npm - local-llm-rn - Versions diffs - 1.0.0 - Mend

local-llm-rn 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (626) hide show

package/cpp/CMakeLists.txt +285 -0
package/cpp/common/CMakeLists.txt +149 -0
package/cpp/common/arg.cpp +3799 -0
package/cpp/common/arg.h +131 -0
package/cpp/common/base64.hpp +392 -0
package/cpp/common/build-info.cpp.in +4 -0
package/cpp/common/chat-parser-xml-toolcall.cpp +879 -0
package/cpp/common/chat-parser-xml-toolcall.h +45 -0
package/cpp/common/chat-parser.cpp +1649 -0
package/cpp/common/chat-parser.h +133 -0
package/cpp/common/chat-peg-parser.cpp +124 -0
package/cpp/common/chat-peg-parser.h +105 -0
package/cpp/common/chat.cpp +3355 -0
package/cpp/common/chat.h +252 -0
package/cpp/common/common.cpp +1824 -0
package/cpp/common/common.h +930 -0
package/cpp/common/console.cpp +1137 -0
package/cpp/common/console.h +41 -0
package/cpp/common/debug.cpp +167 -0
package/cpp/common/debug.h +43 -0
package/cpp/common/download.cpp +792 -0
package/cpp/common/download.h +84 -0
package/cpp/common/http.h +84 -0
package/cpp/common/jinja/README.md +88 -0
package/cpp/common/jinja/caps.cpp +285 -0
package/cpp/common/jinja/caps.h +30 -0
package/cpp/common/jinja/lexer.cpp +341 -0
package/cpp/common/jinja/lexer.h +157 -0
package/cpp/common/jinja/parser.cpp +591 -0
package/cpp/common/jinja/parser.h +21 -0
package/cpp/common/jinja/runtime.cpp +867 -0
package/cpp/common/jinja/runtime.h +638 -0
package/cpp/common/jinja/string.cpp +213 -0
package/cpp/common/jinja/string.h +61 -0
package/cpp/common/jinja/utils.h +149 -0
package/cpp/common/jinja/value.cpp +1393 -0
package/cpp/common/jinja/value.h +756 -0
package/cpp/common/json-partial.cpp +324 -0
package/cpp/common/json-partial.h +39 -0
package/cpp/common/json-schema-to-grammar.cpp +1153 -0
package/cpp/common/json-schema-to-grammar.h +43 -0
package/cpp/common/llguidance.cpp +258 -0
package/cpp/common/log.cpp +446 -0
package/cpp/common/log.h +119 -0
package/cpp/common/ngram-cache.cpp +285 -0
package/cpp/common/ngram-cache.h +101 -0
package/cpp/common/ngram-map.cpp +530 -0
package/cpp/common/ngram-map.h +115 -0
package/cpp/common/ngram-mod.cpp +60 -0
package/cpp/common/ngram-mod.h +38 -0
package/cpp/common/peg-parser.cpp +1712 -0
package/cpp/common/peg-parser.h +459 -0
package/cpp/common/preset.cpp +483 -0
package/cpp/common/preset.h +83 -0
package/cpp/common/regex-partial.cpp +204 -0
package/cpp/common/regex-partial.h +56 -0
package/cpp/common/sampling.cpp +745 -0
package/cpp/common/sampling.h +119 -0
package/cpp/common/speculative.cpp +1074 -0
package/cpp/common/speculative.h +41 -0
package/cpp/common/unicode.cpp +64 -0
package/cpp/common/unicode.h +22 -0
package/cpp/ggml/CMakeLists.txt +494 -0
package/cpp/ggml/cmake/GitVars.cmake +22 -0
package/cpp/ggml/cmake/common.cmake +50 -0
package/cpp/ggml/cmake/ggml-config.cmake.in +191 -0
package/cpp/ggml/include/ggml-alloc.h +85 -0
package/cpp/ggml/include/ggml-backend.h +373 -0
package/cpp/ggml/include/ggml-blas.h +25 -0
package/cpp/ggml/include/ggml-cann.h +123 -0
package/cpp/ggml/include/ggml-cpp.h +39 -0
package/cpp/ggml/include/ggml-cpu.h +151 -0
package/cpp/ggml/include/ggml-cuda.h +47 -0
package/cpp/ggml/include/ggml-hexagon.h +19 -0
package/cpp/ggml/include/ggml-metal.h +61 -0
package/cpp/ggml/include/ggml-opencl.h +26 -0
package/cpp/ggml/include/ggml-opt.h +256 -0
package/cpp/ggml/include/ggml-rpc.h +30 -0
package/cpp/ggml/include/ggml-sycl.h +49 -0
package/cpp/ggml/include/ggml-virtgpu.h +14 -0
package/cpp/ggml/include/ggml-vulkan.h +29 -0
package/cpp/ggml/include/ggml-webgpu.h +19 -0
package/cpp/ggml/include/ggml-zdnn.h +17 -0
package/cpp/ggml/include/ggml-zendnn.h +22 -0
package/cpp/ggml/include/ggml.h +2753 -0
package/cpp/ggml/include/gguf.h +204 -0
package/cpp/ggml/src/CMakeLists.txt +492 -0
package/cpp/ggml/src/ggml-alloc.c +1244 -0
package/cpp/ggml/src/ggml-backend-dl.cpp +48 -0
package/cpp/ggml/src/ggml-backend-dl.h +45 -0
package/cpp/ggml/src/ggml-backend-impl.h +255 -0
package/cpp/ggml/src/ggml-backend-reg.cpp +566 -0
package/cpp/ggml/src/ggml-backend.cpp +2270 -0
package/cpp/ggml/src/ggml-blas/CMakeLists.txt +101 -0
package/cpp/ggml/src/ggml-blas/ggml-blas.cpp +518 -0
package/cpp/ggml/src/ggml-common.h +1878 -0
package/cpp/ggml/src/ggml-cpu/CMakeLists.txt +691 -0
package/cpp/ggml/src/ggml-cpu/amx/amx.cpp +247 -0
package/cpp/ggml/src/ggml-cpu/amx/amx.h +8 -0
package/cpp/ggml/src/ggml-cpu/amx/common.h +91 -0
package/cpp/ggml/src/ggml-cpu/amx/mmq.cpp +2512 -0
package/cpp/ggml/src/ggml-cpu/amx/mmq.h +10 -0
package/cpp/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +98 -0
package/cpp/ggml/src/ggml-cpu/arch/arm/quants.c +4052 -0
package/cpp/ggml/src/ggml-cpu/arch/arm/repack.cpp +4935 -0
package/cpp/ggml/src/ggml-cpu/arch/loongarch/quants.c +2159 -0
package/cpp/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
package/cpp/ggml/src/ggml-cpu/arch/powerpc/quants.c +2305 -0
package/cpp/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
package/cpp/ggml/src/ggml-cpu/arch/riscv/quants.c +2726 -0
package/cpp/ggml/src/ggml-cpu/arch/riscv/repack.cpp +342 -0
package/cpp/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
package/cpp/ggml/src/ggml-cpu/arch/s390/quants.c +1468 -0
package/cpp/ggml/src/ggml-cpu/arch/wasm/quants.c +1221 -0
package/cpp/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
package/cpp/ggml/src/ggml-cpu/arch/x86/quants.c +3820 -0
package/cpp/ggml/src/ggml-cpu/arch/x86/repack.cpp +6307 -0
package/cpp/ggml/src/ggml-cpu/arch-fallback.h +313 -0
package/cpp/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
package/cpp/ggml/src/ggml-cpu/binary-ops.h +16 -0
package/cpp/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
package/cpp/ggml/src/ggml-cpu/common.h +95 -0
package/cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +529 -0
package/cpp/ggml/src/ggml-cpu/ggml-cpu.c +3734 -0
package/cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +701 -0
package/cpp/ggml/src/ggml-cpu/hbm.cpp +55 -0
package/cpp/ggml/src/ggml-cpu/hbm.h +8 -0
package/cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +938 -0
package/cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +90 -0
package/cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +798 -0
package/cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
package/cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4033 -0
package/cpp/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
package/cpp/ggml/src/ggml-cpu/ops.cpp +10978 -0
package/cpp/ggml/src/ggml-cpu/ops.h +116 -0
package/cpp/ggml/src/ggml-cpu/quants.c +1193 -0
package/cpp/ggml/src/ggml-cpu/quants.h +97 -0
package/cpp/ggml/src/ggml-cpu/repack.cpp +3316 -0
package/cpp/ggml/src/ggml-cpu/repack.h +173 -0
package/cpp/ggml/src/ggml-cpu/simd-gemm.h +136 -0
package/cpp/ggml/src/ggml-cpu/simd-mappings.h +1279 -0
package/cpp/ggml/src/ggml-cpu/spacemit/ime.cpp +1025 -0
package/cpp/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
package/cpp/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
package/cpp/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
package/cpp/ggml/src/ggml-cpu/traits.cpp +36 -0
package/cpp/ggml/src/ggml-cpu/traits.h +38 -0
package/cpp/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
package/cpp/ggml/src/ggml-cpu/unary-ops.h +35 -0
package/cpp/ggml/src/ggml-cpu/vec.cpp +629 -0
package/cpp/ggml/src/ggml-cpu/vec.h +1585 -0
package/cpp/ggml/src/ggml-hexagon/CMakeLists.txt +117 -0
package/cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp +3232 -0
package/cpp/ggml/src/ggml-hexagon/htp/CMakeLists.txt +45 -0
package/cpp/ggml/src/ggml-hexagon/htp/act-ops.c +815 -0
package/cpp/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
package/cpp/ggml/src/ggml-hexagon/htp/binary-ops.c +827 -0
package/cpp/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +157 -0
package/cpp/ggml/src/ggml-hexagon/htp/cpy-ops.c +251 -0
package/cpp/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +666 -0
package/cpp/ggml/src/ggml-hexagon/htp/get-rows-ops.c +111 -0
package/cpp/ggml/src/ggml-hexagon/htp/hex-dma.c +63 -0
package/cpp/ggml/src/ggml-hexagon/htp/hex-dma.h +182 -0
package/cpp/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
package/cpp/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
package/cpp/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
package/cpp/ggml/src/ggml-hexagon/htp/htp-ctx.h +35 -0
package/cpp/ggml/src/ggml-hexagon/htp/htp-msg.h +154 -0
package/cpp/ggml/src/ggml-hexagon/htp/htp-ops.h +65 -0
package/cpp/ggml/src/ggml-hexagon/htp/htp_iface.idl +16 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-arith.h +470 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-base.h +173 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-div.h +116 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-inverse.h +176 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-reduce.h +266 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
package/cpp/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -0
package/cpp/ggml/src/ggml-hexagon/htp/main.c +1150 -0
package/cpp/ggml/src/ggml-hexagon/htp/matmul-ops.c +2595 -0
package/cpp/ggml/src/ggml-hexagon/htp/rope-ops.c +498 -0
package/cpp/ggml/src/ggml-hexagon/htp/set-rows-ops.c +167 -0
package/cpp/ggml/src/ggml-hexagon/htp/softmax-ops.c +421 -0
package/cpp/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +130 -0
package/cpp/ggml/src/ggml-hexagon/htp/unary-ops.c +384 -0
package/cpp/ggml/src/ggml-hexagon/htp/worker-pool.c +293 -0
package/cpp/ggml/src/ggml-hexagon/htp/worker-pool.h +57 -0
package/cpp/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
package/cpp/ggml/src/ggml-hexagon/htp-drv.h +121 -0
package/cpp/ggml/src/ggml-hexagon/libdl.h +79 -0
package/cpp/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
package/cpp/ggml/src/ggml-hexagon/op-desc.h +153 -0
package/cpp/ggml/src/ggml-impl.h +724 -0
package/cpp/ggml/src/ggml-metal/CMakeLists.txt +124 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-common.cpp +457 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-context.m +702 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-device.cpp +1890 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-device.h +290 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-device.m +1749 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-impl.h +1054 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-ops.cpp +4370 -0
package/cpp/ggml/src/ggml-metal/ggml-metal-ops.h +94 -0
package/cpp/ggml/src/ggml-metal/ggml-metal.cpp +937 -0
package/cpp/ggml/src/ggml-metal/ggml-metal.metal +9819 -0
package/cpp/ggml/src/ggml-musa/CMakeLists.txt +125 -0
package/cpp/ggml/src/ggml-musa/mudnn.cu +112 -0
package/cpp/ggml/src/ggml-musa/mudnn.cuh +12 -0
package/cpp/ggml/src/ggml-opencl/CMakeLists.txt +150 -0
package/cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +11553 -0
package/cpp/ggml/src/ggml-opencl/kernels/add.cl +190 -0
package/cpp/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
package/cpp/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
package/cpp/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
package/cpp/ggml/src/ggml-opencl/kernels/concat.cl +51 -0
package/cpp/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
package/cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
package/cpp/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
package/cpp/ggml/src/ggml-opencl/kernels/cvt.cl +417 -0
package/cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
package/cpp/ggml/src/ggml-opencl/kernels/div.cl +138 -0
package/cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
package/cpp/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
package/cpp/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +371 -0
package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
package/cpp/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
package/cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
package/cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl +268 -0
package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl +274 -0
package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
package/cpp/ggml/src/ggml-opencl/kernels/get_rows.cl +187 -0
package/cpp/ggml/src/ggml-opencl/kernels/glu.cl +378 -0
package/cpp/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
package/cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
package/cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
package/cpp/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +147 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
package/cpp/ggml/src/ggml-opencl/kernels/norm.cl +161 -0
package/cpp/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
package/cpp/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
package/cpp/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
package/cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl +190 -0
package/cpp/ggml/src/ggml-opencl/kernels/rope.cl +747 -0
package/cpp/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
package/cpp/ggml/src/ggml-opencl/kernels/set_rows.cl +208 -0
package/cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
package/cpp/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
package/cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
package/cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
package/cpp/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
package/cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
package/cpp/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
package/cpp/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
package/cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
package/cpp/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
package/cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
package/cpp/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
package/cpp/ggml/src/ggml-opencl/kernels/transpose.cl +117 -0
package/cpp/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
package/cpp/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
package/cpp/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
package/cpp/ggml/src/ggml-opt.cpp +1093 -0
package/cpp/ggml/src/ggml-quants.c +5325 -0
package/cpp/ggml/src/ggml-quants.h +106 -0
package/cpp/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
package/cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +2118 -0
package/cpp/ggml/src/ggml-threading.cpp +12 -0
package/cpp/ggml/src/ggml-threading.h +14 -0
package/cpp/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
package/cpp/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
package/cpp/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
package/cpp/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
package/cpp/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
package/cpp/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
package/cpp/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
package/cpp/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
package/cpp/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
package/cpp/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
package/cpp/ggml/src/ggml-webgpu/CMakeLists.txt +80 -0
package/cpp/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1231 -0
package/cpp/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3150 -0
package/cpp/ggml/src/ggml-webgpu/pre_wgsl.hpp +778 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +107 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +923 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +107 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +182 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +636 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +668 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +323 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +713 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +103 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +138 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +188 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +194 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +123 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/rope.tmpl.wgsl +295 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +63 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +109 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl +345 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
package/cpp/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
package/cpp/ggml/src/ggml-zdnn/common.hpp +59 -0
package/cpp/ggml/src/ggml-zdnn/ggml-zdnn.cpp +633 -0
package/cpp/ggml/src/ggml-zdnn/mmf.cpp +80 -0
package/cpp/ggml/src/ggml-zdnn/mmf.hpp +12 -0
package/cpp/ggml/src/ggml-zdnn/utils.cpp +79 -0
package/cpp/ggml/src/ggml-zdnn/utils.hpp +19 -0
package/cpp/ggml/src/ggml-zendnn/CMakeLists.txt +92 -0
package/cpp/ggml/src/ggml-zendnn/ggml-zendnn.cpp +469 -0
package/cpp/ggml/src/ggml.c +7669 -0
package/cpp/ggml/src/ggml.cpp +26 -0
package/cpp/ggml/src/gguf.cpp +1699 -0
package/cpp/include/llama-cpp.h +32 -0
package/cpp/include/llama.h +1568 -0
package/cpp/mtmd/CMakeLists.txt +98 -0
package/cpp/mtmd/README.md +63 -0
package/cpp/mtmd/clip-graph.h +117 -0
package/cpp/mtmd/clip-impl.h +586 -0
package/cpp/mtmd/clip-model.h +390 -0
package/cpp/mtmd/clip.cpp +4154 -0
package/cpp/mtmd/clip.h +121 -0
package/cpp/mtmd/deprecation-warning.cpp +22 -0
package/cpp/mtmd/legacy-models/convert_image_encoder_to_gguf.py +412 -0
package/cpp/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py +280 -0
package/cpp/mtmd/legacy-models/glmedge-surgery.py +33 -0
package/cpp/mtmd/legacy-models/llava_surgery.py +38 -0
package/cpp/mtmd/legacy-models/llava_surgery_v2.py +180 -0
package/cpp/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py +892 -0
package/cpp/mtmd/legacy-models/minicpmv-surgery.py +47 -0
package/cpp/mtmd/models/cogvlm.cpp +98 -0
package/cpp/mtmd/models/conformer.cpp +216 -0
package/cpp/mtmd/models/glm4v.cpp +122 -0
package/cpp/mtmd/models/internvl.cpp +69 -0
package/cpp/mtmd/models/kimik25.cpp +101 -0
package/cpp/mtmd/models/kimivl.cpp +63 -0
package/cpp/mtmd/models/llama4.cpp +96 -0
package/cpp/mtmd/models/llava.cpp +374 -0
package/cpp/mtmd/models/minicpmv.cpp +114 -0
package/cpp/mtmd/models/mobilenetv5.cpp +451 -0
package/cpp/mtmd/models/models.h +128 -0
package/cpp/mtmd/models/nemotron-v2-vl.cpp +35 -0
package/cpp/mtmd/models/paddleocr.cpp +52 -0
package/cpp/mtmd/models/pixtral.cpp +86 -0
package/cpp/mtmd/models/qwen2vl.cpp +183 -0
package/cpp/mtmd/models/qwen3vl.cpp +193 -0
package/cpp/mtmd/models/siglip.cpp +86 -0
package/cpp/mtmd/models/whisper-enc.cpp +115 -0
package/cpp/mtmd/models/youtuvl.cpp +179 -0
package/cpp/mtmd/mtmd-audio.cpp +730 -0
package/cpp/mtmd/mtmd-audio.h +113 -0
package/cpp/mtmd/mtmd-cli.cpp +437 -0
package/cpp/mtmd/mtmd-helper.cpp +521 -0
package/cpp/mtmd/mtmd-helper.h +96 -0
package/cpp/mtmd/mtmd.cpp +1156 -0
package/cpp/mtmd/mtmd.h +319 -0
package/cpp/mtmd/requirements.txt +5 -0
package/cpp/mtmd/test-1.jpeg +0 -0
package/cpp/mtmd/test-2.mp3 +0 -0
package/cpp/mtmd/tests.sh +192 -0
package/cpp/src/CMakeLists.txt +169 -0
package/cpp/src/llama-adapter.cpp +488 -0
package/cpp/src/llama-adapter.h +89 -0
package/cpp/src/llama-arch.cpp +2855 -0
package/cpp/src/llama-arch.h +619 -0
package/cpp/src/llama-batch.cpp +917 -0
package/cpp/src/llama-batch.h +173 -0
package/cpp/src/llama-chat.cpp +896 -0
package/cpp/src/llama-chat.h +71 -0
package/cpp/src/llama-context.cpp +3512 -0
package/cpp/src/llama-context.h +359 -0
package/cpp/src/llama-cparams.cpp +5 -0
package/cpp/src/llama-cparams.h +44 -0
package/cpp/src/llama-grammar.cpp +1464 -0
package/cpp/src/llama-grammar.h +194 -0
package/cpp/src/llama-graph.cpp +2685 -0
package/cpp/src/llama-graph.h +1026 -0
package/cpp/src/llama-hparams.cpp +234 -0
package/cpp/src/llama-hparams.h +339 -0
package/cpp/src/llama-impl.cpp +171 -0
package/cpp/src/llama-impl.h +73 -0
package/cpp/src/llama-io.cpp +15 -0
package/cpp/src/llama-io.h +35 -0
package/cpp/src/llama-kv-cache-iswa.cpp +330 -0
package/cpp/src/llama-kv-cache-iswa.h +137 -0
package/cpp/src/llama-kv-cache.cpp +2271 -0
package/cpp/src/llama-kv-cache.h +388 -0
package/cpp/src/llama-kv-cells.h +533 -0
package/cpp/src/llama-memory-hybrid-iswa.cpp +275 -0
package/cpp/src/llama-memory-hybrid-iswa.h +140 -0
package/cpp/src/llama-memory-hybrid.cpp +268 -0
package/cpp/src/llama-memory-hybrid.h +139 -0
package/cpp/src/llama-memory-recurrent.cpp +1165 -0
package/cpp/src/llama-memory-recurrent.h +182 -0
package/cpp/src/llama-memory.cpp +59 -0
package/cpp/src/llama-memory.h +122 -0
package/cpp/src/llama-mmap.cpp +785 -0
package/cpp/src/llama-mmap.h +92 -0
package/cpp/src/llama-model-loader.cpp +1414 -0
package/cpp/src/llama-model-loader.h +203 -0
package/cpp/src/llama-model-saver.cpp +286 -0
package/cpp/src/llama-model-saver.h +37 -0
package/cpp/src/llama-model.cpp +9253 -0
package/cpp/src/llama-model.h +576 -0
package/cpp/src/llama-quant.cpp +1119 -0
package/cpp/src/llama-quant.h +1 -0
package/cpp/src/llama-sampler.cpp +3885 -0
package/cpp/src/llama-sampler.h +42 -0
package/cpp/src/llama-vocab.cpp +3970 -0
package/cpp/src/llama-vocab.h +187 -0
package/cpp/src/llama.cpp +1313 -0
package/cpp/src/models/afmoe.cpp +191 -0
package/cpp/src/models/apertus.cpp +125 -0
package/cpp/src/models/arcee.cpp +135 -0
package/cpp/src/models/arctic.cpp +138 -0
package/cpp/src/models/arwkv7.cpp +86 -0
package/cpp/src/models/baichuan.cpp +122 -0
package/cpp/src/models/bailingmoe.cpp +144 -0
package/cpp/src/models/bailingmoe2.cpp +135 -0
package/cpp/src/models/bert.cpp +178 -0
package/cpp/src/models/bitnet.cpp +160 -0
package/cpp/src/models/bloom.cpp +101 -0
package/cpp/src/models/chameleon.cpp +178 -0
package/cpp/src/models/chatglm.cpp +132 -0
package/cpp/src/models/codeshell.cpp +111 -0
package/cpp/src/models/cogvlm.cpp +102 -0
package/cpp/src/models/cohere2-iswa.cpp +134 -0
package/cpp/src/models/command-r.cpp +122 -0
package/cpp/src/models/dbrx.cpp +123 -0
package/cpp/src/models/deci.cpp +135 -0
package/cpp/src/models/deepseek.cpp +144 -0
package/cpp/src/models/deepseek2.cpp +262 -0
package/cpp/src/models/delta-net-base.cpp +376 -0
package/cpp/src/models/dots1.cpp +134 -0
package/cpp/src/models/dream.cpp +105 -0
package/cpp/src/models/ernie4-5-moe.cpp +150 -0
package/cpp/src/models/ernie4-5.cpp +110 -0
package/cpp/src/models/eurobert.cpp +97 -0
package/cpp/src/models/exaone-moe.cpp +146 -0
package/cpp/src/models/exaone.cpp +114 -0
package/cpp/src/models/exaone4.cpp +123 -0
package/cpp/src/models/falcon-h1.cpp +111 -0
package/cpp/src/models/falcon.cpp +120 -0
package/cpp/src/models/gemma-embedding.cpp +116 -0
package/cpp/src/models/gemma.cpp +112 -0
package/cpp/src/models/gemma2-iswa.cpp +128 -0
package/cpp/src/models/gemma3.cpp +155 -0
package/cpp/src/models/gemma3n-iswa.cpp +384 -0
package/cpp/src/models/glm4-moe.cpp +170 -0
package/cpp/src/models/glm4.cpp +157 -0
package/cpp/src/models/gpt2.cpp +105 -0
package/cpp/src/models/gptneox.cpp +144 -0
package/cpp/src/models/granite-hybrid.cpp +196 -0
package/cpp/src/models/granite.cpp +211 -0
package/cpp/src/models/grok.cpp +159 -0
package/cpp/src/models/grovemoe.cpp +141 -0
package/cpp/src/models/hunyuan-dense.cpp +132 -0
package/cpp/src/models/hunyuan-moe.cpp +154 -0
package/cpp/src/models/internlm2.cpp +120 -0
package/cpp/src/models/jais.cpp +86 -0
package/cpp/src/models/jais2.cpp +123 -0
package/cpp/src/models/jamba.cpp +106 -0
package/cpp/src/models/kimi-linear.cpp +392 -0
package/cpp/src/models/lfm2.cpp +190 -0
package/cpp/src/models/llada-moe.cpp +122 -0
package/cpp/src/models/llada.cpp +99 -0
package/cpp/src/models/llama-iswa.cpp +178 -0
package/cpp/src/models/llama.cpp +168 -0
package/cpp/src/models/maincoder.cpp +117 -0
package/cpp/src/models/mamba-base.cpp +285 -0
package/cpp/src/models/mamba.cpp +54 -0
package/cpp/src/models/mimo2-iswa.cpp +123 -0
package/cpp/src/models/minicpm3.cpp +200 -0
package/cpp/src/models/minimax-m2.cpp +124 -0
package/cpp/src/models/mistral3.cpp +160 -0
package/cpp/src/models/models.h +684 -0
package/cpp/src/models/modern-bert.cpp +109 -0
package/cpp/src/models/mpt.cpp +126 -0
package/cpp/src/models/nemotron-h.cpp +148 -0
package/cpp/src/models/nemotron.cpp +122 -0
package/cpp/src/models/neo-bert.cpp +104 -0
package/cpp/src/models/olmo.cpp +121 -0
package/cpp/src/models/olmo2.cpp +150 -0
package/cpp/src/models/olmoe.cpp +124 -0
package/cpp/src/models/openai-moe-iswa.cpp +127 -0
package/cpp/src/models/openelm.cpp +124 -0
package/cpp/src/models/orion.cpp +123 -0
package/cpp/src/models/paddleocr.cpp +122 -0
package/cpp/src/models/pangu-embedded.cpp +121 -0
package/cpp/src/models/phi2.cpp +121 -0
package/cpp/src/models/phi3.cpp +152 -0
package/cpp/src/models/plamo.cpp +110 -0
package/cpp/src/models/plamo2.cpp +318 -0
package/cpp/src/models/plamo3.cpp +128 -0
package/cpp/src/models/plm.cpp +169 -0
package/cpp/src/models/qwen.cpp +108 -0
package/cpp/src/models/qwen2.cpp +126 -0
package/cpp/src/models/qwen2moe.cpp +151 -0
package/cpp/src/models/qwen2vl.cpp +117 -0
package/cpp/src/models/qwen3.cpp +117 -0
package/cpp/src/models/qwen35.cpp +386 -0
package/cpp/src/models/qwen35moe.cpp +420 -0
package/cpp/src/models/qwen3moe.cpp +124 -0
package/cpp/src/models/qwen3next.cpp +525 -0
package/cpp/src/models/qwen3vl-moe.cpp +140 -0
package/cpp/src/models/qwen3vl.cpp +132 -0
package/cpp/src/models/refact.cpp +94 -0
package/cpp/src/models/rnd1.cpp +126 -0
package/cpp/src/models/rwkv6-base.cpp +164 -0
package/cpp/src/models/rwkv6.cpp +94 -0
package/cpp/src/models/rwkv6qwen2.cpp +86 -0
package/cpp/src/models/rwkv7-base.cpp +137 -0
package/cpp/src/models/rwkv7.cpp +90 -0
package/cpp/src/models/seed-oss.cpp +124 -0
package/cpp/src/models/smallthinker.cpp +126 -0
package/cpp/src/models/smollm3.cpp +128 -0
package/cpp/src/models/stablelm.cpp +146 -0
package/cpp/src/models/starcoder.cpp +100 -0
package/cpp/src/models/starcoder2.cpp +121 -0
package/cpp/src/models/step35-iswa.cpp +168 -0
package/cpp/src/models/t5-dec.cpp +166 -0
package/cpp/src/models/t5-enc.cpp +96 -0
package/cpp/src/models/wavtokenizer-dec.cpp +149 -0
package/cpp/src/models/xverse.cpp +108 -0
package/cpp/src/unicode-data.cpp +7034 -0
package/cpp/src/unicode-data.h +20 -0
package/cpp/src/unicode.cpp +1103 -0
package/cpp/src/unicode.h +111 -0
package/cpp/vendor/nlohmann/json.hpp +25526 -0
package/cpp/vendor/nlohmann/json_fwd.hpp +187 -0
package/cpp/vendor/stb/stb_image.h +7988 -0
package/ios/LocalLLM-Bridging-Header.h +2 -0
package/ios/LocalLLM.h +5 -0
package/ios/LocalLLM.mm +1267 -0
package/local-llm-rn.podspec +60 -0
package/package.json +35 -0
package/src/NativeLocalLLM.ts +73 -0
package/src/device.ts +50 -0
package/src/download-adapter.ts +17 -0
package/src/index.ts +21 -0
package/src/native-bridge.ts +142 -0
package/src/rn-downloader.ts +37 -0

package/cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl ADDED Viewed

@@ -0,0 +1,156 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define QK_MXFP4 32
+#define N_SIMDGROUP 4
+#define SIMDGROUP_WIDTH 64
+static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) {
+    ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00;
+    fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00;
+    fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00;
+    fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00;
+    bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0;
+    fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0;
+    fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0;
+    fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0;
+    fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0;
+    sign_a.lo = (fp4x8.s0 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s0 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s0 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s0 & 0x8000;
+    fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0;
+    fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0;
+    ushort2 fp16_packed_a_1, fp16_packed_b_1;
+    fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00;
+    fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00;
+    fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00;
+    fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00;
+    bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0;
+    fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0;
+    fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0;
+    fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0;
+    fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0;
+    sign_a.lo = (fp4x8.s1 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s1 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s1 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s1 & 0x8000;
+    fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1;
+    fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1;
+    return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1));
+}
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemv_moe_mxfp4_f32(
+    __global uint4 * src0_q,
+    __global uchar * src0_e,
+    __read_only image1d_buffer_t src1,
+    __global uint * src2,
+    __global float * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    int           ne11
+) {
+    uint i01  = get_global_id(0);
+    uint i20  = get_global_id(2);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+    uint i11 = i20 % ne11;
+    uint expert_id = src2[i20];
+    uint expert_offset = expert_id * ne00 * ne01 / 32;
+    __private float sum = 0.0f; // each thread calculate partial sum of one output
+    // loop along ne00 in block granularity, skip 4 blocks every iter
+    for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) {
+        // load one block of q
+        uint4 regQ = src0_q[expert_offset + ib00 * ne01 + i01];
+        uint offset = i11 * ne00 / 4 + ib00 * 8;
+        half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0));
+        float4 shared_y4;
+        shared_y4 = read_imagef(src1, (offset + 0));
+        float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+        shared_y4 = read_imagef(src1, (offset + 4));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1));
+        shared_y4 = read_imagef(src1, (offset + 1));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+        shared_y4 = read_imagef(src1, (offset + 5));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2));
+        shared_y4 = read_imagef(src1, (offset + 2));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+        shared_y4 = read_imagef(src1, (offset + 6));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3));
+        shared_y4 = read_imagef(src1, (offset + 3));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+        shared_y4 = read_imagef(src1, (offset + 7));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+        uchar regE = src0_e[ib00 * ne01 + i01 + expert_offset];
+        sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum;
+    if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum;
+    if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+    // 1 outputs per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01 + i20 * ne01] = sum;
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl ADDED Viewed

@@ -0,0 +1,268 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+// assume
+#define QK4_0 32
+#define N_SIMDGROUP 4
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C (0)
+        uint K,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+    __private uint4     regA;
+    __private half2     regS;
+    __private float8    regB;
+    __private float2 totalSum = (float2)(0.0f);
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+    // reduction in local memory, assumes #wave=4
+    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+}