode (==1) from batch (>=2), the
+// batch side refined into {2,3,4,5}: Q>1 reuses one K/V load across rows, so it only pays
+// off once ne01 aligns with Q. ne11 (KV length) is bucketed too, as the Q>1 crossover is
+// head-size dependent (small dk crosses late, large dk wins even at short KV).
+constexpr int FA_VEC_NE11_BUCKETS[] = { 1024, 4096, 16384 };
+constexpr int FA_VEC_NE01_BUCKETS[] = { 2, 3, 4, 5 };
+
+int fa_vec_ne11_bucket(int64_t ne11);
+int fa_vec_ne01_bucket(int64_t ne01);
+
+// NE baked into each (dk,dv) baseline instantiation in kernels/fa.metal.
+// Hand-maintained mirror; keep in sync with those instantiations (run_fa_vec_tune_check
+// exercises every (Q,NE), so a missing instantiation surfaces there).
+int fa_vec_baseline_ne(int dk, int dv);
+
+// Tuned table has two row kinds. Exact rows key a (ne11_b, ne01_b) bucket. Default rows
+// collapse ne11 over one ne01 domain: ne11_b == FA_VEC_NE11_DEFAULT and ne01_b holds the
+// domain. fa_vec_pick tries exact bucket -> domain default -> baseline; short KV
+// (ne11 < FA_VEC_NE11_BUCKETS[0]) always uses baseline.
+constexpr int8_t FA_VEC_NE11_DEFAULT  = -1;
+constexpr int8_t FA_VEC_DOMAIN_DECODE = 0;  // ne01 == 1
+constexpr int8_t FA_VEC_DOMAIN_BATCH  = 1;  // ne01 >= 2
+
+struct fa_vec_key_t {
+    int8_t  device_id;
+    int8_t  dtype;
+    int16_t dk;
+    int16_t dv;
+    int8_t  ne11_b;
+    int8_t  ne01_b;
+};
+
+static_assert(sizeof(fa_vec_key_t) == 8, "fa_vec_key_t must be tightly packed for memcmp");
+
+struct fa_vec_cfg_t {
+    int8_t Q;
+    int8_t NE;
+};
+
+struct fa_vec_entry_t {
+    fa_vec_key_t key;
+    fa_vec_cfg_t cfg;
+};
+
+// test/tune-only override; when set, fa_vec_pick returns it directly.
+void         fa_vec_set_override(fa_vec_cfg_t cfg);
+void         fa_vec_clear_override();
+fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
+
+// device_id selects a per-SKU row; on a miss, gpu_family (0 if unknown) maps to a representative
+// SKU and the table is retried. No match -> baseline.
+fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
+
+}  // namespace ggml_metal_tuning
@@ -6,6 +6,7 @@
 #include "ggml-metal-device.h"
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
+#include "ggml-metal-tuning.h"

 #include <mutex>
 #include <string>
@@ -868,10 +869,46 @@ static ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t
    GGML_UNUSED(reg);
 }

+// test/tune-only override for the FA vec (Q, NE) selection, reached via proc_address.
+static void ggml_backend_metal_tuning_set_fa_vec_override(int Q, int NE) {
+    ggml_metal_tuning::fa_vec_set_override({ (int8_t) Q, (int8_t) NE });
+}
+
+static void ggml_backend_metal_tuning_clear_fa_vec_override(void) {
+    ggml_metal_tuning::fa_vec_clear_override();
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne11_bucket(int64_t ne11) {
+    return ggml_metal_tuning::fa_vec_ne11_bucket(ne11);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne01_bucket(int64_t ne01) {
+    return ggml_metal_tuning::fa_vec_ne01_bucket(ne01);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_baseline_ne(int dk, int dv) {
+    return ggml_metal_tuning::fa_vec_baseline_ne(dk, dv);
+}
+
 static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
    if (strcmp(name, "ggml_backend_get_features") == 0) {
        return (void *)ggml_backend_metal_get_features;
    }
+    if (strcmp(name, "ggml_backend_metal_tuning_set_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_set_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_clear_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_clear_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne11_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne11_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne01_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne01_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_baseline_ne") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_baseline_ne;
+    }

    return NULL;

@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;
@@ -0,0 +1,226 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}
@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
@@ -0,0 +1,719 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 16; i++) {
+        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif
@@ -0,0 +1,438 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template <typename T>
+kernel void kernel_snake(
+        constant ggml_metal_kargs_snake & args,
+        device const T     * x,
+        device const float * a,
+        device const float * inv_b,
+        device       T     * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T * args.C) {
+        return;
+    }
+
+    const int   c  = idx / args.T;  // x is [T, C], a / inv_b collapse to [1, C]
+    const float xi = float(x[idx]);
+    const float si = sin(a[c] * xi);
+    dst[idx] = T(xi + si * si * inv_b[c]);
+}
+
+template [[host_name("kernel_snake_f32")]]  kernel void kernel_snake<float>(constant ggml_metal_kargs_snake &, device const float *, device const float *, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_snake_f16")]]  kernel void kernel_snake<half>(constant ggml_metal_kargs_snake &, device const half *, device const float *, device const float *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_snake_bf16")]] kernel void kernel_snake<bfloat>(constant ggml_metal_kargs_snake &, device const bfloat *, device const float *, device const float *, device bfloat *, uint, uint, uint);
+#endif
+
+template<int N>
+kernel void kernel_fwht_f32(
+        constant ggml_metal_kargs_fwht & args,
+        device const float * src,
+        device float * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort3  ntg[[threads_per_threadgroup]]) {
+
+    constexpr int NW = N_SIMDWIDTH;
+    constexpr int NE = N / NW;
+
+    const float scale = 1.0f / sqrt((float) N);
+
+    const int sg_per_tg = ntg.x / NW;
+    const int64_t r = tgpig.x * sg_per_tg + sgitg;
+    if (r >= args.nrows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    const int lane = tiisg;
+
+    float reg[NE];
+    for (int i = 0; i < NE; i++) {
+        reg[i] = src[i*NW + lane]*scale;
+    }
+    for (int i = 1; i < NW; i *= 2) {
+        for (int j = 0; j < NE; j++) {
+            const float val = reg[j];
+            const float val2 = simd_shuffle_xor(val, i);
+            reg[j] = (lane & i) == 0 ? val2 + val : val2 - val;
+        }
+    }
+
+    for (int i = NW; i < N; i *= 2) {
+        const int step = i / NW;
+        for (int j = 0; j < NE; j += (2 * step)) {
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k ];
+                const float y = reg[j + k + step];
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+    for (int i = 0; i < NE; i++) {
+        dst[i*NW + lane] = reg[i];
+    }
+}
+
+typedef decltype(kernel_fwht_f32<64>) kernel_fwht_t;
+
+template [[host_name("kernel_fwht_f32_64")]]  kernel kernel_fwht_t kernel_fwht_f32<64>;
+template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f32<128>;
+template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
+
@@ -0,0 +1,842 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}
@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}
@@ -0,0 +1,234 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q2_0(device const float * src, device block_q2_0 & dst) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK2_0; j++) {
+        float a = fabs(src[j]);
+        if (a > amax) amax = a;
+    }
+    const float d = amax;
+    dst.d = d;
+
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    for (int j = 0; j < QK2_0 / 4; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK2_0; j++) {
+        int q = (int)round(src[j] * id) + 1;
+        q = max(0, min(3, q));
+        dst.qs[j / 4] |= (q << (2 * (j % 4)));
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
@@ -0,0 +1,393 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q2_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK2_0,  block_q2_0,   quantize_q2_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q2_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_0,    4, dequantize_q2_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename TD>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) src_row[ind];
+    }
+}
+
+typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
+template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
+template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
+template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
+#endif
+
+template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
+template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
+
@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
@@ -0,0 +1,318 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;
@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;
@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}
@@ -0,0 +1,279 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}
@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif
@@ -0,0 +1,360 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;
@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device const char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}
@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}
@@ -373,6 +373,7 @@ class Keys:
        FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
        PROJECTION_DIM      = "clip.audio.projection_dim"
        BLOCK_COUNT         = "clip.audio.block_count"
+        SUBSAMPLING_FACTOR  = "clip.audio.subsampling_factor"
        CHUNK_SIZE          = "clip.audio.chunk_size"
        CONV_KERNEL_SIZE    = "clip.audio.conv_kernel_size"
        MAX_POS_EMB         = "clip.audio.max_pos_emb"
@@ -1002,6 +1003,10 @@ class MODEL_TENSOR(IntEnum):
    A_ENC_CONV_NORM        = auto() # SSM conv
    A_ENC_CONV_PW1         = auto()
    A_ENC_CONV_PW2         = auto()
+    A_ENC_CONV_NORM_MEAN   = auto() # parakeet
+    A_ENC_CONV_NORM_VAR    = auto() # parakeet
+    A_ENC_MEL_FILTERS      = auto() # parakeet
+    A_ENC_WINDOW           = auto() # parakeet
    A_CTC_OUT              = auto()
    A_CTC_OUT_MID          = auto()
    A_ENC_ATTN_REL_POS_EMB = auto()
@@ -1591,6 +1596,10 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.A_ENC_CONV_NORM:           "a.blk.{bid}.conv_norm",
    MODEL_TENSOR.A_ENC_CONV_PW1:            "a.blk.{bid}.conv_pw1",
    MODEL_TENSOR.A_ENC_CONV_PW2:            "a.blk.{bid}.conv_pw2",
+    MODEL_TENSOR.A_ENC_CONV_NORM_MEAN:      "a.blk.{bid}.conv_norm_mean",
+    MODEL_TENSOR.A_ENC_CONV_NORM_VAR:       "a.blk.{bid}.conv_norm_var",
+    MODEL_TENSOR.A_ENC_MEL_FILTERS:         "a.mel_filters",
+    MODEL_TENSOR.A_ENC_WINDOW:              "a.window",
    MODEL_TENSOR.A_CTC_OUT:                 "a.enc_ctc_out",
    MODEL_TENSOR.A_CTC_OUT_MID:             "a.enc_ctc_out_mid",
    MODEL_TENSOR.A_ENC_ATTN_REL_POS_EMB:    "a.blk.{bid}.attn_rel_pos_emb",
@@ -1810,6 +1819,10 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.A_ENC_CONV_NORM,
        MODEL_TENSOR.A_ENC_CONV_PW1,
        MODEL_TENSOR.A_ENC_CONV_PW2,
+        MODEL_TENSOR.A_ENC_CONV_NORM_MEAN,
+        MODEL_TENSOR.A_ENC_CONV_NORM_VAR,
+        MODEL_TENSOR.A_ENC_MEL_FILTERS,
+        MODEL_TENSOR.A_ENC_WINDOW,
        MODEL_TENSOR.A_MM_INP_PROJ,
        MODEL_TENSOR.A_MM_SOFT_EMB_NORM,
        MODEL_TENSOR.A_MM_EMBEDDING,
@@ -4861,6 +4874,7 @@ class VisionProjectorType:
    YOUTUVL = "youtuvl"
    NEMOTRON_V2_VL = "nemotron_v2_vl"
    HUNYUANVL      = "hunyuanvl"
+    PARAKEET       = "parakeet"  # audio
    MINIMAXM3      = "minimax_m3"
    MINICPMV4_6    = "minicpmv4_6"
    GRANITE_SPEECH = "granite_speech"  # audio
@@ -1374,6 +1374,9 @@ class GGUFWriter:
    def add_audio_stack_factor(self, value: int) -> None:
        self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)

+    def add_audio_subsampling_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.SUBSAMPLING_FACTOR, value)
+
    def add_audio_chunk_size(self, value: int) -> None:
        self.add_uint32(Keys.ClipAudio.CHUNK_SIZE, value)

@@ -2107,6 +2107,7 @@ class TensorNameMap:
            "conformer.pre_encode.conv.{bid}", # lfm2
            "model.audio_tower.subsample_conv_projection.conv_{bid}.conv", # gemma3n
            "conformer.subsample_conv_projection.layer{bid}.conv", # gemma4
+            "sound_encoder.encoder.subsampling.layers.{bid}", # parakeet
            "encoder.conv{bid}", # mimo-audio-tokenizer
        ),

@@ -2140,6 +2141,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.self_attn.linear_q", # lfm2
            "conformer.layers.{bid}.attention.attn.q_proj", # gemma3n
            "conformer.layers.{bid}.self_attn.q_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.q_proj", # parakeet
            "encoder.layers.{bid}.attn.to_q", # granite_speech
            "encoder.layers.{bid}.self_attn.q_proj", # mimo-audio-tokenizer
        ),
@@ -2149,6 +2151,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.self_attn.linear_k", # lfm2
            "conformer.layers.{bid}.attention.attn.k_proj", # gemma3n
            "conformer.layers.{bid}.self_attn.k_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.k_proj", # parakeet
            "encoder.layers.{bid}.attn.to_k", # granite_speech (split from to_kv)
            "encoder.layers.{bid}.self_attn.k_proj", # mimo-audio-tokenizer
        ),
@@ -2158,6 +2161,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.self_attn.linear_v", # lfm2
            "conformer.layers.{bid}.attention.attn.v_proj", # gemma3n
            "conformer.layers.{bid}.self_attn.v_proj", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.v_proj", # parakeet
            "encoder.layers.{bid}.attn.to_v", # granite_speech (split from to_kv)
            "encoder.layers.{bid}.self_attn.v_proj", # mimo-audio-tokenizer
        ),
@@ -2187,6 +2191,7 @@ class TensorNameMap:
            "audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
            "conformer.layers.{bid}.norm_self_att", # lfm2
            "conformer.layers.{bid}.attention.pre_attn_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_self_att", # parakeet
            "encoder.layers.{bid}.attn.pre_norm", # granite_speech
            "encoder.layers.{bid}.self_attn_layer_norm", # mimo-audio-tokenizer
        ),
@@ -2196,6 +2201,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.self_attn.linear_out", # lfm2
            "conformer.layers.{bid}.attention.post", # gemma3n
            "conformer.layers.{bid}.self_attn.post", # gemma4
+            "sound_encoder.encoder.layers.{bid}.self_attn.o_proj", # parakeet
            "encoder.layers.{bid}.attn.to_out", # granite_speech
            "encoder.layers.{bid}.self_attn.out_proj", # mimo-audio-tokenizer
        ),
@@ -2204,6 +2210,7 @@ class TensorNameMap:
            "audio_tower.layers.{bid}.final_layer_norm", # ultravox
            "conformer.layers.{bid}.norm_out", # lfm2
            "conformer.layers.{bid}.attention.post_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_out", # parakeet
            "encoder.layers.{bid}.post_norm", # granite_speech
            "encoder.layers.{bid}.final_layer_norm", # mimo-audio-tokenizer
        ),
@@ -2212,6 +2219,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.norm_feed_forward1", # lfm2
            "conformer.layers.{bid}.ffw_layer_start.pre_layer_norm", # gemma3n
            "conformer.layers.{bid}.feed_forward1.pre_layer_norm", # gemma4
+            "sound_encoder.encoder.layers.{bid}.norm_feed_forward1", # parakeet
            "encoder.layers.{bid}.ff1.pre_norm", # granite_speech
        ),

@@ -2229,6 +2237,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.feed_forward1.linear1", # lfm2
            "conformer.layers.{bid}.ffw_layer_start.ffw_layer_1", # gemma3n
            "conformer.layers.{bid}.feed_forward1.ffw_layer_1", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward1.linear1", # parakeet
            "encoder.layers.{bid}.ff1.up_proj", # granite_speech
            "encoder.layers.{bid}.fc1", # mimo-audio-tokenizer
        ),
@@ -2240,6 +2249,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.feed_forward1.linear2", # lfm2
            "conformer.layers.{bid}.ffw_layer_start.ffw_layer_2", # gemma3n
            "conformer.layers.{bid}.feed_forward1.ffw_layer_2", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward1.linear2", # parakeet
            "encoder.layers.{bid}.ff1.down_proj", # granite_speech
            "encoder.layers.{bid}.fc2", # mimo-audio-tokenizer
        ),
@@ -2248,6 +2258,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.feed_forward2.linear1", # lfm2
            "conformer.layers.{bid}.ffw_layer_end.ffw_layer_1", # gemma3n
            "conformer.layers.{bid}.feed_forward2.ffw_layer_1", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward2.linear1", # parakeet
            "encoder.layers.{bid}.ff2.up_proj", # granite_speech
        ),

@@ -2255,6 +2266,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.feed_forward2.linear2", # lfm2
            "conformer.layers.{bid}.ffw_layer_end.ffw_layer_2", # gemma3n
            "conformer.layers.{bid}.feed_forward2.ffw_layer_2", # gemma4
+            "sound_encoder.encoder.layers.{bid}.feed_forward2.linear2", # parakeet
            "encoder.layers.{bid}.ff2.down_proj", # granite_speech
        ),

@@ -2262,6 +2274,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.norm_feed_forward2", # lfm2
            "conformer.layers.{bid}.ffw_layer_end.pre_layer_norm", # gemma3n
            "conformer.layers.{bid}.feed_forward2.pre_layer_norm", # gemma4
+            "sound_encoder.encoder.layers.{bid}.norm_feed_forward2", # parakeet
            "encoder.layers.{bid}.ff2.pre_norm", # granite_speech
        ),

@@ -2290,20 +2303,24 @@ class TensorNameMap:
        MODEL_TENSOR.A_ENC_LINEAR_POS: (
            "conformer.layers.{bid}.self_attn.linear_pos", # lfm2
            "conformer.layers.{bid}.attention.attn.relative_position_embedding.pos_proj", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.self_attn.relative_k_proj", # parakeet
        ),

        MODEL_TENSOR.A_ENC_POS_BIAS_U: (
            "conformer.layers.{bid}.self_attn.pos_bias_u", # lfm2
+            "sound_encoder.encoder.layers.{bid}.self_attn.bias_u", # parakeet
        ),

        MODEL_TENSOR.A_ENC_POS_BIAS_V: (
            "conformer.layers.{bid}.self_attn.pos_bias_v", # lfm2
+            "sound_encoder.encoder.layers.{bid}.self_attn.bias_v", # parakeet
        ),

        MODEL_TENSOR.A_ENC_OUT: (
            "conformer.pre_encode.out", # lfm2
            "model.audio_tower.subsample_conv_projection.input_proj_linear", # gemma3n (note: it should be A_ENC_INP_PROJ, this is a mistake; it should be corrected in C++ code when it's supported)
            "conformer.output_proj", # gemma4
+            "sound_encoder.encoder.subsampling.linear", # parakeet
        ),

        # note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors
@@ -2313,6 +2330,7 @@ class TensorNameMap:
            "audio.multi_modal_projector.linear_{bid}", # ultravox, meralion
            "audio_adapter.model.{bid}", # lfm2
            "audio_tower.proj{bid}", # qwen3omni
+            "sound_projection.linear{bid}", # parakeet (linear1, linear2)
        ),

        MODEL_TENSOR.A_MMPROJ_FC: (
@@ -2323,6 +2341,7 @@ class TensorNameMap:

        MODEL_TENSOR.A_MM_NORM_PRE: (
            "audio.multi_modal_projector.ln_pre", # ultravox
+            "sound_projection.norm", # parakeet
        ),

        MODEL_TENSOR.A_MM_NORM_MID: (
@@ -2368,30 +2387,43 @@ class TensorNameMap:
        MODEL_TENSOR.A_ENC_CONV_DW: (
            "conformer.layers.{bid}.conv.depthwise_conv", # lfm2
            "conformer.layers.{bid}.lconv1d.depthwise_conv1d", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.depthwise_conv", # parakeet
            "encoder.layers.{bid}.conv.depth_conv.conv", # granite_speech
        ),

        MODEL_TENSOR.A_ENC_CONV_NORM: (
            "conformer.layers.{bid}.conv.batch_norm", # lfm2
            "conformer.layers.{bid}.lconv1d.pre_layer_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.norm", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV_NORM_MEAN: (
+            "sound_encoder.encoder.layers.{bid}.conv.norm.running_mean", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV_NORM_VAR: (
+            "sound_encoder.encoder.layers.{bid}.conv.norm.running_var", # parakeet
            "encoder.layers.{bid}.conv.batch_norm", # granite_speech
        ),

        MODEL_TENSOR.A_ENC_CONV_PW1: (
            "conformer.layers.{bid}.conv.pointwise_conv1", # lfm2
            "conformer.layers.{bid}.lconv1d.linear_start", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv1", # parakeet
            "encoder.layers.{bid}.conv.up_conv", # granite_speech
        ),

        MODEL_TENSOR.A_ENC_CONV_PW2: (
            "conformer.layers.{bid}.conv.pointwise_conv2", # lfm2
            "conformer.layers.{bid}.lconv1d.linear_end", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv2", # parakeet
            "encoder.layers.{bid}.conv.down_conv", # granite_speech
        ),

        MODEL_TENSOR.A_ENC_NORM_CONV: (
            "conformer.layers.{bid}.norm_conv", # lfm2
            "conformer.layers.{bid}.lconv1d.conv_norm", # gemma3n
+            "sound_encoder.encoder.layers.{bid}.norm_conv", # parakeet
            "encoder.layers.{bid}.conv.norm", # granite_speech
        ),

@@ -2403,6 +2435,14 @@ class TensorNameMap:
            "conformer.layers.{bid}.attention.attn.per_dim_scale", # gemma4
        ),

+        MODEL_TENSOR.A_ENC_MEL_FILTERS: (
+            "sound_encoder.encoder.feature_extractor.featurizer.fb", # parakeet
+        ),
+
+        MODEL_TENSOR.A_ENC_WINDOW: (
+            "sound_encoder.encoder.feature_extractor.featurizer.window", # parakeet
+        ),
+
        MODEL_TENSOR.A_MM_EMBEDDING: (
            "model.embed_audio.embedding", # gemma3n
        ),
@@ -478,6 +478,7 @@ enum test_mode {
    MODE_PERF,
    MODE_GRAD,
    MODE_SUPPORT,
+    MODE_TUNE,
 };

 // Output format support similar to llama-bench
@@ -10111,8 +10112,482 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_from_file(const c
    return test_cases;
 }

+// ---- FA vec (Q,NE) tuning: numerical correctness check + drift guard ----
+// metal proc_address bridges (resolved by string, not symbol linkage)
+using set_fa_vec_override_t   = void (*)(int, int);
+using clear_fa_vec_override_t = void (*)(void);
+using fa_vec_bucket_t         = int  (*)(int64_t);  // runtime ne11/ne01 bucketers, shared via proc bridge
+using fa_vec_baseline_ne_t    = int  (*)(int, int); // runtime (dk,dv) -> baseline NE
+
+// legal NE for a (dk,dv): NL = 32/NE, require (dk/4)%NL==0 && (dv/4)%NL==0
+static std::vector<int> fa_vec_legal_ne(int dk, int dv) {
+    std::vector<int> r;
+    for (int ne : {1, 2, 4}) {
+        const int nl = 32 / ne;
+        if ((dk/4) % nl == 0 && (dv/4) % nl == 0) {
+            r.push_back(ne);
+        }
+    }
+    return r;
+}
+
+// Baseline-path smoke test: with no override, fa_vec_pick returns (1, baseline_ne) and
+// dispatches the no-suffix baseline kernel. Checks baseline correctness across all 10 shapes
+// (NE is numerically transparent; the (Q,NE) kernels are covered by run_fa_vec_tune_check).
+static bool run_fa_vec_drift_guard(ggml_backend_t backend_metal, ggml_backend_t backend_cpu) {
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { {32,32},{64,64},{96,96},{128,128},{192,192},
+                               {192,128},{256,256},{320,256},{512,512},{576,512} };
+    bool ok = true;
+    for (auto s : shapes) {
+        // no override + short KV (ne11 < FA_VEC_NE11_BUCKETS[0]) -> fa_vec_pick returns baseline;
+        // compare metal vs CPU-ref
+        test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/512, /*nb=*/8,
+                               /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                               GGML_TYPE_F16, GGML_TYPE_F16);
+        auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+        if (st == test_status_t::FAIL) {
+            printf("DRIFT/baseline FAIL dk=%d dv=%d\n", s.dk, s.dv);
+            ok = false;
+        }
+    }
+    return ok;
+}
+
+// Numerical gate: for each (dk,dv) x legal (Q,NE), force the override and compare metal
+// vs CPU-ref. The ne01/ne11/mask/sinks points exercise the rebased body's padded rows
+// (ne01 not a multiple of Q), per-qq sinks, skip-INF, kvpad, and the nsg-dependent shmem
+// offsets / parallel-reduce stride (ne11 -> nsg 1/2/4).
+static bool run_fa_vec_tune_check(ggml_backend_t backend_metal, ggml_backend_t backend_cpu) {
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_metal));
+    auto set_ov   = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto clear_ov = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    if (!set_ov || !clear_ov) { printf("metal fa_vec override proc unavailable\n"); return false; }
+
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { {32,32},{64,64},{96,96},{128,128},{192,192},
+                               {192,128},{256,256},{320,256},{512,512},{576,512} };
+    const int  ne01_pts[]   = { 1, 2, 3, 8, 17 };       // 1, Q+/-1, 2Q-1, prime; vec upper bound < 20
+    const int  ne11_pts[]   = { 512, 4096, 8192 };      // drives adaptive nsg 1/2/4
+    const int  ne11_kvpad[] = { 4097, 8193 };           // has_kvpad (non-32-multiple kv)
+    const bool mask_pts[]   = { true, false };
+    const bool sinks_pts[]  = { false, true };
+
+    bool ok = true;
+    int n_run = 0, n_fail = 0;
+    for (auto s : shapes) {
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : {1, 2, 4}) {
+                for (bool mask : mask_pts) {
+                    for (bool sinks : sinks_pts) {
+                        for (int ne01 : ne01_pts) {
+                            for (int ne11 : ne11_pts) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/ne11, /*nb=*/ne01,
+                                                       mask, sinks, 0.0f, 0.0f, GGML_PREC_F32,
+                                                       GGML_TYPE_F16, GGML_TYPE_F16);
+                                auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+                                if (st == test_status_t::FAIL) {
+                                    printf("FAIL dk=%d dv=%d Q=%d ne=%d ne01=%d ne11=%d mask=%d sinks=%d\n",
+                                           s.dk, s.dv, Q, ne, ne01, ne11, (int) mask, (int) sinks);
+                                    ok = false; n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    // kvpad pass (non-32-multiple kv); dk128/256 only to keep the case count down
+    for (auto s : shapes) {
+        if (!((s.dk == 128 && s.dv == 128) || (s.dk == 256 && s.dv == 256))) {
+            continue;
+        }
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : {1, 2, 4}) {
+                for (int ne11 : ne11_kvpad) {
+                    set_ov(Q, ne);
+                    test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1,1}, /*kv=*/ne11, /*nb=*/8,
+                                           /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                                           GGML_TYPE_F16, GGML_TYPE_F16);
+                    auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                    clear_ov();
+                    if (st == test_status_t::FAIL) {
+                        printf("FAIL(kvpad) dk=%d dv=%d Q=%d ne=%d ne11=%d\n", s.dk, s.dv, Q, ne, ne11);
+                        ok = false; n_fail++;
+                    }
+                    n_run++;
+                }
+            }
+        }
+    }
+    // quantized K/V numerical check: the rebased body's dequant path is Q-generic
+    // (dequant once, reuse across Q rows); confirm it stays correct for every quant precision.
+    const ggml_type qtypes[] = { GGML_TYPE_Q4_0, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0 };
+    const int  q_ne01[] = { 1, 2, 3, 8 };
+    const int  q_ne11[] = { 512, 4096, 8192 };
+    for (ggml_type qt : qtypes) {
+        for (auto s : shapes) {
+            for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+                for (int Q : { 1, 2, 4 }) {
+                    for (bool sinks : { false, true }) {
+                        for (int ne01 : q_ne01) {
+                            for (int ne11 : q_ne11) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, {1, 1}, /*kv=*/ne11, /*nb=*/ne01,
+                                                       /*mask=*/true, sinks, 0.0f, 0.0f, GGML_PREC_F32, qt, qt);
+                                auto st = tc.eval(backend_metal, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+                                if (st == test_status_t::FAIL) {
+                                    printf("FAIL(quant) type=%s dk=%d dv=%d Q=%d ne=%d ne01=%d ne11=%d sinks=%d\n",
+                                           ggml_type_name(qt), s.dk, s.dv, Q, ne, ne01, ne11, (int) sinks);
+                                    ok = false; n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    printf("fa_vec tune-check: %d cases run, %d failed\n", n_run, n_fail);
+    return ok;
+}
+
+// A prebuilt FA op graph for one (dk,dv,ne01,ne11) cell. The op is replicated n_runs
+// times so a single graph_compute amortizes dispatch/sync overhead. The graph is reused
+// across (Q,NE) overrides (the override only changes the pipeline picked at encode time),
+// which avoids re-allocating/re-initializing the (large) K/V tensors per candidate.
+struct fa_perf_cell {
+    ggml_context_ptr        ctx;
+    ggml_backend_buffer_ptr buf;
+    ggml_cgraph *           gf     = nullptr;
+    int                     n_runs = 0;
+    bool                    ok     = false;
+};
+
+static fa_perf_cell fa_build_perf_cell(ggml_backend_t backend, int dk, int dv, int ne01, int ne11,
+                                       ggml_type type_kv = GGML_TYPE_F16) {
+    fa_perf_cell cell;
+
+    // GQA shape (nr23=[8,1]) matching real spec-decode / verify workloads: enough query
+    // heads to keep the GPU busy so the Q>1 K/V-reuse benefit is visible (and comparable
+    // to the documented #23114 numbers). nh here is the number of KV heads.
+    test_flash_attn_ext tc(dk, dv, /*nh=*/4, { 8, 1 }, /*kv=*/ne11, /*nb=*/ne01,
+                           /*mask=*/true, /*sinks=*/false, 0.0f, 0.0f, GGML_PREC_F32,
+                           type_kv, type_kv);
+
+    const size_t graph_nodes = 1024;
+    ggml_init_params params = {
+        /* .mem_size  = */ ggml_tensor_overhead() * 128 + ggml_graph_overhead_custom(graph_nodes, false),
+        /* .mem_base  = */ NULL,
+        /* .no_alloc  = */ true,
+    };
+    cell.ctx.reset(ggml_init(params));
+    GGML_ASSERT(cell.ctx);
+
+    ggml_tensor * out = tc.build_graph(cell.ctx.get());
+    if (!ggml_backend_supports_op(backend, out)) {
+        return cell;
+    }
+
+    cell.buf.reset(ggml_backend_alloc_ctx_tensors(cell.ctx.get(), backend));
+    if (cell.buf == NULL) {
+        return cell;
+    }
+
+    tc.initialize_tensors(cell.ctx.get());
+
+    cell.gf = ggml_new_graph_custom(cell.ctx.get(), graph_nodes, false);
+    ggml_build_forward_expand(cell.gf, out);
+
+    // replicate the op to amortize overhead (target ~50 GFLOP/compute, capped to bound graph size)
+    cell.n_runs = 1;
+    if (tc.op_flops(out) > 0) {
+        const uint64_t target_flops = 50ULL * 1000 * 1000 * 1000;
+        const int      cap          = 512;
+        const int      by_flops     = (int) std::min<int64_t>(cap, (int64_t) (target_flops / tc.op_flops(out)));
+        cell.n_runs = std::max(1, std::min<int>(by_flops, (int) (ggml_graph_size(cell.gf) - ggml_graph_n_nodes(cell.gf))));
+    }
+    for (int i = 1; i < cell.n_runs; ++i) {
+        ggml_graph_add_node(cell.gf, out);
+    }
+    cell.ok = true;
+    return cell;
+}
+
+// Median per-op GPU time (us) for the currently-set override over the prebuilt cell graph.
+static double time_fa_cell_median(ggml_backend_t backend, const fa_perf_cell & cell, int reps) {
+    if (!cell.ok) {
+        return -1.0;
+    }
+    ggml_backend_graph_compute(backend, cell.gf);  // warmup (compiles the pipeline for the override)
+    ggml_backend_synchronize(backend);
+
+    std::vector<double> samples;
+    samples.reserve(reps);
+    for (int r = 0; r < reps; ++r) {
+        const int64_t t0 = ggml_time_us();
+        ggml_backend_graph_compute(backend, cell.gf);
+        ggml_backend_synchronize(backend);
+        samples.push_back((double) (ggml_time_us() - t0));
+    }
+    std::nth_element(samples.begin(), samples.begin() + samples.size() / 2, samples.end());
+    return samples[samples.size() / 2] / cell.n_runs;
+}
+
+// Perf sweep over the (Q,NE) grid, emitting pasteable fa_vec_tuned_table rows.
+// nsg/nwg are left to the ops.cpp adaptive heuristic (not part of the table).
+static bool run_fa_vec_tune_perf(ggml_backend_t backend_metal) {
+    auto * reg    = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_metal));
+    auto   set_ov      = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto   clr_ov      = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    // Share the runtime's bucketers + baseline NE via read-only proc bridges, so the emitted
+    // (ne11_b, ne01_b) keys match fa_vec_pick and can't silently desync from FA_VEC_*_BUCKETS.
+    auto   ne11_bucket = (fa_vec_bucket_t)      ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_ne11_bucket");
+    auto   ne01_bucket = (fa_vec_bucket_t)      ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_ne01_bucket");
+    auto   baseline_ne = (fa_vec_baseline_ne_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_fa_vec_baseline_ne");
+    if (!set_ov || !clr_ov || !ne11_bucket || !ne01_bucket || !baseline_ne) {
+        printf("metal fa_vec tuning procs unavailable\n");
+        return false;
+    }
+
+    struct shape_t { int dk, dv; };
+    const shape_t shapes[] = { { 32, 32 }, { 64, 64 }, { 96, 96 }, { 128, 128 }, { 192, 192 },
+                               { 192, 128 }, { 256, 256 }, { 320, 256 }, { 512, 512 }, { 576, 512 } };
+    const int ne11_rep[] = { 512, 2048, 8192, 32768 };  // ne11 bucket representatives
+    const int ne01_rep[] = { 1, 2, 3, 4, 5, 6, 7, 8, 16 };  // PR grid: point-bucket reps (1-4) + tail mod-4 cycle (5-8) + large anchor (16); vec serves ne01<20
+    const int REPS = 7;                                 // odd -> exact median
+
+    printf("# fa_vec perf sweep — replace GGML_METAL_DEVICE_M4_MAX with this machine's device\n");
+    printf("# [1] per-bucket timings (us, winner *): '=> cfg Nx' beats baseline, else '=> baseline'\n");
+    printf("# [2] a pasteable fa_vec_tuned_table block (domain defaults + exceptions) is printed after [1]\n");
+
+    struct cand_t { int Q, NE; double t; };  // one timed (Q,NE) candidate
+    struct pt_t   { int dk, dv, ne11, ne01;   // one swept grid point with its candidate times
+                    std::vector<cand_t> cs; double base_t; };
+
+    // group_split compression knobs (see ggml-metal-tuning.h for the row / lookup semantics)
+    const double TUNE_TAU   = 0.05;  // max POINTWISE regret (worst point in a bucket) to ride a domain
+                                     // default instead of emitting the bucket's own exception row
+    const double TUNE_THETA = 1.05;  // min AGGREGATE bucket speedup vs baseline to tune the bucket at all
+
+    auto type_token = [](ggml_type t) -> const char * {
+        switch (t) {
+            case GGML_TYPE_Q4_0: return "GGML_TYPE_Q4_0";
+            case GGML_TYPE_Q4_1: return "GGML_TYPE_Q4_1";
+            case GGML_TYPE_Q5_0: return "GGML_TYPE_Q5_0";
+            case GGML_TYPE_Q5_1: return "GGML_TYPE_Q5_1";
+            case GGML_TYPE_Q8_0: return "GGML_TYPE_Q8_0";
+            default:             return "GGML_TYPE_F16";
+        }
+    };
+
+    const ggml_type types[] = { GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
+                                GGML_TYPE_Q5_0, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0 };
+    for (ggml_type type_kv : types) {
+        printf("\n### dtype=%s\n", ggml_type_name(type_kv));
+        std::vector<pt_t> pts;  // one per swept (shape, ne11, ne01); bucketed + compressed below
+
+        for (auto s : shapes) {
+            const int base_ne = baseline_ne(s.dk, s.dv);
+            const std::vector<int> legal = fa_vec_legal_ne(s.dk, s.dv);
+            for (int ne11 : ne11_rep) {
+                for (int ne01 : ne01_rep) {
+                    fa_perf_cell cell = fa_build_perf_cell(backend_metal, s.dk, s.dv, ne01, ne11, type_kv);
+                    if (!cell.ok) {
+                        continue;
+                    }
+
+                    std::vector<cand_t> cs;
+                    for (int ne : legal) {
+                        for (int Q : { 1, 2, 4 }) {
+                            cs.push_back({ Q, ne, 0.0 });
+                        }
+                    }
+                    // randomize config order to decorrelate thermal throttling across the cell
+                    std::shuffle(cs.begin(), cs.end(), std::mt19937(1234));
+
+                    double anchor = 0.0;  // periodic baseline re-measure -> throttling detection
+                    for (size_t i = 0; i < cs.size(); ++i) {
+                        set_ov(cs[i].Q, cs[i].NE);
+                        cs[i].t = time_fa_cell_median(backend_metal, cell, REPS);
+                        clr_ov();
+
+                        if (i % 4 == 0) {
+                            const double a = time_fa_cell_median(backend_metal, cell, REPS);
+                            if (anchor > 0.0) {
+                                double drift = (a - anchor) / anchor;
+                                if (drift < 0) {
+                                    drift = -drift;
+                                }
+                                if (drift > 0.10) {
+                                    printf("# WARN throttling? anchor drift %.1f%% dk=%d ne11=%d\n", 100.0 * drift, s.dk, ne11);
+                                }
+                            }
+                            anchor = (anchor > 0.0) ? std::min(anchor, a) : a;
+                        }
+                    }
+
+                    std::sort(cs.begin(), cs.end(), [](const cand_t & a, const cand_t & b) {
+                        return a.Q != b.Q ? a.Q < b.Q : a.NE < b.NE;
+                    });
+                    cand_t best = cs[0];
+                    for (const auto & c : cs) {
+                        if (c.t > 0.0 && (best.t <= 0.0 || c.t < best.t)) {
+                            best = c;
+                        }
+                    }
+                    double base_t = 0.0;  // the swept (Q=1, base_ne) candidate == baseline kernel
+                    for (const auto & c : cs) {
+                        if (c.Q == 1 && c.NE == base_ne) { base_t = c.t; break; }
+                    }
+                    const bool keep = best.t > 0.0 && base_t > 0.0 && best.t < base_t * 0.98;
+
+                    printf("# dtype=%s dk=%d dv=%d ne11=%d ne01=%d:", ggml_type_name(type_kv), s.dk, s.dv, ne11, ne01);
+                    for (const auto & c : cs) {
+                        printf("  Q%dNE%d=%.1f%s", c.Q, c.NE, c.t, (c.Q == best.Q && c.NE == best.NE) ? "*" : "");
+                    }
+                    if (keep) {
+                        printf("  => Q%d,NE%d  %.2fx\n", best.Q, best.NE, base_t / best.t);
+                    } else {
+                        printf("  => baseline\n");
+                    }
+                    pts.push_back({ s.dk, s.dv, ne11, ne01, cs, base_t });
+                }
+            }
+        }
+
+        // [2] Compress into pasteable rows. Per (dk,dv) and ne01 domain {decode==1, batch>=2}, emit
+        // one ne11-collapsed default cfg (fewest rows, ne11_b=-1) plus a per-bucket exception wherever
+        // the default's pointwise regret vs the bucket's min-max-regret target, or its aggregate
+        // slowdown vs baseline, exceeds TUNE_TAU. The target still carries an intrinsic per-point regret
+        // from lumping ne01 into one tail bucket — TUNE_TAU bounds resolved-vs-target, not vs-oracle.
+        std::vector<std::string> rows_out;
+        char rbuf[192];
+        for (auto s : shapes) {
+            const int base_ne = baseline_ne(s.dk, s.dv);
+            std::vector<cand_t> cfgs;  // candidate list (identical for every grid point of this shape)
+            int base_i = 0;
+            for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+                for (int Q : { 1, 2, 4 }) {
+                    if (Q == 1 && ne == base_ne) { base_i = (int) cfgs.size(); }
+                    cfgs.push_back({ Q, ne, 0.0 });
+                }
+            }
+            auto cfg_time = [&](const pt_t & p, int i) {
+                for (const auto & c : p.cs) { if (c.Q == cfgs[i].Q && c.NE == cfgs[i].NE) { return c.t; } }
+                return 0.0;
+            };
+
+            // Bucket the grid points using the runtime's bucketers (via proc), so keys match fa_vec_pick.
+            // Short-KV points (ne11 bucket 0) are dropped — the runtime serves those from baseline. Each
+            // kept bucket picks a min-max-regret target (gated by the aggregate TUNE_THETA benefit gate).
+            struct bkt_t { int b11, b01, Ti; std::vector<double> agg; double base_agg;
+                           std::vector<const pt_t *> bp; };  // bp kept for the pointwise ride test below
+            std::set<std::pair<int, int>> seen;
+            for (const auto & p : pts) {
+                if (p.dk != s.dk || p.dv != s.dv) { continue; }
+                const int b11 = ne11_bucket(p.ne11);
+                if (b11 == 0) { continue; }
+                seen.insert({ b11, ne01_bucket(p.ne01) });
+            }
+            std::vector<bkt_t> bks;
+            for (const auto & bb : seen) {
+                const int b11 = bb.first, b01 = bb.second;
+                std::vector<const pt_t *> bp;
+                for (const auto & p : pts) {
+                    if (p.dk == s.dk && p.dv == s.dv && ne11_bucket(p.ne11) == b11 && ne01_bucket(p.ne01) == b01) {
+                        bp.push_back(&p);
+                    }
+                }
+                std::vector<double> agg(cfgs.size(), 0.0), worst(cfgs.size(), 0.0);
+                for (const auto * p : bp) {
+                    double bestt = 0.0;
+                    for (size_t i = 0; i < cfgs.size(); ++i) {
+                        double t = cfg_time(*p, (int) i);
+                        if (t > 0.0 && (bestt == 0.0 || t < bestt)) { bestt = t; }
+                    }
+                    for (size_t i = 0; i < cfgs.size(); ++i) {
+                        double t = cfg_time(*p, (int) i);
+                        agg[i] += t;
+                        if (t > 0.0 && bestt > 0.0) { worst[i] = std::max(worst[i], t / bestt); }
+                    }
+                }
+                int robust = 0;
+                for (size_t i = 1; i < cfgs.size(); ++i) {
+                    if (worst[i] < worst[robust] ||
+                        (worst[i] == worst[robust] && (cfgs[i].Q < cfgs[robust].Q ||
+                        (cfgs[i].Q == cfgs[robust].Q && cfgs[i].NE < cfgs[robust].NE)))) {
+                        robust = (int) i;
+                    }
+                }
+                const bool tune = robust != base_i && agg[base_i] / agg[robust] >= TUNE_THETA;
+                bks.push_back({ b11, b01, tune ? robust : base_i, agg, agg[base_i], bp });
+            }
+
+            // Pointwise regret of default cfg d vs the bucket target Ti. Per-point (not aggregate):
+            // a ratio-of-sums lets a default that wins on aligned ne01 (8, 16) hide a large penalty
+            // on a misaligned point (ne01=5), so the ride/exception decision must be per point.
+            auto reg_pointwise = [&](const bkt_t * b, int d) {
+                double r = 0.0;
+                for (const auto * p : b->bp) {
+                    const double td = cfg_time(*p, d), tT = cfg_time(*p, b->Ti);
+                    if (td > 0.0 && tT > 0.0) { r = std::max(r, td / tT - 1.0); }
+                }
+                return r;
+            };
+
+            for (int dom = 0; dom <= 1; ++dom) {  // 0 = decode (ne01==1), 1 = batch (ne01>=2)
+                std::vector<const bkt_t *> db;
+                for (const auto & b : bks) { if ((dom == 0) == (b.b01 == 0)) { db.push_back(&b); } }
+                if (db.empty()) { continue; }
+
+                // default cfg = the one minimizing (#rows, total achieved time, Q, NE)
+                int bestD = -1, bestRows = 1 << 30; double bestTot = 0.0;
+                for (size_t d = 0; d < cfgs.size(); ++d) {
+                    int rows = ((int) d != base_i) ? 1 : 0; double tot = 0.0;
+                    for (const auto * b : db) {
+                        const double reg  = reg_pointwise(b, (int) d);       // pointwise vs bucket target
+                        const double slow = b->agg[d] / b->base_agg - 1.0;   // aggregate vs baseline (safety net)
+                        if (reg > TUNE_TAU || slow > TUNE_TAU) { rows++; tot += b->agg[b->Ti]; }
+                        else                                   {         tot += b->agg[d];      }
+                    }
+                    const bool better = bestD < 0 || rows < bestRows ||
+                        (rows == bestRows && (tot < bestTot ||
+                        (tot == bestTot && (cfgs[d].Q < cfgs[bestD].Q ||
+                        (cfgs[d].Q == cfgs[bestD].Q && cfgs[d].NE < cfgs[bestD].NE)))));
+                    if (better) { bestD = (int) d; bestRows = rows; bestTot = tot; }
+                }
+
+                const int dom_id = (dom == 0) ? 0 : 1;  // FA_VEC_DOMAIN_DECODE / FA_VEC_DOMAIN_BATCH
+                if (bestD != base_i) {
+                    snprintf(rbuf, sizeof(rbuf), "    { { GGML_METAL_DEVICE_M4_MAX, %s, %d, %d, -1, %d }, { %d, %d } },",
+                             type_token(type_kv), s.dk, s.dv, dom_id, cfgs[bestD].Q, cfgs[bestD].NE);
+                    rows_out.emplace_back(rbuf);
+                }
+                for (const auto * b : db) {
+                    const double reg  = reg_pointwise(b, bestD);            // pointwise vs bucket target
+                    const double slow = b->agg[bestD] / b->base_agg - 1.0;  // aggregate vs baseline
+                    if (reg <= TUNE_TAU && slow <= TUNE_TAU) { continue; }  // rides the default / baseline
+                    snprintf(rbuf, sizeof(rbuf), "    { { GGML_METAL_DEVICE_M4_MAX, %s, %d, %d, %d, %d }, { %d, %d } },",
+                             type_token(type_kv), s.dk, s.dv, b->b11, b->b01, cfgs[b->Ti].Q, cfgs[b->Ti].NE);
+                    rows_out.emplace_back(rbuf);
+                }
+            }
+        }
+        printf("\n    // ---- %s: %zu rows ----\n", ggml_type_name(type_kv), rows_out.size());
+        for (const auto & r : rows_out) { printf("%s\n", r.c_str()); }
+    }  // for type_kv
+    return true;
+}
+
 static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mode mode, const char * op_names_filter, const char * params_filter,
-                         printer * output_printer, const char * test_file_path, int parallel_workers) {
+                         printer * output_printer, const char * test_file_path, int parallel_workers, bool tune_perf = false) {
    auto filter_test_cases = [](std::vector<std::unique_ptr<test_case>> & test_cases, const char * params_filter) {
        if (params_filter == nullptr) {
            return;
@@ -10142,6 +10617,9 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
        case MODE_PERF:
            test_cases = make_test_cases_perf();
            break;
+        case MODE_TUNE:
+            // MODE_TUNE routes to its own dispatch below; no generic test_cases needed
+            break;
        }
    } else {
        test_cases = make_test_cases_from_file(test_file_path);
@@ -10277,6 +10755,29 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
        return true;
    }

+    if (mode == MODE_TUNE) {
+        // self-create a CPU backend with the reference implementation as golden.
+        // (backend_cpu in the MODE_TEST block above is out of scope here.)
+        ggml_backend_t backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, NULL);
+        GGML_ASSERT(backend_cpu != NULL);
+        {
+            using ggml_backend_cpu_set_use_ref_t = void (*)(ggml_backend_t, bool);
+            auto * cpu_reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
+            auto * set_use_ref = (ggml_backend_cpu_set_use_ref_t) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_cpu_set_use_ref");
+            if (set_use_ref) {
+                set_use_ref(backend_cpu, true);
+            }
+        }
+
+        // backend here is the MODE_TUNE-selected metal backend (-b MTL0)
+        const bool ok = tune_perf
+            ? run_fa_vec_tune_perf(backend)
+            : (run_fa_vec_drift_guard(backend, backend_cpu) && run_fa_vec_tune_check(backend, backend_cpu));
+
+        ggml_backend_free(backend_cpu);
+        return ok;
+    }
+
    if (mode == MODE_SUPPORT) {
        // Filter out fusion cases
        test_cases.erase(
@@ -10402,6 +10903,7 @@ static void usage(char ** argv) {
    printf("      - grad (compare gradients from backpropagation with method of finite differences)\n");
    printf("      - perf (performance evaluation)\n");
    printf("      - support (probe backend operation support)\n");
+    printf("      - tune (FA vec (Q,NE) numerical correctness check vs CPU reference; --tune-perf for the perf sweep)\n");
    printf("    op names for -o are as given by ggml_op_desc() (e.g. ADD, MUL_MAT, etc),\n");
    printf("        optionally including the full test case string (e.g. \"ADD(type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1)\")\n");
    printf("    --output specifies output format (default: console, options: console, sql, csv)\n");
@@ -10419,6 +10921,7 @@ int main(int argc, char ** argv) {
    const char * params_filter = nullptr;
    const char * test_file_path = nullptr;
    int parallel_workers = 1;
+    bool tune_perf = false;

    for (int i = 1; i < argc; i++) {
        if (strcmp(argv[i], "test") == 0) {
@@ -10429,6 +10932,10 @@ int main(int argc, char ** argv) {
            mode = MODE_GRAD;
        } else if (strcmp(argv[i], "support") == 0) {
            mode = MODE_SUPPORT;
+        } else if (strcmp(argv[i], "tune") == 0) {
+            mode = MODE_TUNE;
+        } else if (strcmp(argv[i], "--tune-perf") == 0) {
+            tune_perf = true;
        } else if (strcmp(argv[i], "-o") == 0) {
            if (i + 1 < argc) {
                op_names_filter = argv[++i];
@@ -10536,7 +11043,7 @@ int main(int argc, char ** argv) {
                                                             false, "", ggml_backend_dev_description(dev),
                                                             total / 1024 / 1024, free / 1024 / 1024, true));

-        bool ok = test_backend(backend.get(), dev, mode, op_names_filter, params_filter, output_printer.get(), test_file_path, parallel_workers);
+        bool ok = test_backend(backend.get(), dev, mode, op_names_filter, params_filter, output_printer.get(), test_file_path, parallel_workers, tune_perf);

        if (ok) {
            n_ok++;
@@ -60,6 +60,7 @@ add_library(mtmd
            models/mobilenetv5.cpp
            models/youtuvl.cpp
            models/yasa2.cpp
+            models/parakeet.cpp
            )

 set_target_properties(mtmd PROPERTIES
@@ -88,6 +88,7 @@
 #define KEY_A_ATTN_WINDOW_SIZE     "clip.audio.window_size"          // mimo-audio-tokenizer: sliding-window radius
 #define KEY_A_LOCAL_BLOCK_COUNT    "clip.audio.local_block_count"    // mimo-v2.5: input_local_transformer layer count
 #define KEY_A_LOCAL_GROUP_SIZE     "clip.audio.local_group_size"     // mimo-v2.5: input_local_transformer grouping size
+#define KEY_AUDIO_SUBSAMPLING_FACTOR "clip.audio.subsampling_factor"

 //
 // tensor name constants
@@ -338,6 +339,12 @@
 #define TN_YASA_STAGE_DOWN_CONV  "v.stage.%d.down.conv.%s"
 #define TN_YASA_STAGE_BLK        "v.stage.%d.blk.%d.%s.%s"

+// parakeet
+#define TN_MEL_FILTERS           "a.mel_filters"
+#define TN_WINDOW                "a.window"
+#define TN_CONV_NORM_MEAN        "%s.blk.%d.conv_norm_mean"
+#define TN_CONV_NORM_VAR         "%s.blk.%d.conv_norm_var"
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))

@@ -392,6 +399,7 @@ enum projector_type {
    PROJECTOR_TYPE_KIMIK25,
    PROJECTOR_TYPE_NEMOTRON_V2_VL,
    PROJECTOR_TYPE_HUNYUANVL,
+    PROJECTOR_TYPE_PARAKEET,
    PROJECTOR_TYPE_EXAONE4_5,
    PROJECTOR_TYPE_MINICPMV4_6,
    PROJECTOR_TYPE_GRANITE_SPEECH,
@@ -455,6 +463,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
    { PROJECTOR_TYPE_MINIMAX_M3,        "minimax_m3"},
    { PROJECTOR_TYPE_GRANITE4_VISION,   "granite4_vision"},
    { PROJECTOR_TYPE_MIMO_AUDIO,        "mimo_audio"},
+    { PROJECTOR_TYPE_PARAKEET,          "parakeet"},
 };

 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -110,6 +110,8 @@ struct clip_hparams {
    // audio
    int32_t n_mel_bins = 0; // whisper preprocessor
    int32_t proj_stack_factor = 0; // ultravox
+    int32_t subsampling_factor = 0; // parakeet
+
    int32_t audio_chunk_size           = 0;
    int32_t audio_conv_kernel_size     = 0;
    int32_t audio_max_pos_emb          = 0;
@@ -124,6 +126,10 @@ struct clip_hparams {
    int32_t audio_window_len  = -1;
    int32_t audio_hop_len     = -1;

+    // parakeet
+    std::vector<float> mel_filters;
+    std::vector<float> window;
+
    // mimo-audio-tokenizer: residual vector quantizer
    int32_t rvq_num_quantizers = 0;
    std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)
@@ -245,14 +251,16 @@ struct clip_layer {
    ggml_tensor * norm_conv_b   = nullptr;
    ggml_tensor * linear_pos_w  = nullptr;

-    ggml_tensor * conv_norm_w   = nullptr;
-    ggml_tensor * conv_norm_b   = nullptr;
-    ggml_tensor * conv_dw_w     = nullptr;
-    ggml_tensor * conv_dw_b     = nullptr;
-    ggml_tensor * conv_pw1_w    = nullptr;
-    ggml_tensor * conv_pw1_b    = nullptr;
-    ggml_tensor * conv_pw2_w    = nullptr;
-    ggml_tensor * conv_pw2_b    = nullptr;
+    ggml_tensor * conv_norm_w    = nullptr;
+    ggml_tensor * conv_norm_b    = nullptr;
+    ggml_tensor * conv_norm_mean = nullptr;  // parakeet
+    ggml_tensor * conv_norm_var  = nullptr;  // parakeet
+    ggml_tensor * conv_dw_w      = nullptr;
+    ggml_tensor * conv_dw_b      = nullptr;
+    ggml_tensor * conv_pw1_w     = nullptr;
+    ggml_tensor * conv_pw1_b     = nullptr;
+    ggml_tensor * conv_pw2_w     = nullptr;
+    ggml_tensor * conv_pw2_b     = nullptr;

    // gemma4 audio conformer per-layer
    ggml_tensor * attn_pre_norm_w   = nullptr;
@@ -1033,6 +1033,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
            {
                builder = std::make_unique<clip_graph_yasa2>(ctx, img);
            } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                builder = std::make_unique<clip_graph_parakeet>(ctx, img);
+            } break;
        case PROJECTOR_TYPE_GRANITE4_VISION:
            {
                builder = std::make_unique<clip_graph_granite4_vision>(ctx, img);
@@ -1356,6 +1360,20 @@ struct clip_model_loader {
                    {
                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                    } break;
+                case PROJECTOR_TYPE_PARAKEET:
+                    {
+                        get_u32(KEY_AUDIO_SUBSAMPLING_FACTOR, hparams.subsampling_factor);
+                        GGML_ASSERT(hparams.subsampling_factor == 8 &&
+                            "subsampling_factor must match the conv strides in clip_graph_parakeet::build()");
+                        get_u32(KEY_A_CONV_KERNEL_SIZE,       hparams.audio_conv_kernel_size);
+                        GGML_ASSERT(hparams.audio_conv_kernel_size > 0 && hparams.audio_conv_kernel_size % 2 == 1 &&
+                            "audio_conv_kernel_size must be a positive odd integer");
+                        hparams.audio_chunk_len    = 0;
+                        hparams.audio_sample_rate  = 16000;
+                        hparams.audio_n_fft        = 512;
+                        hparams.audio_window_len   = 400;
+                        hparams.audio_hop_len      = 160;
+                    } break;
                case PROJECTOR_TYPE_IDEFICS3:
                    {
                        // use default llava-uhd preprocessing params
@@ -1893,16 +1911,46 @@ struct clip_model_loader {
            return cur;
        };

-        auto get_scalar = [&](const std::string & name, float default_val) {
+        auto get_vector = [&](const std::string & name) {
+            std::vector<float> result;
            auto it = tensor_offset.find(name);
            if (it == tensor_offset.end()) {
+                return result;
+            }
+
+            const int64_t idx = gguf_find_tensor(ctx_gguf.get(), name.c_str());
+            if (idx < 0) {
+                throw std::runtime_error(string_format("%s: failed to find tensor %s\n", __func__, name.c_str()));
+            }
+
+            if (const auto type = gguf_get_tensor_type(ctx_gguf.get(), idx); type != GGML_TYPE_F32) {
+                throw std::runtime_error(string_format("%s: %s must be %s, was %s\n", __func__,
+                            name.c_str(), ggml_type_name(GGML_TYPE_F32), ggml_type_name(type)));
+            }
+
+            const size_t n_bytes = gguf_get_tensor_size(ctx_gguf.get(), idx);
+            if (n_bytes == 0) {
+                throw std::runtime_error(string_format("%s: tensor %s is empty\n", __func__, name.c_str()));
+            }
+
+            const size_t n_elems = n_bytes / sizeof(float);
+            result.resize(n_elems);
+            fin.seekg(it->second, std::ios::beg);
+            fin.read(reinterpret_cast<char*>(result.data()), n_bytes);
+            return result;
+        };
+
+        auto get_scalar = [&](const std::string & name, float default_val) {
+            auto v = get_vector(name);
+            if (v.empty()) {
                return default_val;
            }
-            size_t offset = it->second;
-            fin.seekg(offset, std::ios::beg);
-            float value;
-            fin.read(reinterpret_cast<char*>(&value), sizeof(float));
-            return value;
+            if (v.size() != 1) {
+                throw std::runtime_error(string_format("%s: expected scalar tensor '%s' but got %d elements\n",
+                            __func__, name.c_str(), (int) v.size()));
+            }
+
+            return v[0];
        };

        model.class_embedding = get_tensor(TN_CLASS_EMBD, false);
@@ -2800,6 +2848,68 @@ struct clip_model_loader {
                        layer.conv_pw2_b   = get_tensor(string_format(TN_CONV_PW2,  prefix, il, "bias"));
                    }
                } break;
+            case PROJECTOR_TYPE_PARAKEET:
+                {
+
+                    hparams.mel_filters = get_vector(TN_MEL_FILTERS);
+                    hparams.window      = get_vector(TN_WINDOW);
+
+                    // Subsampling layers (conv1d)
+                    for (int i : {0, 2, 3, 5, 6}) {
+                        model.pre_encode_conv_X_w[i] = get_tensor(string_format(TN_CONV1D, i, "weight"));
+                        model.pre_encode_conv_X_b[i] = get_tensor(string_format(TN_CONV1D, i, "bias"));
+                    }
+                    model.pre_encode_out_w = get_tensor(string_format(TN_PRE_ENCODE_OUT, "weight"));
+                    model.pre_encode_out_b = get_tensor(string_format(TN_PRE_ENCODE_OUT, "bias"));
+
+                    // Projection layers
+                    model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"), false);
+                    model.mm_0_w        = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"), false);
+                    model.mm_1_w        = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"), false);
+
+                    // Encoder layers
+                    for (int il = 0; il < hparams.n_layer; ++il) {
+                        auto & layer = model.layers[il];
+
+                        // Attention (from shared above)
+
+                        // Relative position encoding
+                        layer.linear_pos_w = get_tensor(string_format(TN_LINEAR_POS, prefix, il, "weight"));
+                        layer.pos_bias_u   = get_tensor(string_format(TN_POS_BIAS_U, prefix, il));
+                        layer.pos_bias_v   = get_tensor(string_format(TN_POS_BIAS_V, prefix, il));
+
+                        // Convolution module
+                        layer.conv_pw1_w     = get_tensor(string_format(TN_CONV_PW1,       prefix, il, "weight"));
+                        layer.conv_pw1_b     = get_tensor(string_format(TN_CONV_PW1,       prefix, il, "bias"), false);
+                        layer.conv_dw_w      = get_tensor(string_format(TN_CONV_DW,        prefix, il, "weight"));
+                        layer.conv_dw_b      = get_tensor(string_format(TN_CONV_DW,        prefix, il, "bias"), false);
+                        layer.conv_norm_w    = get_tensor(string_format(TN_CONV_NORM,      prefix, il, "weight"));
+                        layer.conv_norm_b    = get_tensor(string_format(TN_CONV_NORM,      prefix, il, "bias"));
+                        layer.conv_norm_mean = get_tensor(string_format(TN_CONV_NORM_MEAN, prefix, il));
+                        layer.conv_norm_var  = get_tensor(string_format(TN_CONV_NORM_VAR,  prefix, il));
+                        layer.conv_pw2_w     = get_tensor(string_format(TN_CONV_PW2,       prefix, il, "weight"));
+                        layer.conv_pw2_b     = get_tensor(string_format(TN_CONV_PW2,       prefix, il, "bias"), false);
+
+                        // Feed-forward networks
+                        layer.ff_norm_w   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "weight"));
+                        layer.ff_norm_b   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "bias"));
+
+                        layer.ff_norm_1_w = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "weight"));
+                        layer.ff_norm_1_b = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "bias"));
+                        layer.ff_up_1_w   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "weight"));
+                        layer.ff_up_1_b   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "bias"), false);
+                        layer.ff_down_1_w = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "weight"));
+                        layer.ff_down_1_b = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "bias"), false);
+
+                        // Layer norms
+                        layer.norm_conv_w = get_tensor(string_format(TN_NORM_CONV, prefix, il, "weight"));
+                        layer.norm_conv_b = get_tensor(string_format(TN_NORM_CONV, prefix, il, "bias"));
+                    }
+
+                    model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
+                    model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    model.mm_model_mlp_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
+                } break;
            case PROJECTOR_TYPE_GRANITE_SPEECH:
                {
                    model.inp_proj_w     = get_tensor(string_format(TN_INP_PROJ,    "weight"));
@@ -3645,6 +3755,10 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
                }
                n_patches = n;
            } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                n_patches = (img->nx() + (params.subsampling_factor - 1)) / params.subsampling_factor;
+            } break;
        case PROJECTOR_TYPE_GEMMA4UA:
            {
                n_patches = img->nx();  // no downsampling: one token per raw waveform frame
@@ -4558,6 +4672,88 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
                }
                set_input_f32("pos_emb", pos_emb);
            } break;
+        case PROJECTOR_TYPE_PARAKEET:
+            {
+                GGML_ASSERT(imgs.entries.size() == 1);
+                struct ggml_tensor * attn_mask = ggml_graph_get_tensor(gf, "attn_mask");
+                const int n_q = attn_mask->ne[1];
+                const int n_k = attn_mask->ne[0];
+                const int n_frames = imgs.entries.front().nx();
+                const int n_tokens_real = (n_frames + hparams.subsampling_factor-1) / hparams.subsampling_factor;
+                const float mask_value  = -1e30f;
+
+                std::vector<float> mask_data(n_q * n_k);
+                if (n_k == n_q) {
+                    // full attention: mask keys that are padding
+                    for (int q = 0; q < n_q; ++q) {
+                        for (int k = 0; k < n_k; ++k) {
+                            mask_data[q * n_k + k] = (k >= n_tokens_real) ? mask_value : 0.0f;
+                        }
+                    }
+                } else {
+                    // local attention: mask keys outside the valid window
+                    const int att_left = n_k / 2;
+                    for (int q = 0; q < n_q; ++q) {
+                        for (int k = 0; k < n_k; ++k) {
+                            const int key = q - att_left + k;
+                            mask_data[q * n_k + k] = (key >= 0 && key < n_tokens_real) ? 0.0f : mask_value;
+                        }
+                    }
+                }
+                set_input_f32(attn_mask->name, mask_data);
+
+                // local attention skew mask: zeroes out the probs that were
+                // computed for keys outside the valid sliding window.
+                if (struct ggml_tensor * local_mask = ggml_graph_get_tensor(gf, "local_mask")) {
+                    const int lm_k = local_mask->ne[0];
+                    const int lm_q = local_mask->ne[1];
+                    const int window_size = lm_k - lm_q + 1;
+                    std::vector<float> lm_data(lm_q * lm_k);
+                    for (int q = 0; q < lm_q; ++q) {
+                        for (int k = 0; k < lm_k; ++k) {
+                            const int rel = k - q;
+                            lm_data[q * lm_k + k] = (rel >= 0 && rel < window_size) ? 1.0f : 0.0f;
+                        }
+                    }
+                    set_input_f32(local_mask->name, lm_data);
+                }
+
+                // Generate rotation frequencies for relative positional encoding.
+                {
+                    const int n_state     = hparams.n_embd;
+                    const int d_half      = n_state / 2;
+                    const float log_10000 = logf(10000.0f);
+                    std::vector<float> freqs(d_half);
+                    for (int k = 0; k < d_half; ++k) {
+                        freqs[k] = expf(-(float(k * 2) * log_10000 / float(n_state)));
+                    }
+                    set_input_f32("pos_freqs", freqs);
+                }
+
+                // Generate relative positional distance values which scaled by
+                // the frequency to produce the angles for sin/cos.
+                {
+                    // window_size is only known after graph construction since it depends on
+                    // n_time from the conv output, so we read it back from the graph tensor.
+                    struct ggml_tensor * rel_pos = ggml_graph_get_tensor(gf, "rel_positions");
+                    const int window_size = rel_pos->ne[1];
+                    std::vector<float> pos(window_size);
+                    // local attention: window is fixed at [att_left, att_right]
+                    // full attention: window covers the full sequence, centered
+                    if (ggml_graph_get_tensor(gf, "local_mask")) {
+                        const int att_left = window_size / 2;
+                        for (int t = 0; t < window_size; ++t) {
+                            pos[t] = float(att_left - t);
+                        }
+                    } else {
+                        const int n_time = (window_size + 1) / 2;
+                        for (int t = 0; t < window_size; ++t) {
+                            pos[t] = float(n_time - 1 - t);
+                        }
+                    }
+                    set_input_f32(rel_pos->name, pos);
+                }
+            } break;
        case PROJECTOR_TYPE_GRANITE_SPEECH:
            {
                const int context_size = ctx->model.hparams.audio_chunk_size;
@@ -4841,6 +5037,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
            return ctx->model.mm_ffn_down_w->ne[1];
        case PROJECTOR_TYPE_MIMO_AUDIO:
            return ctx->model.mm_2_w->ne[1];
+        case PROJECTOR_TYPE_PARAKEET:
+            return ctx->model.mm_1_w->ne[1];
        default:
            GGML_ABORT("Unknown projector type");
    }
@@ -222,6 +222,11 @@ struct clip_graph_kimik25 : clip_graph {
    ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
 };

+struct clip_graph_parakeet : clip_graph {
+    clip_graph_parakeet(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_exaone4_5 : clip_graph {
    clip_graph_exaone4_5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
    ggml_cgraph * build() override;
@@ -0,0 +1,421 @@
+#include "models.h"
+
+static constexpr int PARAKEET_LOCAL_ATTN_THRESHOLD = 8192;
+static constexpr int PARAKEET_LOCAL_ATTN_WINDOW    = 128;
+
+// conv subsampling + conformer encoder
+ggml_cgraph * clip_graph_parakeet::build() {
+
+    // Conv subsampling
+    ggml_tensor * inp = build_inp_raw(1);
+    inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+
+    // [freq, time, channels, batch]
+    ggml_tensor * cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[0], inp, 2, 2, 1, 1, 1, 1);
+    cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[0]);
+    cb(cur, "pre_conv_0", -1);
+
+    cur = ggml_relu(ctx0, cur);
+    cb(cur, "pre_conv_0_relu", -1);
+
+    // [freq, time, channels, batch]
+    cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[2], cur, 2, 2, 1, 1, 1, 1);
+    cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[2]);
+    cb(cur, "pre_conv_2", -1);
+
+    // [freq, time, channels, batch]
+    cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[3], cur, 1, 1, 0, 0, 1, 1);
+    cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[3]);
+    cb(cur, "pre_conv_3", -1);
+
+    cur = ggml_relu(ctx0, cur);
+    cb(cur, "pre_conv_3_relu", -1);
+
+    // [freq, time, channels, batch]
+    cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[5], cur, 2, 2, 1, 1, 1, 1);
+    cb(cur, "pre_conv_5_direct", -1);
+    cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[5]);
+    cb(cur, "pre_conv_5", -1);
+
+    // [freq, time, channels, batch]
+    cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[6], cur, 1, 1, 0, 0, 1, 1);
+    cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[6]);
+    cb(cur, "pre_conv_6", -1);
+
+    cur = ggml_relu(ctx0, cur);
+    cb(cur, "pre_conv_6_relu", -1);
+
+    // [freq, time, chan]
+    cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+    // [freq, chan, time]
+    cur = ggml_cont(ctx0, cur);
+
+    const int n_freq   = cur->ne[0];
+    const int n_chan   = cur->ne[1];
+    const int n_frames = cur->ne[2];
+
+    // [freq, time, chan, batch] -> [(freq * chan), time]
+    cur = ggml_reshape_2d(ctx0, cur, n_freq * n_chan, n_frames);
+
+    cur = build_mm(model.pre_encode_out_w, cur);
+    cur = ggml_add(ctx0, cur, model.pre_encode_out_b);
+
+    ggml_set_name(cur, "pre_enc_out");
+
+    // Encoder
+
+    const auto & hparams  = model.hparams;
+    const int n_layer     = hparams.n_layer;
+    const int n_state     = hparams.n_embd;
+    const float fc_factor = 0.5f;
+
+    const int  n_time      = cur->ne[1];
+    const bool local_attn  = n_time > PARAKEET_LOCAL_ATTN_THRESHOLD;
+    const int  att_left    = local_attn ? PARAKEET_LOCAL_ATTN_WINDOW : n_time - 1;
+    const int  att_right   = local_attn ? PARAKEET_LOCAL_ATTN_WINDOW : n_time - 1;
+    const int  window_size = local_attn ? att_left + att_right + 1 : 2 * n_time - 1;
+    const int  d_half      = n_state / 2;
+    const int  mask_dim    = local_attn ? window_size : n_time;
+
+    // mask [key, n_time]
+    struct ggml_tensor * attn_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, mask_dim, n_time);
+    ggml_set_name(attn_mask, "attn_mask");
+    ggml_set_input(attn_mask);
+
+    struct ggml_tensor * local_mask = nullptr;
+    if (local_attn) {
+        const int chunk = att_left + att_right;
+        local_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, chunk + window_size - 1, chunk);
+        ggml_set_name(local_mask, "local_mask");
+        ggml_set_input(local_mask);
+    }
+
+    struct ggml_tensor * pos_freqs = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, d_half);
+    ggml_set_name(pos_freqs, "pos_freqs");
+    ggml_set_input(pos_freqs);
+
+    struct ggml_tensor * rel_positions = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, window_size);
+    ggml_set_name(rel_positions, "rel_positions");
+    ggml_set_input(rel_positions);
+
+    struct ggml_tensor * freqs = ggml_repeat_4d(ctx0, pos_freqs, d_half, window_size, 1, 1);
+    struct ggml_tensor * theta = ggml_mul(ctx0, freqs, rel_positions);
+
+    struct ggml_tensor * sin = ggml_reshape_3d(ctx0, ggml_sin(ctx0, theta), 1, d_half, window_size);
+    struct ggml_tensor * cos = ggml_reshape_3d(ctx0, ggml_cos(ctx0, theta), 1, d_half, window_size);
+    struct ggml_tensor * pos_emb = ggml_reshape_2d(ctx0, ggml_cont(ctx0, ggml_concat(ctx0, sin, cos, 0)), n_state, window_size);
+    ggml_set_name(pos_emb, "pos_emb");
+
+    for (int il = 0; il < n_layer; ++il) {
+        const auto & layer = model.layers[il];
+        // FFN1
+        {
+            struct ggml_tensor * residual = cur;
+            ggml_format_name(cur, "enc_%d_res", il);
+
+            // norm
+            cur = ggml_norm(ctx0, cur, hparams.eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ff_norm_w), layer.ff_norm_b);
+            ggml_format_name(cur, "enc_%d_ffn_norm_1", il);
+
+            cur = build_ffn(cur, layer.ff_up_w, nullptr, nullptr, nullptr, layer.ff_down_w, nullptr, FFN_SILU, il);
+            ggml_format_name(cur, "enc_%d_ffn_1", il);
+
+            cur = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
+            ggml_format_name(cur, "enc_%d_res_ffn", il);
+        }
+
+        // self attention block using relative positional encoding from model.position_embedding.
+        {
+            // [feat, time_frames, 1, 1]
+            struct ggml_tensor * residual = cur;
+
+            cur = ggml_norm(ctx0, cur, hparams.eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_1_w), layer.ln_1_b);
+            ggml_format_name(cur, "enc_%d_attn_norm", il);
+
+            const int n_head = hparams.n_head;
+            const int d_head = n_state / n_head;
+
+            // [feat, time_frames, 1, 1]
+            struct ggml_tensor * Q_cur = build_mm(layer.q_w, cur);
+            struct ggml_tensor * K_cur = build_mm(layer.k_w, cur);
+            struct ggml_tensor * V_cur = build_mm(layer.v_w, cur);
+
+            // [d_head, n_heads, n_time, 1]
+            Q_cur = ggml_reshape_3d(ctx0, Q_cur, d_head, n_head, n_time);
+            K_cur = ggml_reshape_3d(ctx0, K_cur, d_head, n_head, n_time);
+            V_cur = ggml_reshape_3d(ctx0, V_cur, d_head, n_head, n_time);
+
+            // [n_state, window_size]
+            struct ggml_tensor * pos = build_mm(layer.linear_pos_w, pos_emb);
+            // [feat, head, window_size, 1]
+            pos = ggml_reshape_3d(ctx0, pos, d_head, n_head, pos_emb->ne[1]);
+            // [feat, window_size, head, 1]
+            pos = ggml_cont(ctx0, ggml_permute(ctx0, pos, 0, 2, 1, 3));
+            ggml_format_name(pos, "enc_%d_attn_pos", il);
+
+            if (local_attn) {
+                const int  chunk         = att_left + att_right;
+                const int  n_group       = (n_time + chunk - 1) / chunk;
+                const int  n_time_padded = n_group * chunk;
+                const int  n_kv_chunk    = chunk + window_size - 1;
+                const int  n_kv_dense    = n_kv_chunk * n_group;
+                const bool need_padding  = n_time_padded > n_time;
+
+                Q_cur = ggml_cont(ctx0, ggml_permute(ctx0, Q_cur, 0, 2, 1, 3));
+                K_cur = ggml_cont(ctx0, ggml_permute(ctx0, K_cur, 0, 2, 1, 3));
+                V_cur = ggml_cont(ctx0, ggml_permute(ctx0, V_cur, 0, 2, 1, 3));
+
+                // content bias
+                struct ggml_tensor * bias_u = ggml_reshape_3d(ctx0, layer.pos_bias_u, d_head, 1, n_head);
+                struct ggml_tensor * Q_u = ggml_add(ctx0, Q_cur, bias_u);
+
+                // position bias
+                struct ggml_tensor * bias_v = ggml_reshape_3d(ctx0, layer.pos_bias_v, d_head, 1, n_head);
+                struct ggml_tensor * Q_v = ggml_add(ctx0, Q_cur, bias_v);
+
+                // right pad the time dimension
+                struct ggml_tensor * Q_u_padded = need_padding ?
+                    ggml_pad_ext(ctx0, Q_u, 0, 0, 0, n_time_padded - n_time, 0, 0, 0, 0) : Q_u;
+                Q_u_padded = ggml_reshape_4d(ctx0, Q_u_padded, d_head, chunk, n_group, n_head);
+
+                // pad front and back for the first and last time frames
+                struct ggml_tensor * K_padded = ggml_pad_ext(ctx0, K_cur, 0, 0, att_left, att_right, 0, 0, 0, 0);
+                if (n_kv_dense > K_padded->ne[1]) {
+                    K_padded = ggml_pad_ext(ctx0, K_padded, 0, 0, 0, n_kv_dense - K_padded->ne[1], 0, 0, 0, 0);
+                }
+
+                // sliding window view: each group spans n_kv_chunk keys but steps by chunk
+                struct ggml_tensor * K_chunk = ggml_view_4d(ctx0, K_padded,
+                        d_head, n_kv_chunk, n_group, n_head,
+                        K_padded->nb[1],
+                        (size_t) chunk * K_padded->nb[1],
+                        K_padded->nb[2],
+                        0);
+                K_chunk = ggml_cont(ctx0, K_chunk);
+
+                struct ggml_tensor * content_scores = ggml_mul_mat(ctx0, K_chunk, Q_u_padded);
+
+                // trim the dense output down to window_size scores per query
+                content_scores = ggml_view_4d(ctx0, content_scores,
+                        window_size, chunk, n_group, n_head,
+                        (size_t) (chunk + window_size) * content_scores->nb[0],
+                        content_scores->nb[2],
+                        content_scores->nb[3],
+                        0);
+                content_scores = ggml_cont(ctx0, content_scores);
+
+                // ungroup: [window_size, n_time_padded, n_head]
+                content_scores = ggml_reshape_3d(ctx0, content_scores, window_size, n_time_padded, n_head);
+                if (need_padding) {
+                    content_scores = ggml_view_3d(ctx0, content_scores,
+                            window_size, n_time, n_head,
+                            content_scores->nb[1],
+                            content_scores->nb[2],
+                            0);
+                }
+
+                // Q_v: [d_head, time, head]
+                Q_v = ggml_cont(ctx0, ggml_permute(ctx0, Q_v, 0, 2, 1, 3));
+                struct ggml_tensor * rel_pos_scores = ggml_mul_mat(ctx0, pos, Q_v);
+
+                struct ggml_tensor * attn_scores = ggml_add(ctx0, content_scores, rel_pos_scores);
+                attn_scores = ggml_soft_max_ext(ctx0, attn_scores, attn_mask, 1.0f / std::sqrt(d_head), 0.0f);
+                ggml_format_name(attn_scores, "enc_%d_attn_probs", il);
+
+                // expand probs back to n_kv_chunk width for the V matmul
+                struct ggml_tensor * probs_padded = need_padding ?
+                    ggml_pad_ext(ctx0, attn_scores, 0, 0, 0, n_time_padded - n_time, 0, 0, 0, 0) : attn_scores;
+
+                probs_padded = ggml_reshape_4d(ctx0, probs_padded, window_size, chunk, n_group, n_head);
+                probs_padded = ggml_pad_ext(ctx0, probs_padded, 0, chunk, 0, 0, 0, 0, 0, 0);
+                probs_padded = ggml_view_4d(ctx0, probs_padded,
+                        n_kv_chunk, chunk, n_group, n_head,
+                        (size_t) n_kv_chunk * probs_padded->nb[0],
+                        probs_padded->nb[2],
+                        probs_padded->nb[3],
+                        0);
+                probs_padded = ggml_cont(ctx0, probs_padded);
+                probs_padded = ggml_mul(ctx0, probs_padded, local_mask);
+
+                struct ggml_tensor * V_padded = ggml_pad_ext(ctx0, V_cur, 0, 0, att_left, att_right, 0, 0, 0, 0);
+                if (n_kv_dense > V_padded->ne[1]) {
+                    V_padded = ggml_pad_ext(ctx0, V_padded, 0, 0, 0, n_kv_dense - V_padded->ne[1], 0, 0, 0, 0);
+                }
+                V_padded = ggml_cont(ctx0, ggml_transpose(ctx0, V_padded));
+
+                struct ggml_tensor * V_chunk = ggml_view_4d(ctx0, V_padded,
+                        n_kv_chunk, d_head, n_group, n_head,
+                        V_padded->nb[1],
+                        (size_t) chunk * V_padded->nb[0],
+                        V_padded->nb[2],
+                        0);
+                V_chunk = ggml_cont(ctx0, V_chunk);
+
+                cur = ggml_mul_mat(ctx0, V_chunk, probs_padded);
+                cur = ggml_reshape_3d(ctx0, cur, d_head, n_time_padded, n_head);
+                if (need_padding) {
+                    cur = ggml_view_3d(ctx0, cur, d_head, n_time, n_head, cur->nb[1], cur->nb[2], 0);
+                }
+                cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));
+                cur = ggml_reshape_2d(ctx0, cur, n_state, n_time);
+                cur = build_mm(layer.o_w, cur);
+            } else {
+                // full attention
+                struct ggml_tensor * Q_u = ggml_add(ctx0, Q_cur, layer.pos_bias_u);
+                ggml_format_name(Q_u, "enc_%d_attn_q_u", il);
+
+                struct ggml_tensor * K_prep = ggml_permute(ctx0, K_cur, 0, 2, 1, 3);
+                struct ggml_tensor * Q_prep = ggml_permute(ctx0, Q_u,   0, 2, 1, 3);
+                struct ggml_tensor * content_scores = ggml_mul_mat(ctx0, K_prep, Q_prep);
+                ggml_format_name(content_scores, "enc_%d_attn_content_scores", il);
+
+                struct ggml_tensor * Q_v = ggml_add(ctx0, Q_cur, layer.pos_bias_v);
+                ggml_format_name(Q_v, "enc_%d_attn_q_v", il);
+
+                Q_v = ggml_permute(ctx0, Q_v, 0, 2, 1, 3);
+                Q_v = ggml_cont(ctx0, Q_v);
+                ggml_format_name(Q_v, "enc_%d_attn_q_v_perm", il);
+
+                struct ggml_tensor * rel_pos_scores = ggml_mul_mat(ctx0, pos, Q_v);
+                ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos", il);
+
+                // Relative positional shift
+                {
+                    const auto pos_window = rel_pos_scores->ne[0];
+                    const auto n_frame    = rel_pos_scores->ne[1];
+                    const auto n_head     = rel_pos_scores->ne[2];
+
+                    rel_pos_scores = ggml_pad(ctx0, rel_pos_scores, 1, 0, 0, 0);
+                    rel_pos_scores = ggml_roll(ctx0, rel_pos_scores, 1, 0, 0, 0);
+
+                    rel_pos_scores = ggml_reshape_3d(ctx0, rel_pos_scores, n_frame, pos_window + 1, n_head);
+                    rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
+                    ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_reshaped", il);
+
+                    int center = pos_window / 2;
+                    size_t offset = rel_pos_scores->nb[0] * (center+1);
+
+                    rel_pos_scores = ggml_view_3d(ctx0, rel_pos_scores,
+                                                  n_frame, pos_window, n_head,
+                                                  (pos_window) * 4,
+                                                  rel_pos_scores->nb[2],
+                                                  offset);
+                    rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
+                    ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_shifted", il);
+
+                    rel_pos_scores = ggml_view_3d(ctx0, rel_pos_scores,
+                                                  content_scores->ne[0],
+                                                  content_scores->ne[1],
+                                                  rel_pos_scores->ne[2],
+                                                  rel_pos_scores->nb[1],
+                                                  rel_pos_scores->nb[2],
+                                                  0);
+                    rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
+                    ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_shifted_view", il);
+                }
+
+                struct ggml_tensor * attn_scores = ggml_add(ctx0, content_scores, rel_pos_scores);
+                ggml_format_name(attn_scores, "enc_%d_attn_scores", il);
+                attn_scores = ggml_scale(ctx0, attn_scores, 1.0f / std::sqrt(d_head));
+                attn_scores = ggml_add(ctx0, attn_scores, attn_mask);
+                ggml_format_name(attn_scores, "enc_%d_attn_scores_scaled", il);
+
+                struct ggml_tensor * probs = ggml_soft_max(ctx0, attn_scores);
+                ggml_format_name(probs, "enc_%d_attn_probs", il);
+
+                V_cur = ggml_cont(ctx0, ggml_permute(ctx0, V_cur, 1, 2, 0, 3));
+                ggml_format_name(V_cur, "enc_%d_attn_v_cur", il);
+                cur = ggml_mul_mat(ctx0, probs, V_cur);
+                ggml_format_name(cur, "enc_%d_attn_inp", il);
+
+                cur = ggml_permute(ctx0, cur, 2, 0, 1, 3);
+                cur = ggml_cont_2d(ctx0, cur, n_state, n_time);
+                cur = build_mm(layer.o_w, cur);
+            }
+            ggml_format_name(cur, "enc_%d_attn_out", il);
+
+            cur = ggml_add(ctx0, residual, cur);
+            ggml_format_name(cur, "enc_%d_attn_res", il);
+        }
+
+        // Convolution
+        {
+            struct ggml_tensor * residual = cur;
+            ggml_format_name(cur, "enc_%d_residual_conv", il);
+
+            cur = ggml_norm(ctx0, cur, hparams.eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.norm_conv_w), layer.norm_conv_b);
+            ggml_format_name(cur, "enc_%d_norm_conv", il);
+
+            // pointwise 1d convolution:
+            cur = build_mm(layer.conv_pw1_w, cur);
+            ggml_format_name(cur, "enc_%d_conv_pw1", il);
+
+            {
+                int64_t d = cur->ne[0] / 2;
+                struct ggml_tensor * signal = ggml_view_2d(ctx0, cur, d, cur->ne[1], cur->nb[1], 0);
+                struct ggml_tensor * gate   = ggml_view_2d(ctx0, cur, d, cur->ne[1], cur->nb[1], d * cur->nb[0]);
+
+                cur = ggml_mul(ctx0, signal, ggml_sigmoid(ctx0, gate));
+                ggml_format_name(cur, "enc_%d_conv_glu", il);
+            }
+
+            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+            // use ggml_ssm_conv for f32 precision
+            const int dw_pad = (hparams.audio_conv_kernel_size - 1) / 2;
+            cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
+            cur = ggml_roll(ctx0, cur, dw_pad, 0, 0, 0);
+            cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
+            ggml_format_name(cur, "enc_%d_conv_dw_pad", il);
+
+            cur = ggml_ssm_conv(ctx0, cur, layer.conv_dw_w);
+            ggml_format_name(cur, "enc_%d_conv_1d_dw", il);
+
+            cur = ggml_sub(ctx0, cur, layer.conv_norm_mean);
+            struct ggml_tensor * std = ggml_sqrt(ctx0, layer.conv_norm_var);
+            cur = ggml_div(ctx0, cur, std);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.conv_norm_w), layer.conv_norm_b);
+            ggml_format_name(cur, "enc_%d_conv_bn", il);
+
+            cur = ggml_silu(ctx0, cur);
+            ggml_format_name(cur, "enc_%d_conv_silu", il);
+
+            cur = build_mm(layer.conv_pw2_w, cur);
+            ggml_format_name(cur, "enc_%d_conv_pw2", il);
+
+            cur = ggml_add(ctx0, residual, cur);
+            ggml_format_name(cur, "enc_%d_conv_res", il);
+        }
+
+        // FFN2
+        {
+            struct ggml_tensor * residual = cur;
+            cur = ggml_norm(ctx0, cur, hparams.eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ff_norm_1_w), layer.ff_norm_1_b);
+            ggml_format_name(cur, "enc_%d_ffn_norm_2", il);
+
+            cur = build_ffn(cur, layer.ff_up_1_w, nullptr, nullptr, nullptr, layer.ff_down_1_w, nullptr, FFN_SILU, il);
+            cur = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, 0.5));
+            ggml_format_name(cur, "enc_%d_ffn_res", il);
+        }
+
+        cur = ggml_norm(ctx0, cur, hparams.eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_2_w), layer.ln_2_b);
+    }
+
+    cb(cur, "encoder_out", -1);
+
+    cur = ggml_rms_norm(ctx0, cur, 1e-6);
+    cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
+    cb(cur, "sound_projection.norm", -1);
+
+    cur = build_ffn(cur, model.mm_0_w, model.mm_0_b, nullptr, nullptr, model.mm_1_w, model.mm_1_b, FFN_RELU_SQR, -1);
+    cb(cur, "projected", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}
@@ -1022,6 +1022,209 @@ bool mtmd_audio_preprocessor_gemma4a::preprocess(const float *                 s
 }

 //
+// mtmd_audio_preprocessor_parakeet implementation
+//
+
+void mtmd_audio_preprocessor_parakeet::worker_thread(
+                             int   ith,
+                     const float * window_func,
+                             int   window_size,
+        const std::vector<float> & samples,
+                             int   n_samples,
+                             int   frame_size,
+                             int   frame_step,
+                             int   n_threads,
+                             int   n_fft_bins,
+          const mtmd_audio_cache & cache,
+                  mtmd_audio_mel & mel) {
+    std::vector<float> fft_in(frame_size * 2, 0.0);
+    std::vector<float> fft_out(frame_size * 2 * 2 * 2);
+
+    int n_fb = n_fft_bins;
+    int i = ith;
+
+    GGML_ASSERT(n_fb == 1 + (frame_size / 2));
+
+    const double eps = 5.960464477539063e-08;
+
+    for (; i < std::min(n_samples / frame_step + 1, (int) mel.n_len); i += n_threads) {
+        const int offset = i * frame_step;
+        const int window_pad_left = (frame_size - window_size) / 2;
+
+        // Zero-pad left.
+        std::fill(fft_in.begin(), fft_in.begin() + window_pad_left, 0.0f);
+
+        // Apply windowed samples in the center.
+        const int n_to_process = std::min({window_size, n_samples - offset});
+        for (int j = 0; j < n_to_process; j++) {
+            fft_in[window_pad_left + j] = window_func[j] * samples[offset + window_pad_left + j];
+        }
+
+        // Zero-pad right.
+        std::fill(fft_in.begin() + window_pad_left + n_to_process, fft_in.begin() + frame_size, 0.0f);
+
+        // FFT.
+        fft(cache, fft_in.data(), frame_size, fft_out.data());
+
+        // Calculate modulus^2 of complex numbers.
+        for (int j = 0; j < n_fb; j++) {
+            fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
+        }
+
+        // mel spectrogram.
+        for (int j = 0; j < mel.n_mel; j++) {
+            double sum = 0.0;
+            int k = 0;
+            for (k = 0; k < n_fb - 3; k += 4) {
+                sum +=
+                    fft_out[k + 0] * cache.filters.data[j * n_fb + k + 0] +
+                    fft_out[k + 1] * cache.filters.data[j * n_fb + k + 1] +
+                    fft_out[k + 2] * cache.filters.data[j * n_fb + k + 2] +
+                    fft_out[k + 3] * cache.filters.data[j * n_fb + k + 3];
+            }
+            for (; k < n_fb; k++) {
+                sum += fft_out[k] * cache.filters.data[j * n_fb + k];
+            }
+            mel.data[j * mel.n_len + i] = std::log(sum + eps);
+        }
+    }
+
+    // Otherwise fft_out are all zero.
+    const double empty_sum = std::log(eps);
+    for (; i < mel.n_len; i += n_threads) {
+        for (int j = 0; j < mel.n_mel; j++) {
+            mel.data[j * mel.n_len + i] = empty_sum;
+        }
+    }
+}
+
+void mtmd_audio_preprocessor_parakeet::initialize() {
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+
+    const size_t n_fft = hparams.audio_n_fft / 2 + 1;
+    GGML_ASSERT(hparams.mel_filters.size() == (size_t)hparams.n_mel_bins * n_fft);
+    cache.filters.n_mel = hparams.n_mel_bins;
+    cache.filters.n_fft = n_fft;
+    cache.filters.data  = hparams.mel_filters;
+
+    GGML_ASSERT(hparams.window.size() == (size_t)hparams.audio_window_len);
+    GGML_ASSERT(hparams.window.size() <= (size_t) hparams.audio_n_fft);
+    cache.hann_window = hparams.window;
+}
+
+bool mtmd_audio_preprocessor_parakeet::preprocess(const float * samples,
+                                                       size_t   n_samples_in,
+                                  std::vector<mtmd_audio_mel> & output) {
+    if (n_samples_in == 0) {
+        return false;
+    }
+
+    filter_params params;
+    params.n_mel            = hparams.n_mel_bins;
+    params.n_fft_bins       = 1 + (hparams.audio_n_fft / 2);
+    params.hann_window_size = hparams.audio_window_len;
+    params.hop_length       = hparams.audio_hop_len;
+    params.sample_rate      = hparams.audio_sample_rate;
+
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
+
+    const float * window_func = cache.hann_window.data();
+    const int     window_size = params.hann_window_size;
+    const int     frame_size  = (params.n_fft_bins - 1) * 2;
+    const int     frame_step  = params.hop_length;
+
+    // Apply preemphasis filter (high-pass): x[i] = x[i] - 0.97 * x[i-1]
+    std::vector<float> samples_preprocessed(samples, samples + n_samples_in);
+    {
+        const float preemph = 0.97f;
+        for (int i = n_samples_in - 1; i > 0; i--) {
+            samples_preprocessed[i] = samples_preprocessed[i] - preemph * samples_preprocessed[i - 1];
+        }
+    }
+
+    // Parakeet uses centered constant padding
+    const size_t pad = (size_t)(frame_size / 2);
+    std::vector<float> samples_padded(n_samples_in + 2 * pad, 0.0f);
+    std::copy(samples_preprocessed.begin(), samples_preprocessed.end(), samples_padded.begin() + pad);
+
+    mtmd_audio_mel out_full;
+    out_full.n_mel = params.n_mel;
+    out_full.n_len = (samples_padded.size() - frame_size) / frame_step + 1;
+    out_full.n_len_org = out_full.n_len;
+    out_full.data.resize(out_full.n_mel * out_full.n_len);
+
+    const int n_threads = 4;
+    std::vector<std::thread> workers(n_threads - 1);
+    for (int iw = 0; iw < n_threads - 1; ++iw) {
+        workers[iw] = std::thread(
+            worker_thread, iw + 1,
+            window_func,
+            window_size,
+            std::cref(samples_padded),
+            samples_padded.size(),
+            frame_size,
+            frame_step,
+            n_threads,
+            params.n_fft_bins,
+            std::cref(cache),
+            std::ref(out_full)
+        );
+    }
+
+    worker_thread(0,
+            window_func,
+            window_size,
+            samples_padded,
+            samples_padded.size(),
+            frame_size,
+            frame_step,
+            n_threads,
+            params.n_fft_bins,
+            cache,
+            out_full);
+
+    for (int iw = 0; iw < n_threads - 1; ++iw) {
+        workers[iw].join();
+    }
+
+    // Per-feature normalization (only on valid frames)
+    {
+        const double eps = 1e-5;
+        int valid_frames = n_samples_in / frame_step;
+
+        for (int j = 0; j < out_full.n_mel; j++) {
+            double sum = 0.0;
+            double sq_diff_sum = 0.0;
+
+            // Calculate Mean ONLY on valid audio frames
+            for (int i = 0; i < valid_frames; i++) {
+                sum += (double)out_full.data[j * out_full.n_len + i];
+            }
+            double mean = sum / valid_frames;
+
+            // Calculate Variance ONLY on valid audio frames
+            for (int i = 0; i < valid_frames; i++) {
+                double diff = (double)out_full.data[j * out_full.n_len + i] - mean;
+                sq_diff_sum += diff * diff;
+            }
+
+            double std_dev = std::sqrt(sq_diff_sum / (valid_frames - 1.0));
+            double denominator = std_dev + eps;
+
+            // Apply to ALL frames (including the padded ones)
+            for (int i = 0; i < out_full.n_len; i++) {
+                out_full.data[j * out_full.n_len + i] = (float)((out_full.data[j * out_full.n_len + i] - mean) / denominator);
+            }
+        }
+    }
+
+    output.push_back(std::move(out_full));
+    return true;
+}
+
+
 // mtmd_audio_preprocessor_gemma4ua
 //

@@ -120,6 +120,21 @@ struct mtmd_audio_preprocessor_mimo_audio : mtmd_audio_preprocessor {
    mtmd_audio_cache cache;
 };

+struct mtmd_audio_preprocessor_parakeet : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_parakeet(clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) { }
+    void initialize() override;
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+
+    static void worker_thread(int ith, const float * window_func, int window_size,
+                              const std::vector<float> & samples, int n_samples,
+                              int frame_size, int frame_step, int n_threads,
+                              int n_fft_bins,
+                              const mtmd_audio_cache & cache, mtmd_audio_mel & mel);
+};
+
 //
 // streaming ISTFT - converts spectrogram frames back to audio one frame at a time
 //
@@ -724,6 +724,10 @@ struct mtmd_context {
                    aud_end = "<audio|>";
                    audio_preproc = std::make_unique<mtmd_audio_preprocessor_gemma4a>(ctx_a);
                } break;
+            case PROJECTOR_TYPE_PARAKEET:
+                {
+                    audio_preproc = std::make_unique<mtmd_audio_preprocessor_parakeet>(ctx_a);
+                } break;
            case PROJECTOR_TYPE_GEMMA4UA:
                {
                    aud_beg = "<|audio>";
@@ -11,8 +11,7 @@
 		classifyToolResult,
 		formatJsonPretty,
 		parseToolResultWithImages,
-		type AgenticSection,
-		type ToolResultLine
+		type AgenticSection
 	} from '$lib/utils';
 	import { getBuiltinToolUi } from '$lib/constants/built-in-tools';
 	import type { DatabaseMessageExtra } from '$lib/types';
@@ -29,11 +28,10 @@
 	let { section, open, isStreaming, attachments, onToggle }: Props = $props();

 	const title = $derived(getBuiltinToolUi(section.toolName)?.label ?? section.toolName ?? '');
-
-	const parsedLines: ToolResultLine[] = $derived(
+	const outputKind = $derived(classifyToolResult(section.toolResult));
+	const parsedLines = $derived(
 		section.toolResult ? parseToolResultWithImages(section.toolResult, attachments) : []
 	);
-	const outputKind = $derived(classifyToolResult(section.toolResult));
 </script>

 <ToolCallBlock {section} {open} {isStreaming} meta={null} {title} {onToggle}>
@@ -15,7 +15,6 @@
 	let { section, open, isStreaming, onToggle }: Props = $props();

 	const editFileMeta = $derived(parseEditFileMeta(section));
-
 	const editDiffs = $derived(
 		(editFileMeta?.edits ?? []).map((edit) => computeLineDiff(edit.oldText, edit.newText))
 	);
@@ -27,7 +27,7 @@
 	const isStreamingCall = $derived(section.type === AgenticSectionType.TOOL_CALL_STREAMING);
 	const showSpinner = $derived(isPending || (isStreamingCall && isStreaming));

-	const results: SearchResult[] = $derived(extractSearchResults(section.toolResult));
+	const results = $derived(extractSearchResults(section.toolResult));
 	const query = $derived(extractSearchQuery(section.toolArgs));

 	// Same icon-resolution chain as ChatMessageToolCallBlockDefault so
@@ -28,6 +28,18 @@ export const LATEX_MATH_AND_CODE_PATTERN =
 /** Regex to capture the content of a $$...\\\\...$$ block (display-formula with line-break) */
 export const LATEX_LINEBREAK_REGEXP = /\$\$([\s\S]*?\\\\[\s\S]*?)\$\$/;

+/**
+ * Matches the unescaped `\[...\]` display-math delimiter and surrounding
+ * context so callers can insert line-breaks around the placeholder or convert
+ * to inline when the formula has a non-empty trailing context (e.g. a table
+ * cell that opens with `\[` and closes with content after `\]`).
+ *
+ * group 1: prefix before `\[`
+ * group 2: formula body
+ * group 3: trailing context after `\]`
+ */
+export const LATEX_DISPLAY_BLOCK_REGEXP = /([\S].*?)\\\[([\s\S]*?)\\\](.*)/g;
+
 /**
 * Cheap gate for `preprocessLaTeX`. Every transformation it performs is triggered
 * by a `$` (inline/display math, currency escaping) or a backslash escape
@@ -36,6 +48,76 @@ export const LATEX_LINEBREAK_REGEXP = /\$\$([\s\S]*?\\\\[\s\S]*?)\$\$/;
 */
 export const LATEX_TRIGGER_REGEXP = /[$\\]/;

+/** Inline LaTeX math delimiter (the dollar sign). */
+export const LATEX_INLINE_DELIMITER = '$';
+
+/** Display LaTeX math delimiter (paired dollar signs). */
+export const LATEX_DISPLAY_DELIMITER = '$$';
+
+/** Matches a single non-whitespace character. */
+export const LATEX_NON_WHITESPACE_REGEXP = /\S/;
+
+/** Matches a character that may appear adjacent to `$`, indicating a non-TeX
+ * context such as an identifier (`var$`, `$var`), currency ($5), or code. */
+export const LATEX_NEIGHBOR_CHAR_REGEXP = /[A-Za-z0-9_$-]/;
+
+/** Matches a single digit (used to detect currency-like `$5`). */
+export const LATEX_DIGIT_REGEXP = /[0-9]/;
+
+/** Matches the leading blockquote prefix (`> ` or `>`) on a markdown line. */
+export const LATEX_BLOCKQUOTE_PREFIX_REGEXP = /^(>\s*)/;
+
+/** Matches the placeholder inserted by the protect/restore pipeline for a
+ * protected LaTeX expression. Group 1 is the index into `latexExpressions`. */
+export const LATEX_PLACEHOLDER_REGEXP = /<<LATEX_(\d+)>>/g;
+
+/** Matches the placeholder inserted by the protect/restore pipeline for a
+ * protected code block. Group 1 is the index into `codeBlocks`. */
+export const CODE_BLOCK_PLACEHOLDER_REGEXP = /<<CODE_BLOCK_(\d+)>>/g;
+
+/** Matches a `$` immediately followed by a digit, which is treated as a
+ * currency amount (e.g. `$5`) and escaped to `\$5` so it isn't parsed as math. */
+export const LATEX_CURRENCY_DOLLAR_REGEXP = /\$(?=\d)/g;
+
+/** Captures remaining `$$...$$`, `\[...\]`, `\(...\)` (only unescaped via
+ * `(?<!\\)`) after the display-block pass has run. Group 1 holds the
+ * matched formula. */
+export const LATEX_PROTECT_REGEXP =
+	/(\$\$[\s\S]*?\$\$|(?<!\\)\\\[[\s\S]*?\\\]|(?<!\\)\\\(.*?\\\))/g;
+
+/** Matches unescaped inline `\(...\)` (at least one char inside) used to
+ * convert `\(` → `$` after the protect pass. */
+export const LATEX_INLINE_CONVERT_REGEXP = /(?<!\\)\\\((.+?)\\\)/g;
+
+/** Matches unescaped display `\[...\]` used to convert `\[` → `$$`
+ * after the protect pass. */
+export const LATEX_DISPLAY_CONVERT_REGEXP = /(?<!\\)\\\[([\s\S]*?)\\\]/g;
+
+/** `\(` — opens an inline LaTeX math block. */
+export const LATEX_INLINE_OPEN = '\\(';
+
+/** `\)` — closes an inline LaTeX math block. */
+export const LATEX_INLINE_CLOSE = '\\)';
+
+/** `\[` — opens a display LaTeX math block. */
+export const LATEX_DISPLAY_OPEN = '\\[';
+
+/** `\]` — closes a display LaTeX math block. */
+export const LATEX_DISPLAY_CLOSE = '\\]';
+
+/** `\` — the LaTeX escape character. */
+export const LATEX_BACKSLASH = '\\';
+
+/** `\$` — dollar sign escaped so it isn't parsed as math (used to disambiguate
+ * currency amounts like `$5`). */
+export const LATEX_CURRENCY_ESCAPE = '\\$';
+
+/** `\ce{` — mhchem chemistry command prefix. */
+export const LATEX_MHCHEM_CE = '\\ce{';
+
+/** `\pu{` — mhchem physics-unit command prefix. */
+export const LATEX_MHCHEM_PU = '\\pu{';
+
 /** map from mchem-regexp to replacement */
 export const MHCHEM_PATTERN_MAP: readonly [RegExp, string][] = [
 	[/(\s)\$\\ce{/g, '$1$\\\\ce{'],
@@ -92,8 +92,17 @@ function deriveSingleTurnSections(

 	// 3. Persisted tool calls (from message.toolCalls field)
 	const toolCalls = parseToolCalls(message.toolCalls);
+
+	// Index tool messages by toolCallId for O(1) lookup instead of O(n) find()
+	const toolMsgById = new Map<string, DatabaseMessage>();
+	for (const tm of toolMessages) {
+		if (tm.toolCallId && !toolMsgById.has(tm.toolCallId)) {
+			toolMsgById.set(tm.toolCallId, tm);
+		}
+	}
+
 	for (const tc of toolCalls) {
-		const resultMsg = toolMessages.find((m) => m.toolCallId === tc.id);
+		const resultMsg = tc.id ? toolMsgById.get(tc.id) : undefined;
 		// Only show as pending/loading if we're actively streaming; otherwise it's just a tool call without result
 		const type = resultMsg
 			? AgenticSectionType.TOOL_CALL
@@ -112,9 +121,10 @@ function deriveSingleTurnSections(
 	}

 	// 4. Streaming tool calls (not yet persisted - currently being received)
+	const persistedIds = new Set(toolCalls.map((t) => t.id).filter(Boolean));
 	for (const tc of streamingToolCalls) {
 		// Skip if already in persisted tool calls
-		if (tc.id && toolCalls.find((t) => t.id === tc.id)) continue;
+		if (tc.id && persistedIds.has(tc.id)) continue;
 		sections.push({
 			type: AgenticSectionType.TOOL_CALL_STREAMING,
 			content: '',
@@ -281,15 +291,31 @@ export function splitSearchSummaryList(
 	return { lines };
 }

+/** Bounded cache for parseToolResultWithImages results. */
+const TOOL_RESULT_LINES_CACHE_MAX_SIZE = 32;
+const toolResultLinesCache = new Map<string, ToolResultLine[]>();
+
 /**
 * Parse tool result text into lines, matching image attachments by name.
+ * Memoized: called per render during streaming on unchanged tool result
+ * strings with unchanged extras.
 */
 export function parseToolResultWithImages(
 	toolResult: string,
 	extras?: DatabaseMessageExtra[]
 ): ToolResultLine[] {
+	// Cache key includes image attachment names so we recompute when
+	// attachments change, even if the count stays the same.
+	const imageNames = (extras ?? [])
+		.filter((e): e is DatabaseMessageExtraImageFile => e.type === AttachmentType.IMAGE)
+		.map((e) => e.name)
+		.join(NEWLINE);
+	const cacheKey = `${imageNames}:${toolResult}`;
+	const cached = toolResultLinesCache.get(cacheKey);
+	if (cached !== undefined) return cached;
+
 	const lines = toolResult.split(NEWLINE);
-	return lines.map((line) => {
+	const result = lines.map((line) => {
 		const match = line.match(ATTACHMENT_SAVED_REGEX);
 		if (!match || !extras) return { text: line };

@@ -301,8 +327,19 @@ export function parseToolResultWithImages(

 		return { text: line, image };
 	});
+
+	if (toolResultLinesCache.size >= TOOL_RESULT_LINES_CACHE_MAX_SIZE) {
+		toolResultLinesCache.delete(toolResultLinesCache.keys().next().value!);
+	}
+	toolResultLinesCache.set(cacheKey, result);
+
+	return result;
 }

+/** Bounded cache for classifyToolResult results. */
+const CLASSIFY_CACHE_MAX_SIZE = 32;
+const classifyCache = new Map<string, ToolResultKind>();
+
 /**
 * Pick a renderer tier for a tool's result content.
 *
@@ -312,25 +349,39 @@ export function parseToolResultWithImages(
 *              through MarkdownContent for proper formatting.
 *   text     - everything else, rendered as plain text lines (with image
 *              attachment resolution as a side effect).
+ * Memoized: called per render during streaming on unchanged content.
 */
 export function classifyToolResult(content: string | undefined): ToolResultKind {
 	if (!content) return ToolResultKind.TEXT;
+
+	const cached = classifyCache.get(content);
+	if (cached !== undefined) return cached;
+
 	const trimmed = content.trim();
 	if (!trimmed) return ToolResultKind.TEXT;

+	let result: ToolResultKind = ToolResultKind.TEXT;
+
 	// Strongest signal: JSON object/array round-trips through JSON.parse.
 	if (TOOL_RESULT_JSON_OPEN_REGEX.test(trimmed)) {
 		try {
 			JSON.parse(trimmed);
-			return ToolResultKind.JSON;
+			result = ToolResultKind.JSON;
 		} catch (error) {
 			console.error('[agentic] tool result looked like JSON but failed to parse:', error);
 		}
 	}

-	if (looksLikeMarkdown(trimmed)) return ToolResultKind.MARKDOWN;
+	if (result === ToolResultKind.TEXT && looksLikeMarkdown(trimmed)) {
+		result = ToolResultKind.MARKDOWN;
+	}

-	return ToolResultKind.TEXT;
+	if (classifyCache.size >= CLASSIFY_CACHE_MAX_SIZE) {
+		classifyCache.delete(classifyCache.keys().next().value!);
+	}
+	classifyCache.set(content, result);
+
+	return result;
 }

 /**
@@ -370,19 +421,35 @@ function looksLikeMarkdown(content: string): boolean {
 	return false;
 }

+/** Bounded cache for parsed tool-call JSON blobs. */
+const TOOL_CALLS_CACHE_MAX_SIZE = 64;
+const toolCallsParseCache = new Map<string, ApiChatCompletionToolCall[]>();
+
 /**
 * Safely parse the toolCalls JSON string from a DatabaseMessage.
+ * Memoized: the same JSON string is re-parsed on every render during
+ * streaming, which is wasted CPU since tool calls don't change mid-stream.
 */
 function parseToolCalls(toolCallsJson?: string): ApiChatCompletionToolCall[] {
 	if (!toolCallsJson) return [];

+	const cached = toolCallsParseCache.get(toolCallsJson);
+	if (cached) return cached;
+
+	let result: ApiChatCompletionToolCall[];
 	try {
 		const parsed = JSON.parse(toolCallsJson);
-
-		return Array.isArray(parsed) ? parsed : [];
+		result = Array.isArray(parsed) ? parsed : [];
 	} catch {
-		return [];
+		result = [];
 	}
+
+	if (toolCallsParseCache.size >= TOOL_CALLS_CACHE_MAX_SIZE) {
+		toolCallsParseCache.delete(toolCallsParseCache.keys().next().value!);
+	}
+	toolCallsParseCache.set(toolCallsJson, result);
+
+	return result;
 }

 /**
@@ -34,6 +34,10 @@ function escapeCode(code: string): string {
 	return code.replace(AMPERSAND_REGEX, '&amp;').replace(LT_REGEX, '&lt;').replace(GT_REGEX, '&gt;');
 }

+/** Bounded cache for highlightCode results. */
+const HIGHLIGHT_CACHE_MAX_SIZE = 64;
+const highlightCache = new Map<string, string>();
+
 /**
 * Highlights code using highlight.js
 * @param code - The code to highlight
@@ -47,23 +51,37 @@ function escapeCode(code: string): string {
 export function highlightCode(code: string, language: string, autoDetect = true): string {
 	if (!code) return '';

+	// Cache key includes language and autoDetect flag since results differ.
+	// During streaming, the same code string may be highlighted repeatedly
+	// (e.g., when text after a code block changes but the code itself doesn't).
+	const cacheKey = `${language}:${autoDetect}:${code}`;
+	const cached = highlightCache.get(cacheKey);
+	if (cached) return cached;
+
 	const trimmed = trimCodePadding(code);
+	let result: string;

 	try {
 		const lang = language.toLowerCase();
 		const isSupported = hljs.getLanguage(lang);

 		if (isSupported) {
-			return hljs.highlight(trimmed, { language: lang }).value;
+			result = hljs.highlight(trimmed, { language: lang }).value;
 		} else if (autoDetect) {
-			return hljs.highlightAuto(trimmed).value;
+			result = hljs.highlightAuto(trimmed).value;
 		} else {
-			return escapeCode(trimmed);
+			result = escapeCode(trimmed);
 		}
 	} catch {
-		// Fallback to escaped plain text
-		return escapeCode(trimmed);
+		result = escapeCode(trimmed);
 	}
+
+	if (highlightCache.size >= HIGHLIGHT_CACHE_MAX_SIZE) {
+		highlightCache.delete(highlightCache.keys().next().value!);
+	}
+	highlightCache.set(cacheKey, result);
+
+	return result;
 }

 export { trimCodePadding };
@@ -1,9 +1,31 @@
 import {
+	CODE_BLOCK_PLACEHOLDER_REGEXP,
 	CODE_BLOCK_REGEXP,
+	LATEX_BACKSLASH,
+	LATEX_BLOCKQUOTE_PREFIX_REGEXP,
+	LATEX_CURRENCY_DOLLAR_REGEXP,
+	LATEX_CURRENCY_ESCAPE,
+	LATEX_DIGIT_REGEXP,
+	LATEX_DISPLAY_BLOCK_REGEXP,
+	LATEX_DISPLAY_CLOSE,
+	LATEX_DISPLAY_CONVERT_REGEXP,
+	LATEX_DISPLAY_DELIMITER,
+	LATEX_DISPLAY_OPEN,
+	LATEX_INLINE_CLOSE,
+	LATEX_INLINE_CONVERT_REGEXP,
+	LATEX_INLINE_DELIMITER,
+	LATEX_INLINE_OPEN,
 	LATEX_MATH_AND_CODE_PATTERN,
+	LATEX_MHCHEM_CE,
+	LATEX_MHCHEM_PU,
 	LATEX_LINEBREAK_REGEXP,
+	LATEX_NEIGHBOR_CHAR_REGEXP,
+	LATEX_NON_WHITESPACE_REGEXP,
+	LATEX_PLACEHOLDER_REGEXP,
+	LATEX_PROTECT_REGEXP,
 	LATEX_TRIGGER_REGEXP,
-	MHCHEM_PATTERN_MAP
+	MHCHEM_PATTERN_MAP,
+	NEWLINE
 } from '$lib/constants';

 /**
@@ -20,13 +42,13 @@ import {
 * @returns The processed string with LaTeX replaced by placeholders.
 */
 export function maskInlineLaTeX(content: string, latexExpressions: string[]): string {
-	if (!content.includes('$')) {
+	if (!content.includes(LATEX_INLINE_DELIMITER)) {
 		return content;
 	}
 	return content
-		.split('\n')
+		.split(NEWLINE)
 		.map((line) => {
-			if (line.indexOf('$') == -1) {
+			if (line.indexOf(LATEX_INLINE_DELIMITER) == -1) {
 				return line;
 			}

@@ -34,7 +56,7 @@ export function maskInlineLaTeX(content: string, latexExpressions: string[]): st
 			let currentPosition = 0;

 			while (currentPosition < line.length) {
-				const openDollarIndex = line.indexOf('$', currentPosition);
+				const openDollarIndex = line.indexOf(LATEX_INLINE_DELIMITER, currentPosition);

 				if (openDollarIndex == -1) {
 					processedLine += line.slice(currentPosition);
@@ -42,7 +64,7 @@ export function maskInlineLaTeX(content: string, latexExpressions: string[]): st
 				}

 				// Is there a next $-sign?
-				const closeDollarIndex = line.indexOf('$', openDollarIndex + 1);
+				const closeDollarIndex = line.indexOf(LATEX_INLINE_DELIMITER, openDollarIndex + 1);

 				if (closeDollarIndex == -1) {
 					processedLine += line.slice(currentPosition);
@@ -62,14 +84,14 @@ export function maskInlineLaTeX(content: string, latexExpressions: string[]): st
 					shouldSkipAsNonLatex = true;
 				}

-				if (/[A-Za-z0-9_$-]/.test(charBeforeOpen)) {
+				if (LATEX_NEIGHBOR_CHAR_REGEXP.test(charBeforeOpen)) {
 					// Character, digit, $, _ or - before first '$', no TeX.
 					shouldSkipAsNonLatex = true;
 				}

 				if (
-					/[0-9]/.test(charAfterOpen) &&
-					(/[A-Za-z0-9_$-]/.test(charAfterClose) || ' ' == charBeforeClose)
+					LATEX_DIGIT_REGEXP.test(charAfterOpen) &&
+					(LATEX_NEIGHBOR_CHAR_REGEXP.test(charAfterClose) || ' ' == charBeforeClose)
 				) {
 					// First $ seems to belong to an amount.
 					shouldSkipAsNonLatex = true;
@@ -92,7 +114,7 @@ export function maskInlineLaTeX(content: string, latexExpressions: string[]): st

 			return processedLine;
 		})
-		.join('\n');
+		.join(NEWLINE);
 }

 function escapeBrackets(text: string): string {
@@ -107,9 +129,9 @@ function escapeBrackets(text: string): string {
 			if (codeBlock != null) {
 				return codeBlock;
 			} else if (squareBracket != null) {
-				return `$$${squareBracket}$$`;
+				return `${LATEX_DISPLAY_DELIMITER}${squareBracket}${LATEX_DISPLAY_DELIMITER}`;
 			} else if (roundBracket != null) {
-				return `$${roundBracket}$`;
+				return `${LATEX_INLINE_DELIMITER}${roundBracket}${LATEX_INLINE_DELIMITER}`;
 			}

 			return match;
@@ -145,32 +167,49 @@ const doEscapeMhchem = false;
 * preprocessLaTeX("Price: $10. The equation is \\(x^2\\).")
 * // → "Price: $10. The equation is $x^2$."
 */
+/** Bounded cache for preprocessLaTeX results. */
+const LATEX_CACHE_MAX_SIZE = 64;
+const latexCache = new Map<string, string>();
+
 export function preprocessLaTeX(content: string): string {
 	// See also:
 	// https://github.com/danny-avila/LibreChat/blob/main/client/src/utils/latex.ts

+	// Memoize on the input string. During streaming the prefix before an
+	// incomplete code block stays the same across multiple tokens, so the
+	// full protect/restore pipeline would re-run unnecessarily.
+	const cached = latexCache.get(content);
+	if (cached !== undefined) return cached;
+
+	// Save original before the function mutates `content` through steps 0-8
+	const originalContent = content;
+
 	// Every step below keys off a `$` or a backslash escape (\[ \] \( \) \ce{ \pu{).
 	// With neither present the protect/restore passes round-trip the input
 	// unchanged, so skip them: the step 2 scan is O(n^2) in line length and costs
 	// ~90ms on a 26KB single-line message that contains no math at all. This
 	// matters during streaming, where the whole message is reprocessed per frame.
 	if (!LATEX_TRIGGER_REGEXP.test(content)) {
+		if (latexCache.size >= LATEX_CACHE_MAX_SIZE) {
+			latexCache.delete(latexCache.keys().next().value!);
+		}
+		latexCache.set(originalContent, content);
 		return content;
 	}

 	// Step 0: Temporarily remove blockquote markers (>) to process LaTeX correctly
 	// Store the structure so we can restore it later
 	const blockquoteMarkers: Map<number, string> = new Map();
-	const lines = content.split('\n');
+	const lines = content.split(NEWLINE);
 	const processedLines = lines.map((line, index) => {
-		const match = line.match(/^(>\s*)/);
+		const match = line.match(LATEX_BLOCKQUOTE_PREFIX_REGEXP);
 		if (match) {
 			blockquoteMarkers.set(index, match[1]);
 			return line.slice(match[1].length);
 		}
 		return line;
 	});
-	content = processedLines.join('\n');
+	content = processedLines.join(NEWLINE);

 	// Step 1: Protect code blocks
 	const codeBlocks: string[] = [];
@@ -187,58 +226,52 @@ export function preprocessLaTeX(content: string): string {
 	// Match \S...\[...\] and protect them and insert a line-break.
 	// Guarded: with no `\[` present this pattern still probes every start offset,
 	// expanding `.*?` to the end of each line before failing - O(n^2) for nothing.
-	if (content.includes('\\[')) {
-		content = content.replace(
-			/([\S].*?)\\\[([\s\S]*?)\\\](.*)/g,
-			(match, group1, group2, group3) => {
-				// Check if there are characters following the formula (display-formula in a table-cell?)
-				if (group1.endsWith('\\')) {
-					return match; // Backslash before \[, do nothing.
-				}
-				const hasSuffix = /\S/.test(group3);
-				let optBreak;
-
-				if (hasSuffix) {
-					latexExpressions.push(`\\(${group2.trim()}\\)`); // Convert into inline.
-					optBreak = '';
-				} else {
-					latexExpressions.push(`\\[${group2}\\]`);
-					optBreak = '\n';
-				}
-
-				return `${group1}${optBreak}<<LATEX_${latexExpressions.length - 1}>>${optBreak}${group3}`;
+	if (content.includes(LATEX_DISPLAY_OPEN)) {
+		content = content.replace(LATEX_DISPLAY_BLOCK_REGEXP, (match, group1, group2, group3) => {
+			// Check if there are characters following the formula (display-formula in a table-cell?)
+			if (group1.endsWith(LATEX_BACKSLASH)) {
+				return match; // Backslash before \[, do nothing.
 			}
-		);
+			const hasSuffix = LATEX_NON_WHITESPACE_REGEXP.test(group3);
+			let optBreak;
+
+			if (hasSuffix) {
+				latexExpressions.push(`${LATEX_INLINE_OPEN}${group2.trim()}${LATEX_INLINE_CLOSE}`); // Convert into inline.
+				optBreak = '';
+			} else {
+				latexExpressions.push(`${LATEX_DISPLAY_OPEN}${group2}${LATEX_DISPLAY_CLOSE}`);
+				optBreak = NEWLINE;
+			}
+
+			return `${group1}${optBreak}<<LATEX_${latexExpressions.length - 1}>>${optBreak}${group3}`;
+		});
 	}

 	// Match \(...\), \[...\], $$...$$ and protect them
-	content = content.replace(
-		/(\$\$[\s\S]*?\$\$|(?<!\\)\\\[[\s\S]*?\\\]|(?<!\\)\\\(.*?\\\))/g,
-		(match) => {
-			latexExpressions.push(match);
+	content = content.replace(LATEX_PROTECT_REGEXP, (match) => {
+		latexExpressions.push(match);

-			return `<<LATEX_${latexExpressions.length - 1}>>`;
-		}
-	);
+		return `<<LATEX_${latexExpressions.length - 1}>>`;
+	});

 	// Protect inline $...$ but NOT if it looks like money (e.g., $10, $3.99)
 	content = maskInlineLaTeX(content, latexExpressions);

 	// Step 3: Escape standalone $ before digits (currency like $5 → \$5)
 	// (Now that inline math is protected, this will only escape dollars not already protected)
-	content = content.replace(/\$(?=\d)/g, '\\$');
+	content = content.replace(LATEX_CURRENCY_DOLLAR_REGEXP, LATEX_CURRENCY_ESCAPE);

 	// Step 4: Restore protected LaTeX expressions (they are valid)
-	content = content.replace(/<<LATEX_(\d+)>>/g, (_, index) => {
+	content = content.replace(LATEX_PLACEHOLDER_REGEXP, (_, index) => {
 		let expr = latexExpressions[parseInt(index)];
 		const match = expr.match(LATEX_LINEBREAK_REGEXP);
 		if (match) {
 			// Katex: The $$-delimiters should be in their own line
 			// if there are \\-line-breaks.
 			const formula = match[1];
-			const prefix = formula.startsWith('\n') ? '' : '\n';
-			const suffix = formula.endsWith('\n') ? '' : '\n';
-			expr = '$$' + prefix + formula + suffix + '$$';
+			const prefix = formula.startsWith(NEWLINE) ? '' : NEWLINE;
+			const suffix = formula.endsWith(NEWLINE) ? '' : NEWLINE;
+			expr = LATEX_DISPLAY_DELIMITER + prefix + formula + suffix + LATEX_DISPLAY_DELIMITER;
 		}
 		return expr;
 	});
@@ -247,7 +280,7 @@ export function preprocessLaTeX(content: string): string {
 	// This must happen BEFORE restoring code blocks to avoid affecting code content
 	content = escapeBrackets(content);

-	if (doEscapeMhchem && (content.includes('\\ce{') || content.includes('\\pu{'))) {
+	if (doEscapeMhchem && (content.includes(LATEX_MHCHEM_CE) || content.includes(LATEX_MHCHEM_PU))) {
 		content = escapeMhchem(content);
 	}

@@ -257,31 +290,38 @@ export function preprocessLaTeX(content: string): string {
 		// Using the look‑behind pattern `(?<!\\)` we skip matches
 		// that are preceded by a backslash, e.g.
 		// `Definitions\\(also called macros)` (title of chapter 20 in The TeXbook).
-		.replace(/(?<!\\)\\\((.+?)\\\)/g, '$$$1$') // inline
+		.replace(LATEX_INLINE_CONVERT_REGEXP, (_, formula: string) => {
+			return `${LATEX_INLINE_DELIMITER}${formula}${LATEX_INLINE_DELIMITER}`;
+		}) // inline
 		.replace(
 			// Using the look‑behind pattern `(?<!\\)` we skip matches
 			// that are preceded by a backslash, e.g. `\\[4pt]`.
-			/(?<!\\)\\\[([\s\S]*?)\\\]/g, // display, see also PR #16599
-			(_, content: string) => {
-				return `$$${content}$$`;
+			LATEX_DISPLAY_CONVERT_REGEXP, // display, see also PR #16599
+			(_, formula: string) => {
+				return `${LATEX_DISPLAY_DELIMITER}${formula}${LATEX_DISPLAY_DELIMITER}`;
 			}
 		);

 	// Step 7: Restore code blocks
 	// This happens AFTER all LaTeX conversions to preserve code content
-	content = content.replace(/<<CODE_BLOCK_(\d+)>>/g, (_, index) => {
+	content = content.replace(CODE_BLOCK_PLACEHOLDER_REGEXP, (_, index) => {
 		return codeBlocks[parseInt(index)];
 	});

 	// Step 8: Restore blockquote markers
 	if (blockquoteMarkers.size > 0) {
-		const finalLines = content.split('\n');
+		const finalLines = content.split(NEWLINE);
 		const restoredLines = finalLines.map((line, index) => {
 			const marker = blockquoteMarkers.get(index);
 			return marker ? marker + line : line;
 		});
-		content = restoredLines.join('\n');
+		content = restoredLines.join(NEWLINE);
 	}

+	if (latexCache.size >= LATEX_CACHE_MAX_SIZE) {
+		latexCache.delete(latexCache.keys().next().value!);
+	}
+	latexCache.set(originalContent, content);
+
 	return content;
 }
@@ -14,70 +14,94 @@ const JSON_ARRAY_CLOSE = ']';
 // comma when the model cut off mid-key.
 const TRAILING_JSON_PUNCTUATION_REGEX = /,?\s*$/;

+/** Bounded cache for parsePartialJsonArgs results. */
+const PARTIAL_JSON_CACHE_MAX_SIZE = 32;
+const partialJsonCache = new Map<string, Record<string, unknown> | null>();
+
+function cacheResult(input: string, result: Record<string, unknown> | null): void {
+	if (partialJsonCache.size >= PARTIAL_JSON_CACHE_MAX_SIZE) {
+		partialJsonCache.delete(partialJsonCache.keys().next().value!);
+	}
+	partialJsonCache.set(input, result);
+}
+
 // Parse partial tool-arg JSON streamed token-by-token. Closes any
 // unterminated string and dangling open containers (in reverse order),
 // so parsers can still surface keys already received while the call
-// is still in flight.
+// is still in flight. Memoized: the char-by-char scanner runs on every
+// render during streaming even when toolArgs hasn't changed.
 export function parsePartialJsonArgs(toolArgsString: string): Record<string, unknown> | null {
+	const cached = partialJsonCache.get(toolArgsString);
+	if (cached !== undefined) return cached;
+
+	let result: Record<string, unknown> | null;
+
 	try {
 		const parsed: unknown = JSON.parse(toolArgsString);
-		if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
-			return parsed as Record<string, unknown>;
-		}
-		return null;
+		result =
+			parsed && typeof parsed === 'object' && !Array.isArray(parsed)
+				? (parsed as Record<string, unknown>)
+				: null;
 	} catch {
-		let inString = false;
-		let escape = false;
-		const stack: ('{' | '[')[] = [];
+		result = scanPartialJson(toolArgsString);
+	}

-		for (let i = 0; i < toolArgsString.length; i++) {
-			const ch = toolArgsString[i];
-			if (escape) {
-				escape = false;
-				continue;
-			}
-			if (ch === JSON_BACKSLASH && inString) {
-				escape = true;
-				continue;
-			}
-			if (ch === JSON_QUOTE) {
-				inString = !inString;
-				continue;
-			}
-			if (inString) continue;
-			if (ch === JSON_OBJECT_OPEN) stack.push(JSON_OBJECT_OPEN);
-			else if (ch === JSON_OBJECT_CLOSE) {
-				if (stack.length === 0 || stack[stack.length - 1] !== JSON_OBJECT_OPEN) return null;
-				stack.pop();
-			} else if (ch === JSON_ARRAY_OPEN) stack.push(JSON_ARRAY_OPEN);
-			else if (ch === JSON_ARRAY_CLOSE) {
-				if (stack.length === 0 || stack[stack.length - 1] !== JSON_ARRAY_OPEN) return null;
-				stack.pop();
-			}
-		}
+	cacheResult(toolArgsString, result);
+	return result;
+}

-		let completed = toolArgsString;
+/** Char-by-char scanner for unterminated partial JSON. */
+function scanPartialJson(toolArgsString: string): Record<string, unknown> | null {
+	let inString = false;
+	let escape = false;
+	const stack: ('{' | '[')[] = [];
+
+	for (let i = 0; i < toolArgsString.length; i++) {
+		const ch = toolArgsString[i];
 		if (escape) {
-			// Dangling escape at end of partial JSON: escape the trailing
-			// backslash as a literal so we can close the string cleanly.
-			completed += JSON_BACKSLASH;
+			escape = false;
+			continue;
 		}
-		if (inString) completed += JSON_QUOTE;
-		if (!inString) completed = completed.replace(TRAILING_JSON_PUNCTUATION_REGEX, '');
-
-		// Close in reverse nesting order: innermost container first.
-		for (let i = stack.length - 1; i >= 0; i--) {
-			completed += stack[i] === JSON_OBJECT_OPEN ? JSON_OBJECT_CLOSE : JSON_ARRAY_CLOSE;
+		if (ch === JSON_BACKSLASH && inString) {
+			escape = true;
+			continue;
 		}
-
-		try {
-			const parsed: unknown = JSON.parse(completed);
-			if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
-				return parsed as Record<string, unknown>;
-			}
-			return null;
-		} catch {
-			return null;
+		if (ch === JSON_QUOTE) {
+			inString = !inString;
+			continue;
+		}
+		if (inString) continue;
+		if (ch === JSON_OBJECT_OPEN) stack.push(JSON_OBJECT_OPEN);
+		else if (ch === JSON_OBJECT_CLOSE) {
+			if (stack.length === 0 || stack[stack.length - 1] !== JSON_OBJECT_OPEN) return null;
+			stack.pop();
+		} else if (ch === JSON_ARRAY_OPEN) stack.push(JSON_ARRAY_OPEN);
+		else if (ch === JSON_ARRAY_CLOSE) {
+			if (stack.length === 0 || stack[stack.length - 1] !== JSON_ARRAY_OPEN) return null;
+			stack.pop();
 		}
 	}
+
+	let completed = toolArgsString;
+	if (escape) {
+		// Dangling escape at end of partial JSON: escape the trailing
+		// backslash as a literal so we can close the string cleanly.
+		completed += JSON_BACKSLASH;
+	}
+	if (inString) completed += JSON_QUOTE;
+	if (!inString) completed = completed.replace(TRAILING_JSON_PUNCTUATION_REGEX, '');
+
+	// Close in reverse nesting order: innermost container first.
+	for (let i = stack.length - 1; i >= 0; i--) {
+		completed += stack[i] === JSON_OBJECT_OPEN ? JSON_OBJECT_CLOSE : JSON_ARRAY_CLOSE;
+	}
+
+	try {
+		const parsed: unknown = JSON.parse(completed);
+		return parsed && typeof parsed === 'object' && !Array.isArray(parsed)
+			? (parsed as Record<string, unknown>)
+			: null;
+	} catch {
+		return null;
+	}
 }
@@ -155,41 +155,71 @@ function parseChunk(chunk: string): SearchResult | null {
 	return result;
 }

+/** Bounded cache for extractSearchResults results. */
+const SEARCH_RESULTS_CACHE_MAX_SIZE = 32;
+const searchResultsCache = new Map<string, SearchResult[]>();
+
 /**
 * Extract a SearchResult[] from a tool-result string. Returns `[]` when
 * the input does not match the expected shape — useful for branching
 * between dedicated search-results rendering and the generic tool-call
- * block.
+ * block. Memoized: called per render during streaming on unchanged
+ * tool result strings.
 */
 export function extractSearchResults(text: string | undefined | null): SearchResult[] {
 	if (!text) return [];

+	const cached = searchResultsCache.get(text);
+	if (cached) return cached;
+
 	const results: SearchResult[] = [];
 	for (const chunk of splitChunks(text)) {
 		const parsed = parseChunk(chunk);
 		if (parsed) results.push(parsed);
 	}
+
+	if (searchResultsCache.size >= SEARCH_RESULTS_CACHE_MAX_SIZE) {
+		searchResultsCache.delete(searchResultsCache.keys().next().value!);
+	}
+	searchResultsCache.set(text, results);
+
 	return results;
 }

+/** Bounded cache for extractSearchQuery results. */
+const SEARCH_QUERY_CACHE_MAX_SIZE = 32;
+const searchQueryCache = new Map<string, string>();
+
 /**
 * Best-effort extraction of the search query out of a tool call's JSON
 * argument blob. Currently looks for a `query` field (the convention
 * used by Exa and most web-search MCP servers); returns an empty string
- * if it cannot be located.
+ * if it cannot be located. Memoized: called per render during streaming
+ * on unchanged tool args strings.
 */
 export function extractSearchQuery(toolArgs: string | undefined | null): string {
 	if (!toolArgs) return '';
+
+	const cached = searchQueryCache.get(toolArgs);
+	if (cached !== undefined) return cached;
+
+	let result = '';
 	try {
 		const parsed: unknown = JSON.parse(toolArgs);
 		if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
 			const candidate = (parsed as Record<string, unknown>)[SEARCH_TOOL_QUERY_FIELD];
-			if (typeof candidate === 'string') return candidate.trim();
+			if (typeof candidate === 'string') result = candidate.trim();
 		}
 	} catch {
-		return '';
+		result = '';
 	}
-	return '';
+
+	if (searchQueryCache.size >= SEARCH_QUERY_CACHE_MAX_SIZE) {
+		searchQueryCache.delete(searchQueryCache.keys().next().value!);
+	}
+	searchQueryCache.set(toolArgs, result);
+
+	return result;
 }

 /**
@@ -0,0 +1,146 @@
+// Tests for the memoized parseToolCalls and O(1) tool message lookup in
+// deriveAgenticSections. These were added to prevent regressions where
+// streaming text tokens trigger redundant JSON.parse calls on unchanged
+// tool call data.
+
+import { describe, it, expect, vi } from 'vitest';
+import { deriveAgenticSections } from '$lib/utils/agentic';
+import type { ApiChatCompletionToolCall } from '$lib/types/api';
+import type { DatabaseMessage } from '$lib/types/database';
+import { MessageRole, AgenticSectionType } from '$lib/enums';
+
+function makeMessage(overrides: Partial<DatabaseMessage>): DatabaseMessage {
+	return {
+		id: 'm1',
+		convId: 'c1',
+		type: 'text',
+		timestamp: 0,
+		role: MessageRole.ASSISTANT,
+		content: '',
+		parent: null,
+		children: [],
+		...overrides
+	} as DatabaseMessage;
+}
+
+describe('parseToolCalls memoization', () => {
+	it('returns the same array reference for the same JSON string', () => {
+		// parseToolCalls is not exported, but deriveAgenticSections uses it
+		// internally. We verify memoization through behavior: calling
+		// deriveAgenticSections twice with the same toolCalls should not
+		// re-parse (which we verify by checking the returned sections
+		// are equivalent).
+		const toolCallsJson = JSON.stringify([
+			{ id: 'call_1', type: 'function', function: { name: 'test', arguments: '{}' } }
+		]);
+
+		const msg = makeMessage({ content: 'hello', toolCalls: toolCallsJson });
+		const sections1 = deriveAgenticSections(msg, [], [], false);
+		const sections2 = deriveAgenticSections(msg, [], [], false);
+
+		expect(sections1).toHaveLength(sections2.length);
+		expect(sections1[0].type).toBe(sections2[0].type);
+	});
+
+	it('does not re-parse JSON on cache hit', () => {
+		const toolCallsJson = JSON.stringify([
+			{ id: 'call_1', type: 'function', function: { name: 'test', arguments: '{}' } }
+		]);
+
+		const msg = makeMessage({ content: 'hello', toolCalls: toolCallsJson });
+		const spy = vi.spyOn(JSON, 'parse');
+
+		deriveAgenticSections(msg, [], [], false);
+		const callsAfterFirst = spy.mock.calls.length;
+
+		deriveAgenticSections(msg, [], [], false);
+		expect(spy.mock.calls.length).toBe(callsAfterFirst);
+
+		spy.mockRestore();
+	});
+
+	it('handles empty/undefined toolCalls without error', () => {
+		const msg = makeMessage({ content: 'hello' });
+		const sections = deriveAgenticSections(msg, [], [], false);
+
+		expect(sections).toHaveLength(1);
+		expect(sections[0].type).toBe(AgenticSectionType.TEXT);
+	});
+
+	it('handles invalid JSON gracefully', () => {
+		const msg = makeMessage({ content: 'hello', toolCalls: '{invalid' });
+		const sections = deriveAgenticSections(msg, [], [], false);
+
+		// Should return just the text section, no tool call sections
+		expect(sections).toHaveLength(1);
+		expect(sections[0].type).toBe(AgenticSectionType.TEXT);
+	});
+});
+
+describe('deriveAgenticSections O(1) tool message lookup', () => {
+	it('matches tool messages to tool calls by toolCallId', () => {
+		const toolCallsJson = JSON.stringify([
+			{ id: 'call_1', type: 'function', function: { name: 'test_1', arguments: '{}' } },
+			{ id: 'call_2', type: 'function', function: { name: 'test_2', arguments: '{}' } }
+		]);
+
+		const toolMessages = [
+			makeMessage({ role: MessageRole.TOOL, toolCallId: 'call_1', content: 'result_1' }),
+			makeMessage({ role: MessageRole.TOOL, toolCallId: 'call_2', content: 'result_2' })
+		];
+
+		const msg = makeMessage({ content: 'hello', toolCalls: toolCallsJson });
+		const sections = deriveAgenticSections(msg, toolMessages, [], false);
+
+		// Expect: TEXT + 2 TOOL_CALL sections
+		const toolCallSections = sections.filter((s) => s.type === AgenticSectionType.TOOL_CALL);
+		expect(toolCallSections).toHaveLength(2);
+		expect(toolCallSections[0].toolResult).toBe('result_1');
+		expect(toolCallSections[1].toolResult).toBe('result_2');
+	});
+
+	it('handles missing tool messages (pending calls during streaming)', () => {
+		const toolCallsJson = JSON.stringify([
+			{ id: 'call_1', type: 'function', function: { name: 'test', arguments: '{}' } }
+		]);
+
+		const msg = makeMessage({ content: '', toolCalls: toolCallsJson });
+		const sections = deriveAgenticSections(msg, [], [], true);
+
+		const toolCallSection = sections.find((s) => s.type === AgenticSectionType.TOOL_CALL_PENDING);
+		expect(toolCallSection).toBeDefined();
+		expect(toolCallSection?.content).toBe('');
+	});
+
+	it('scales with many tool calls (no O(n^2) blowup)', () => {
+		const N = 100;
+		const toolCalls = Array.from(
+			{ length: N },
+			(_, i): ApiChatCompletionToolCall => ({
+				id: `call_${i}`,
+				type: 'function',
+				function: { name: `tool_${i}`, arguments: '{}' }
+			})
+		);
+		const toolCallsJson = JSON.stringify(toolCalls);
+
+		const toolMessages = Array.from({ length: N }, (_, i) =>
+			makeMessage({
+				role: MessageRole.TOOL,
+				toolCallId: `call_${i}`,
+				content: `result_${i}`
+			})
+		);
+
+		const msg = makeMessage({ content: 'hello', toolCalls: toolCallsJson });
+
+		// If the lookup were still O(n^2), this would be noticeably slow
+		const start = Date.now();
+		const sections = deriveAgenticSections(msg, toolMessages, [], false);
+		const elapsed = Date.now() - start;
+
+		const toolCallSections = sections.filter((s) => s.type === AgenticSectionType.TOOL_CALL);
+		expect(toolCallSections).toHaveLength(N);
+		expect(elapsed).toBeLessThan(100); // Should be fast with O(1) lookup
+	});
+});
Author	SHA1	Message	Date
YiChen Lv	96ece714c4	metal : per-device tuned (Q, NE) for flash-attn vec (#25750 ) * rebase Q-generic FA vec body from `01dc93607` (#23114) * add 53 f16 (Q,NE) flash-attn vec instantiations (vec 80 -> 133) * add FA vec (Q,NE) tuning table + dispatch wiring + SMEM cap fallback * add FA vec (Q,NE) perf sweep * fill tuning result * fold family table into a per-family representative SKU * refactor tuning result format * extend FA vec tuning to quantized KV caches * sync fa vec tuner bucketing with runtime, use pointwise tuning regret * update tuned table * format and cleanup * prefix fa_vec tuning procs with ggml_backend_metal_tuning_, drop unused fa_vec_override_active	2026-07-29 21:46:01 +08:00
Georgi Gerganov	7185fef97d	ggml-metal: FWHT kernel for metal backend (#25924 )	2026-07-28 19:05:42 +03:00
Georgi Gerganov	a9a5d1da9c	metal: fuse snake activation (mul, sin, sqr, mul, add) (#25459 )	2026-07-28 19:02:22 +03:00
Georgi Gerganov	70f70ed5e3	metal : add Q2_0 support (#25419 )	2026-07-28 19:00:04 +03:00
Georgi Gerganov	ccb8445393	metal : add CONV_2D_DW (depthwise convolution) support (#21565 )	2026-07-28 18:48:08 +03:00
Georgi Gerganov	6d75564519	metal : add set_rows with src0 f16 (#25434 )	2026-07-28 18:48:08 +03:00
Georgi Gerganov	2894406768	metal: add col2im_1d op (f32/f16/bf16) (#25176 )	2026-07-28 18:48:08 +03:00
YiChen Lv	8ed30a710c	metal : per-op source split + parallel compile (#24021 ) * preliminary extract common header * op source split * split metallib into 8 libs && load in parallel * derive kernel->library routing from functionNames * x-macro lib list + underscore filenames, dedup QK_NL, MRC fixes * op source split 8 to 20 * improve robustness of source fallback * clean up * change bool -> atomic_bool * only prepend headers that source actually includes * no semaphore, use GCD global queue * dedup library compile path, fix NSError lifetime, rename gla * relocate upstream concat/rope_back/repeat kernel changes into split files * move ggml-common.h from common.h into dequantize.h to shrink binary size --------- Co-authored-by: lvyichen <lvyichen@stepfun.com>	2026-07-28 18:48:00 +03:00
Daniel Bevenius	7e1e28cae3	mtmd : add Nemotron 3 Nano Omni support (parakeet) (#22520 ) * mtmd : add Nemotron 3 Nano Omni support (parakeet) This commit adds support for the subsampling and encoder part of Nemotron Nemo 3 omni model. The Parakeet subsampling/encoder were taken from parakeet.cpp which is currently a pull request against whisper.cpp. I've tried to copy the code a close as possible to hopefully enable easy patching between the these two project later. Refs: https://github.com/ggml-org/whisper.cpp/pull/3735 * mtmd : generate rel pos tensor in graph instead of in conversion [no ci] This commit removes the generation of the relative positional tensor in the model conversion script and instead computes it in the encoder graph. This is only done for the window of positions required for the current audio sample. * mtmd : add clip_get_model to clip API [no ci] This commit adds a function to get access to the clip_model. It also removes the two functions clip_get_mel_filter_tensor, and clip_get_window_tensor(const struct clip_ctx * ctx) which can now use clip_get_model to access the model tensors that it needs. * mtmd : read mel_filters and window into hparams * mtmd : use set_input_f32 lambda [no ci] * mtmd : add better asserts for mel_filters and hann window [no ci] * mtmd : add missing size_t cast * mtmd : change type of pad to size_t * mtmd : zero initialize samples_padded * mtmd : remove unsued ctx member from parakeet preprocessor * mtmd : make log_mel_spectrogram_parakeet_worker_thread private static * mtmd : sync/update parakeeet impl with latest whisper.cpp This commit updates the parakeet code in mtmd to reflect the latest updates to parakeet.cpp in whisper.cpp. A follow up commit will address the currently hardcoded dw_pad and see if we can add n_conv_kernel as a model metadata field. * mtmd : add audio_conv_kernel_size to model conversion This commit updates the model conversion to read the conv_kernel_size field from the sound_config section of the models config.json file. It then uses this field instead of the hardcoded values in parakeet.cpp. * mtmd : cleanup [no ci] * conversion : call super().filter_tensors [no ci] * do not discard result of super filter_tensors * mtmd : use build_mm instead of ggml_mul_mat * mtmd : use build_ffn * mtmd : move and reuse get_vector lambda * mtmd : use build_inp_raw for parakeet * mtmd : throw exception in get_scalar instead of assert * mtmd : fix std::min call * mtmt : use .c_str in throw clause in get_vector * mtmd : check for F32 type and non-empty tensor in get_vector The get_vector lambda is used by get_scalar but also standalone to read in the mel_filters and the window data. Therefor we are not checking for 1D tensors but allowing multiple dimensions. We do have a check in get_scalar to verify the size of the vector. * mtmd : replace hardcoded 1101 for n_tokens_real * mtmd : assert subsampling_factor is 8 This commit adds an assert of the parakeet subsampling factor to check that it is 8. The motivation for this is that this model currently has three convolutions with a stride of 2. If the underlying model updates the subsampling factor these convolution operations will need to be updated and this will produce and error if this occurs. * mtmd : remove unused ggml_tensors attn_pos_w and mm_norm_w * mtmd : remove single thread path This commit removes the single thread path which was a left over from the original parakeet.cpp where n_threads is configurable. * fix some security issues --------- Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com> Co-authored-by: Xuan Son Nguyen <son@huggingface.co>	2026-07-28 17:20:25 +02:00
Aleksander Grygier	6e2bc65fb2	ui: rendering performance follow-up (#26097 )	2026-07-28 17:13:25 +02:00
Julien Jerphanion	ad77bd31a6	docs: Adapt conda-forge package name (#26229 ) Co-authored-by: dev-tinker <dev-tinker@users.noreply.github.com>	2026-07-28 16:51:20 +02:00