whispercpp 1.3.5 → 1.3.7
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.
- checksums.yaml +4 -4
- data/.document +3 -0
- data/.rdoc_options +2 -0
- data/LICENSE +1 -1
- data/README.md +133 -3
- data/Rakefile +18 -3
- data/ext/dependencies.rb +10 -4
- data/ext/dependencies_for_windows.rb +17 -0
- data/ext/extconf.rb +20 -7
- data/ext/options.rb +54 -14
- data/ext/options_for_windows.rb +51 -0
- data/ext/ruby_whisper.c +56 -46
- data/ext/ruby_whisper.h +165 -2
- data/ext/ruby_whisper_context.c +297 -126
- data/ext/ruby_whisper_context_params.c +163 -0
- data/ext/ruby_whisper_log_queue.c +180 -0
- data/ext/ruby_whisper_log_settable.h +47 -0
- data/ext/ruby_whisper_model.c +0 -1
- data/ext/ruby_whisper_parakeet.c +49 -0
- data/ext/ruby_whisper_parakeet_context.c +304 -0
- data/ext/ruby_whisper_parakeet_context_params.c +117 -0
- data/ext/ruby_whisper_parakeet_model.c +84 -0
- data/ext/ruby_whisper_parakeet_params.c +548 -0
- data/ext/ruby_whisper_parakeet_segment.c +157 -0
- data/ext/ruby_whisper_parakeet_token.c +188 -0
- data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
- data/ext/ruby_whisper_params.c +256 -66
- data/ext/ruby_whisper_segment.c +6 -7
- data/ext/ruby_whisper_token.c +29 -9
- data/ext/ruby_whisper_transcribe.cpp +46 -16
- data/ext/ruby_whisper_vad_context.c +48 -1
- data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
- data/ext/ruby_whisper_vad_params.c +0 -1
- data/ext/ruby_whisper_vad_segment.c +0 -1
- data/ext/ruby_whisper_vad_segments.c +0 -1
- data/ext/sources/CMakeLists.txt +41 -3
- data/ext/sources/CMakePresets.json +95 -0
- data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
- data/ext/sources/cmake/parakeet.pc.in +10 -0
- data/ext/sources/cmake/whisper-config.cmake.in +5 -40
- data/ext/sources/cmake/whisper.pc.in +1 -1
- data/ext/sources/examples/CMakeLists.txt +4 -2
- data/ext/sources/examples/bench/bench.cpp +24 -19
- data/ext/sources/examples/cli/cli.cpp +51 -9
- data/ext/sources/examples/common-ggml.cpp +4 -0
- data/ext/sources/examples/common-whisper.cpp +139 -67
- data/ext/sources/examples/common-whisper.h +11 -0
- data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
- data/ext/sources/examples/miniaudio.h +4507 -2131
- data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
- data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
- data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
- data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
- data/ext/sources/examples/server/server.cpp +213 -163
- data/ext/sources/ggml/CMakeLists.txt +29 -15
- data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
- data/ext/sources/ggml/include/ggml-alloc.h +1 -0
- data/ext/sources/ggml/include/ggml-backend.h +73 -11
- data/ext/sources/ggml/include/ggml-cann.h +1 -1
- data/ext/sources/ggml/include/ggml-cpu.h +5 -0
- data/ext/sources/ggml/include/ggml-cuda.h +3 -0
- data/ext/sources/ggml/include/ggml-openvino.h +37 -0
- data/ext/sources/ggml/include/ggml-opt.h +1 -1
- data/ext/sources/ggml/include/ggml-rpc.h +8 -3
- data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
- data/ext/sources/ggml/include/ggml.h +155 -16
- data/ext/sources/ggml/include/gguf.h +10 -2
- data/ext/sources/ggml/src/CMakeLists.txt +25 -5
- data/ext/sources/ggml/src/ggml-alloc.c +9 -10
- data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
- data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
- data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
- data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +40 -86
- data/ext/sources/ggml/src/ggml-backend.cpp +114 -10
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +10 -2
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +1016 -442
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +111 -85
- data/ext/sources/ggml/src/ggml-cann/common.h +23 -14
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +255 -92
- data/ext/sources/ggml/src/ggml-common.h +22 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +68 -34
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +44 -19
- data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +101 -101
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2874 -613
- data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +5480 -840
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1361 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -11
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +186 -36
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +119 -19
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +112 -26
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +153 -16
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +17 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +976 -251
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +671 -266
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1277 -263
- data/ext/sources/ggml/src/ggml-cpu/ops.h +4 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +95 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2893 -679
- data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +226 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +114 -19
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +54 -53
- data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +18 -8
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +73 -28
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +69 -41
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +359 -29
- data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +94 -27
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +20 -9
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +333 -85
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +632 -190
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +162 -49
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +43 -18
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +44 -14
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +241 -23
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +312 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1454 -599
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +397 -183
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +161 -88
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +522 -431
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +139 -72
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +608 -88
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +47 -79
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +134 -27
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +7 -17
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +244 -137
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
- data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
- data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
- data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +96 -40
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -5
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +202 -135
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +86 -2
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +111 -17
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +30 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +84 -46
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1612 -753
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +51 -11
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +361 -261
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +294 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +753 -241
- data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +295 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +471 -296
- data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +159 -53
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +3 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +372 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +137 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +97 -14
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +163 -67
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +308 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +216 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +142 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -1348
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +547 -635
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3556 -1101
- data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +475 -269
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +94 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +222 -217
- data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +432 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +886 -117
- data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
- data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +40 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +28 -9
- data/ext/sources/ggml/src/ggml-impl.h +68 -1
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +409 -83
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +54 -5
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +254 -52
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +254 -23
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +756 -285
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +7 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +359 -133
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1867 -1123
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +71 -4
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +14127 -5314
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +97 -88
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +104 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1978 -67
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
- data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +195 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
- data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +178 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
- data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
- data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
- data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +985 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +380 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1132 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +956 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +149 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +47 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +40 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +317 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +257 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +86 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +880 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.h +143 -0
- data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
- data/ext/sources/ggml/src/ggml-quants.c +385 -119
- data/ext/sources/ggml/src/ggml-quants.h +6 -0
- data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
- data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
- data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
- data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +64 -91
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +4 -1
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
- data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +356 -11
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +184 -14
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +31 -1
- data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
- data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +77 -156
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1181 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +674 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +227 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
- data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
- data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +347 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1134 -236
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
- data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
- data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
- data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
- data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +72 -1
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +228 -53
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
- data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +12 -1
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3250 -940
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +533 -180
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +113 -68
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +412 -222
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +222 -83
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +22 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +51 -14
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +39 -63
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +13 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +27 -11
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -149
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3221 -97
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3493 -1997
- data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +115 -141
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +93 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -44
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +198 -230
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +234 -335
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +871 -42
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +149 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +36 -138
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +151 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +15 -40
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +39 -12
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +24 -15
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +253 -16
- data/ext/sources/ggml/src/ggml.c +268 -52
- data/ext/sources/ggml/src/gguf.cpp +377 -47
- data/ext/sources/include/parakeet.h +342 -0
- data/ext/sources/include/whisper.h +10 -0
- data/ext/sources/media/matmul.png +0 -0
- data/ext/sources/src/CMakeLists.txt +23 -0
- data/ext/sources/src/parakeet-arch.h +188 -0
- data/ext/sources/src/parakeet.cpp +3838 -0
- data/ext/sources/src/whisper.cpp +62 -40
- data/extsources.rb +26 -10
- data/lib/whisper/log_settable.rb +36 -0
- data/lib/whisper/model/uri.rb +13 -1
- data/lib/whisper/output.rb +74 -0
- data/sig/whisper.rbs +445 -55
- data/test/helper.rb +2 -0
- data/test/jfk_reader/jfk_reader.c +50 -7
- data/test/test_callback.rb +1 -0
- data/test/test_context_params.rb +82 -0
- data/test/test_package.rb +6 -5
- data/test/test_parakeet.rb +28 -0
- data/test/test_parakeet_callback.rb +107 -0
- data/test/test_parakeet_context.rb +116 -0
- data/test/test_parakeet_context_params.rb +24 -0
- data/test/test_parakeet_model.rb +21 -0
- data/test/test_parakeet_params.rb +78 -0
- data/test/test_parakeet_segment.rb +42 -0
- data/test/test_parakeet_token.rb +73 -0
- data/test/test_params.rb +2 -0
- data/test/test_token.rb +11 -0
- data/test/test_vad_context.rb +58 -8
- data/test/test_vad_segment.rb +1 -1
- data/test/test_whisper.rb +44 -6
- data/whispercpp.gemspec +2 -2
- metadata +426 -280
- data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
- data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
- data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
- data/ext/sources/bindings/javascript/package.json +0 -26
- data/ext/sources/bindings/javascript/whisper.js +0 -19
- data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
- data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
- data/ext/sources/examples/addon.node/addon.cpp +0 -557
- data/ext/sources/examples/addon.node/index.js +0 -59
- data/ext/sources/examples/addon.node/package.json +0 -16
- data/ext/sources/examples/addon.node/vad-example.js +0 -132
- data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
- data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
- data/ext/sources/examples/coi-serviceworker.js +0 -146
- data/ext/sources/examples/command/CMakeLists.txt +0 -10
- data/ext/sources/examples/command/command.cpp +0 -802
- data/ext/sources/examples/command/commands.txt +0 -9
- data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
- data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
- data/ext/sources/examples/generate-karaoke.sh +0 -57
- data/ext/sources/examples/helpers.js +0 -191
- data/ext/sources/examples/livestream.sh +0 -112
- data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
- data/ext/sources/examples/lsp/lsp.cpp +0 -471
- data/ext/sources/examples/lsp/whisper.vim +0 -362
- data/ext/sources/examples/python/test_whisper_processor.py +0 -7
- data/ext/sources/examples/python/whisper_processor.py +0 -54
- data/ext/sources/examples/server/bench.js +0 -29
- data/ext/sources/examples/server.py +0 -120
- data/ext/sources/examples/stream/CMakeLists.txt +0 -10
- data/ext/sources/examples/stream/stream.cpp +0 -437
- data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
- data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
- data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
- data/ext/sources/examples/sycl/build.sh +0 -22
- data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
- data/ext/sources/examples/sycl/run-whisper.sh +0 -17
- data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -47
- data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -494
- data/ext/sources/examples/talk-llama/llama-adapter.h +0 -88
- data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2559
- data/ext/sources/examples/talk-llama/llama-arch.h +0 -586
- data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -917
- data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
- data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -876
- data/ext/sources/examples/talk-llama/llama-chat.h +0 -70
- data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3645
- data/ext/sources/examples/talk-llama/llama-context.h +0 -360
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
- data/ext/sources/examples/talk-llama/llama-cparams.h +0 -42
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
- data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
- data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2282
- data/ext/sources/examples/talk-llama/llama-graph.h +0 -910
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -241
- data/ext/sources/examples/talk-llama/llama-hparams.h +0 -284
- data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
- data/ext/sources/examples/talk-llama/llama-impl.h +0 -63
- data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
- data/ext/sources/examples/talk-llama/llama-io.h +0 -35
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -328
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2100
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -390
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1167
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
- data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
- data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -735
- data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1247
- data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -176
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -285
- data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -37
- data/ext/sources/examples/talk-llama/llama-model.cpp +0 -8338
- data/ext/sources/examples/talk-llama/llama-model.h +0 -544
- data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1072
- data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +0 -3771
- data/ext/sources/examples/talk-llama/llama-sampling.h +0 -44
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3900
- data/ext/sources/examples/talk-llama/llama-vocab.h +0 -182
- data/ext/sources/examples/talk-llama/llama.cpp +0 -1140
- data/ext/sources/examples/talk-llama/llama.h +0 -1540
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -191
- data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
- data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -138
- data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/bert.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
- data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
- data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
- data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
- data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -259
- data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -134
- data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
- data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -113
- data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
- data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
- data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
- data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
- data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
- data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -196
- data/ext/sources/examples/talk-llama/models/granite.cpp +0 -211
- data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +0 -283
- data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -141
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -154
- data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -175
- data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
- data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/llama.cpp +0 -168
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -55
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -199
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/models.h +0 -569
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -116
- data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
- data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
- data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
- data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -316
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/plm.cpp +0 -168
- data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
- data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
- data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -873
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -149
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -141
- data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -162
- data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
- data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
- data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
- data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
- data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
- data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
- data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
- data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
- data/ext/sources/examples/talk-llama/speak +0 -40
- data/ext/sources/examples/talk-llama/speak.bat +0 -1
- data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
- data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
- data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
- data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
- data/ext/sources/examples/talk-llama/unicode.cpp +0 -1147
- data/ext/sources/examples/talk-llama/unicode.h +0 -111
- data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
- data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
- data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
- data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
- data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
- data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
- data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
- data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
- data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
- data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +0 -157
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -165
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
- data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
- data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
- data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -147
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +0 -907
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +0 -247
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
- data/ext/sources/tests/CMakeLists.txt +0 -112
- data/ext/sources/tests/earnings21/eval.mk +0 -58
- data/ext/sources/tests/earnings21/eval.py +0 -68
- data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
- data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
- data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
- data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
- data/ext/sources/tests/earnings21/requirements.txt +0 -6
- data/ext/sources/tests/en-0-ref.txt +0 -1
- data/ext/sources/tests/en-1-ref.txt +0 -1
- data/ext/sources/tests/en-2-ref.txt +0 -1
- data/ext/sources/tests/es-0-ref.txt +0 -1
- data/ext/sources/tests/librispeech/eval.mk +0 -39
- data/ext/sources/tests/librispeech/eval.py +0 -47
- data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
- data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
- data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
- data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
- data/ext/sources/tests/librispeech/requirements.txt +0 -6
- data/ext/sources/tests/run-tests.sh +0 -130
- data/ext/sources/tests/test-c.c +0 -3
- data/ext/sources/tests/test-vad-full.cpp +0 -56
- data/ext/sources/tests/test-vad.cpp +0 -83
- data/ext/sources/tests/test-whisper.js +0 -58
- data/lib/whisper/context.rb +0 -15
- data/lib/whisper/segment.rb +0 -58
|
@@ -1,735 +0,0 @@
|
|
|
1
|
-
#include "llama-mmap.h"
|
|
2
|
-
|
|
3
|
-
#include "llama-impl.h"
|
|
4
|
-
|
|
5
|
-
#include "ggml.h"
|
|
6
|
-
|
|
7
|
-
#include <cstring>
|
|
8
|
-
#include <climits>
|
|
9
|
-
#include <stdexcept>
|
|
10
|
-
#include <cerrno>
|
|
11
|
-
#include <algorithm>
|
|
12
|
-
|
|
13
|
-
#ifdef __has_include
|
|
14
|
-
#if __has_include(<unistd.h>)
|
|
15
|
-
#include <unistd.h>
|
|
16
|
-
#include <fcntl.h>
|
|
17
|
-
#include <sys/stat.h>
|
|
18
|
-
#if defined(_POSIX_MAPPED_FILES)
|
|
19
|
-
#include <sys/mman.h>
|
|
20
|
-
#endif
|
|
21
|
-
#if defined(_POSIX_MEMLOCK_RANGE)
|
|
22
|
-
#include <sys/resource.h>
|
|
23
|
-
#endif
|
|
24
|
-
#endif
|
|
25
|
-
#endif
|
|
26
|
-
|
|
27
|
-
#if defined(_WIN32)
|
|
28
|
-
#define WIN32_LEAN_AND_MEAN
|
|
29
|
-
#ifndef NOMINMAX
|
|
30
|
-
#define NOMINMAX
|
|
31
|
-
#endif
|
|
32
|
-
#include <windows.h>
|
|
33
|
-
#ifndef PATH_MAX
|
|
34
|
-
#define PATH_MAX MAX_PATH
|
|
35
|
-
#endif
|
|
36
|
-
#include <io.h>
|
|
37
|
-
#endif
|
|
38
|
-
|
|
39
|
-
#if defined(__APPLE__)
|
|
40
|
-
#include <TargetConditionals.h>
|
|
41
|
-
#endif
|
|
42
|
-
|
|
43
|
-
// TODO: consider moving to llama-impl.h if needed in more places
|
|
44
|
-
#if defined(_WIN32)
|
|
45
|
-
static std::string llama_format_win_err(DWORD err) {
|
|
46
|
-
LPSTR buf;
|
|
47
|
-
size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
|
|
48
|
-
NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
|
|
49
|
-
if (!size) {
|
|
50
|
-
return "FormatMessageA failed";
|
|
51
|
-
}
|
|
52
|
-
std::string ret(buf, size);
|
|
53
|
-
LocalFree(buf);
|
|
54
|
-
return ret;
|
|
55
|
-
}
|
|
56
|
-
#endif
|
|
57
|
-
|
|
58
|
-
// llama_file
|
|
59
|
-
|
|
60
|
-
struct llama_file::impl {
|
|
61
|
-
#if defined(_WIN32)
|
|
62
|
-
HANDLE fp_win32;
|
|
63
|
-
std::string GetErrorMessageWin32(DWORD error_code) const {
|
|
64
|
-
std::string ret;
|
|
65
|
-
LPSTR lpMsgBuf = NULL;
|
|
66
|
-
DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
|
|
67
|
-
NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
|
|
68
|
-
if (!bufLen) {
|
|
69
|
-
ret = format("Win32 error code: %lx", error_code);
|
|
70
|
-
} else {
|
|
71
|
-
ret = lpMsgBuf;
|
|
72
|
-
LocalFree(lpMsgBuf);
|
|
73
|
-
}
|
|
74
|
-
|
|
75
|
-
return ret;
|
|
76
|
-
}
|
|
77
|
-
|
|
78
|
-
impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
|
|
79
|
-
fp = ggml_fopen(fname, mode);
|
|
80
|
-
if (fp == NULL) {
|
|
81
|
-
throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
|
|
82
|
-
}
|
|
83
|
-
fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
|
|
84
|
-
seek(0, SEEK_END);
|
|
85
|
-
size = tell();
|
|
86
|
-
seek(0, SEEK_SET);
|
|
87
|
-
}
|
|
88
|
-
|
|
89
|
-
size_t tell() const {
|
|
90
|
-
LARGE_INTEGER li;
|
|
91
|
-
li.QuadPart = 0;
|
|
92
|
-
BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
|
|
93
|
-
if (!ret) {
|
|
94
|
-
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
|
|
95
|
-
}
|
|
96
|
-
|
|
97
|
-
return li.QuadPart;
|
|
98
|
-
}
|
|
99
|
-
|
|
100
|
-
void seek(size_t offset, int whence) const {
|
|
101
|
-
static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
|
|
102
|
-
static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
|
|
103
|
-
static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
|
|
104
|
-
|
|
105
|
-
LARGE_INTEGER li;
|
|
106
|
-
li.QuadPart = offset;
|
|
107
|
-
BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
|
|
108
|
-
if (!ret) {
|
|
109
|
-
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
|
|
110
|
-
}
|
|
111
|
-
}
|
|
112
|
-
|
|
113
|
-
void read_raw(void * ptr, size_t len) {
|
|
114
|
-
size_t bytes_read = 0;
|
|
115
|
-
while (bytes_read < len) {
|
|
116
|
-
size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
|
|
117
|
-
DWORD chunk_read = 0;
|
|
118
|
-
BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
|
|
119
|
-
if (!result) {
|
|
120
|
-
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
|
|
121
|
-
}
|
|
122
|
-
if (chunk_read < chunk_size || chunk_read == 0) {
|
|
123
|
-
throw std::runtime_error("unexpectedly reached end of file");
|
|
124
|
-
}
|
|
125
|
-
|
|
126
|
-
bytes_read += chunk_read;
|
|
127
|
-
}
|
|
128
|
-
}
|
|
129
|
-
|
|
130
|
-
uint32_t read_u32() {
|
|
131
|
-
uint32_t val;
|
|
132
|
-
read_raw(&val, sizeof(val));
|
|
133
|
-
return val;
|
|
134
|
-
}
|
|
135
|
-
|
|
136
|
-
void write_raw(const void * ptr, size_t len) const {
|
|
137
|
-
size_t bytes_written = 0;
|
|
138
|
-
while (bytes_written < len) {
|
|
139
|
-
size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
|
|
140
|
-
DWORD chunk_written = 0;
|
|
141
|
-
BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
|
|
142
|
-
if (!result) {
|
|
143
|
-
throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
|
|
144
|
-
}
|
|
145
|
-
if (chunk_written < chunk_size || chunk_written == 0) {
|
|
146
|
-
throw std::runtime_error("unexpectedly failed to write bytes");
|
|
147
|
-
}
|
|
148
|
-
|
|
149
|
-
bytes_written += chunk_written;
|
|
150
|
-
}
|
|
151
|
-
}
|
|
152
|
-
|
|
153
|
-
void write_u32(uint32_t val) const {
|
|
154
|
-
write_raw(&val, sizeof(val));
|
|
155
|
-
}
|
|
156
|
-
|
|
157
|
-
bool has_direct_io() const {
|
|
158
|
-
return true;
|
|
159
|
-
}
|
|
160
|
-
|
|
161
|
-
~impl() {
|
|
162
|
-
if (fp) {
|
|
163
|
-
std::fclose(fp);
|
|
164
|
-
}
|
|
165
|
-
}
|
|
166
|
-
#else
|
|
167
|
-
impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) : fname(fname) {
|
|
168
|
-
#ifdef __linux__
|
|
169
|
-
// Try unbuffered I/O for read only
|
|
170
|
-
if (use_direct_io && std::strcmp(mode, "rb") == 0) {
|
|
171
|
-
if (init_fd()) {
|
|
172
|
-
return;
|
|
173
|
-
}
|
|
174
|
-
LLAMA_LOG_WARN("Failed to open file '%s' with error: %s. Falling back to buffered I/O",
|
|
175
|
-
fname, strerror(errno));
|
|
176
|
-
}
|
|
177
|
-
#endif
|
|
178
|
-
init_fp(mode);
|
|
179
|
-
}
|
|
180
|
-
|
|
181
|
-
#ifdef __linux__
|
|
182
|
-
bool init_fd() {
|
|
183
|
-
fd = open(fname.c_str(), O_RDONLY | O_DIRECT);
|
|
184
|
-
|
|
185
|
-
if (fd != -1) {
|
|
186
|
-
struct stat file_stats{};
|
|
187
|
-
fstat(fd, &file_stats);
|
|
188
|
-
|
|
189
|
-
size = file_stats.st_size;
|
|
190
|
-
alignment = file_stats.st_blksize;
|
|
191
|
-
|
|
192
|
-
off_t ret = lseek(fd, 0, SEEK_SET);
|
|
193
|
-
if (ret == -1) {
|
|
194
|
-
throw std::runtime_error(format("seek error: %s", strerror(errno)));
|
|
195
|
-
}
|
|
196
|
-
return true;
|
|
197
|
-
}
|
|
198
|
-
return false;
|
|
199
|
-
}
|
|
200
|
-
#endif
|
|
201
|
-
|
|
202
|
-
void init_fp(const char * mode) {
|
|
203
|
-
fp = ggml_fopen(fname.c_str(), mode);
|
|
204
|
-
if (fp == NULL) {
|
|
205
|
-
throw std::runtime_error(format("failed to open %s: %s", fname.c_str(), strerror(errno)));
|
|
206
|
-
}
|
|
207
|
-
seek(0, SEEK_END);
|
|
208
|
-
size = tell();
|
|
209
|
-
seek(0, SEEK_SET);
|
|
210
|
-
}
|
|
211
|
-
|
|
212
|
-
size_t tell() const {
|
|
213
|
-
if (fd == -1) {
|
|
214
|
-
long ret = std::ftell(fp);
|
|
215
|
-
if (ret == -1) {
|
|
216
|
-
throw std::runtime_error(format("ftell error: %s", strerror(errno)));
|
|
217
|
-
}
|
|
218
|
-
|
|
219
|
-
return (size_t) ret;
|
|
220
|
-
}
|
|
221
|
-
|
|
222
|
-
off_t pos = lseek(fd, 0, SEEK_CUR);
|
|
223
|
-
if (pos == -1) {
|
|
224
|
-
throw std::runtime_error(format("lseek error: %s", strerror(errno)));
|
|
225
|
-
}
|
|
226
|
-
return (size_t) pos;
|
|
227
|
-
}
|
|
228
|
-
|
|
229
|
-
void seek(size_t offset, int whence) const {
|
|
230
|
-
off_t ret = 0;
|
|
231
|
-
if (fd == -1) {
|
|
232
|
-
ret = std::fseek(fp, (long) offset, whence);
|
|
233
|
-
} else {
|
|
234
|
-
ret = lseek(fd, offset, whence);
|
|
235
|
-
}
|
|
236
|
-
if (ret == -1) {
|
|
237
|
-
throw std::runtime_error(format("seek error: %s", strerror(errno)));
|
|
238
|
-
}
|
|
239
|
-
}
|
|
240
|
-
|
|
241
|
-
void read_raw_unsafe(void * ptr, size_t len) {
|
|
242
|
-
if (len == 0) {
|
|
243
|
-
return;
|
|
244
|
-
}
|
|
245
|
-
errno = 0;
|
|
246
|
-
if (fd == -1) {
|
|
247
|
-
std::size_t ret = std::fread(ptr, len, 1, fp);
|
|
248
|
-
if (ferror(fp)) {
|
|
249
|
-
throw std::runtime_error(format("read error: %s", strerror(errno)));
|
|
250
|
-
}
|
|
251
|
-
if (ret != 1) {
|
|
252
|
-
throw std::runtime_error("unexpectedly reached end of file");
|
|
253
|
-
}
|
|
254
|
-
} else {
|
|
255
|
-
size_t bytes_read = 0;
|
|
256
|
-
while (bytes_read < len) {
|
|
257
|
-
const size_t to_read = len - bytes_read;
|
|
258
|
-
ssize_t ret = ::read(fd, reinterpret_cast<char *>(ptr) + bytes_read, to_read);
|
|
259
|
-
|
|
260
|
-
if (ret == -1) {
|
|
261
|
-
if (errno == EINTR) {
|
|
262
|
-
continue; // Interrupted by signal, retry
|
|
263
|
-
}
|
|
264
|
-
// Fallback to std::fread in case the DMA controller cannot access the buffer
|
|
265
|
-
if (errno == EFAULT) {
|
|
266
|
-
auto curr_off = tell();
|
|
267
|
-
close(fd);
|
|
268
|
-
fd = -1;
|
|
269
|
-
alignment = 1;
|
|
270
|
-
init_fp("rb");
|
|
271
|
-
seek(curr_off, SEEK_SET);
|
|
272
|
-
read_raw_unsafe(ptr, len);
|
|
273
|
-
return;
|
|
274
|
-
}
|
|
275
|
-
throw std::runtime_error(format("read error: %s", strerror(errno)));
|
|
276
|
-
}
|
|
277
|
-
if (ret == 0) {
|
|
278
|
-
// EOF: allow if this read was only pulling alignment padding past file end
|
|
279
|
-
off_t pos = lseek(fd, 0, SEEK_CUR);
|
|
280
|
-
if (pos != -1 && (size_t) pos == size) {
|
|
281
|
-
std::memset(reinterpret_cast<char *>(ptr) + bytes_read, 0, len - bytes_read);
|
|
282
|
-
return;
|
|
283
|
-
}
|
|
284
|
-
throw std::runtime_error("unexpectedly reached end of file");
|
|
285
|
-
}
|
|
286
|
-
|
|
287
|
-
bytes_read += (size_t) ret;
|
|
288
|
-
}
|
|
289
|
-
}
|
|
290
|
-
}
|
|
291
|
-
|
|
292
|
-
void read_aligned_chunk(void * dest, size_t size) {
|
|
293
|
-
size_t offset = tell();
|
|
294
|
-
off_t aligned_offset = offset & ~(alignment - 1);
|
|
295
|
-
off_t offset_from_alignment = offset - aligned_offset;
|
|
296
|
-
size_t bytes_to_read = (offset_from_alignment + size + alignment - 1) & ~(alignment - 1);
|
|
297
|
-
|
|
298
|
-
void * raw_buffer = nullptr;
|
|
299
|
-
int ret = posix_memalign(&raw_buffer, alignment, bytes_to_read);
|
|
300
|
-
if (ret != 0) {
|
|
301
|
-
throw std::runtime_error(format("posix_memalign failed with error %d", ret));
|
|
302
|
-
}
|
|
303
|
-
|
|
304
|
-
struct aligned_buffer_deleter {
|
|
305
|
-
void operator()(void * p) const { free(p); }
|
|
306
|
-
};
|
|
307
|
-
std::unique_ptr<void, aligned_buffer_deleter> buffer(raw_buffer);
|
|
308
|
-
|
|
309
|
-
seek(aligned_offset, SEEK_SET);
|
|
310
|
-
read_raw_unsafe(buffer.get(), bytes_to_read);
|
|
311
|
-
|
|
312
|
-
uintptr_t actual_data = reinterpret_cast<uintptr_t>(buffer.get()) + offset_from_alignment;
|
|
313
|
-
memcpy(dest, reinterpret_cast<void *>(actual_data), size);
|
|
314
|
-
}
|
|
315
|
-
|
|
316
|
-
void read_raw(void * ptr, size_t len) {
|
|
317
|
-
if (has_direct_io()) {
|
|
318
|
-
read_aligned_chunk(ptr, len);
|
|
319
|
-
} else {
|
|
320
|
-
read_raw_unsafe(ptr, len);
|
|
321
|
-
}
|
|
322
|
-
}
|
|
323
|
-
|
|
324
|
-
uint32_t read_u32() {
|
|
325
|
-
uint32_t ret;
|
|
326
|
-
read_raw(&ret, sizeof(ret));
|
|
327
|
-
return ret;
|
|
328
|
-
}
|
|
329
|
-
|
|
330
|
-
void write_raw(const void * ptr, size_t len) const {
|
|
331
|
-
if (len == 0) {
|
|
332
|
-
return;
|
|
333
|
-
}
|
|
334
|
-
errno = 0;
|
|
335
|
-
size_t ret = std::fwrite(ptr, len, 1, fp);
|
|
336
|
-
if (ret != 1) {
|
|
337
|
-
throw std::runtime_error(format("write error: %s", strerror(errno)));
|
|
338
|
-
}
|
|
339
|
-
}
|
|
340
|
-
|
|
341
|
-
void write_u32(uint32_t val) const {
|
|
342
|
-
write_raw(&val, sizeof(val));
|
|
343
|
-
}
|
|
344
|
-
|
|
345
|
-
bool has_direct_io() const {
|
|
346
|
-
return fd != -1 && alignment > 1;
|
|
347
|
-
}
|
|
348
|
-
|
|
349
|
-
~impl() {
|
|
350
|
-
if (fd != -1) {
|
|
351
|
-
close(fd);
|
|
352
|
-
} else {
|
|
353
|
-
std::fclose(fp);
|
|
354
|
-
}
|
|
355
|
-
}
|
|
356
|
-
int fd = -1;
|
|
357
|
-
std::string fname;
|
|
358
|
-
#endif
|
|
359
|
-
|
|
360
|
-
size_t read_alignment() const {
|
|
361
|
-
return alignment;
|
|
362
|
-
}
|
|
363
|
-
|
|
364
|
-
size_t alignment = 1;
|
|
365
|
-
|
|
366
|
-
FILE * fp{};
|
|
367
|
-
size_t size{};
|
|
368
|
-
};
|
|
369
|
-
|
|
370
|
-
llama_file::llama_file(const char * fname, const char * mode, const bool use_direct_io) :
|
|
371
|
-
pimpl(std::make_unique<impl>(fname, mode, use_direct_io)) {}
|
|
372
|
-
llama_file::~llama_file() = default;
|
|
373
|
-
|
|
374
|
-
size_t llama_file::tell() const { return pimpl->tell(); }
|
|
375
|
-
size_t llama_file::size() const { return pimpl->size; }
|
|
376
|
-
|
|
377
|
-
size_t llama_file::read_alignment() const { return pimpl->read_alignment(); }
|
|
378
|
-
bool llama_file::has_direct_io() const { return pimpl->has_direct_io(); }
|
|
379
|
-
|
|
380
|
-
int llama_file::file_id() const {
|
|
381
|
-
#ifdef _WIN32
|
|
382
|
-
return _fileno(pimpl->fp);
|
|
383
|
-
#else
|
|
384
|
-
#if defined(fileno)
|
|
385
|
-
return fileno(pimpl->fp);
|
|
386
|
-
#else
|
|
387
|
-
return ::fileno(pimpl->fp);
|
|
388
|
-
#endif
|
|
389
|
-
#endif
|
|
390
|
-
}
|
|
391
|
-
|
|
392
|
-
void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
|
|
393
|
-
void llama_file::read_raw(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
|
|
394
|
-
#ifdef _WIN32
|
|
395
|
-
void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
|
|
396
|
-
#else
|
|
397
|
-
void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsafe(ptr, len); }
|
|
398
|
-
#endif
|
|
399
|
-
|
|
400
|
-
uint32_t llama_file::read_u32() { return pimpl->read_u32(); }
|
|
401
|
-
|
|
402
|
-
void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
|
|
403
|
-
void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
|
|
404
|
-
|
|
405
|
-
// llama_mmap
|
|
406
|
-
|
|
407
|
-
struct llama_mmap::impl {
|
|
408
|
-
#ifdef _POSIX_MAPPED_FILES
|
|
409
|
-
std::vector<std::pair<size_t, size_t>> mapped_fragments;
|
|
410
|
-
|
|
411
|
-
impl(struct llama_file * file, size_t prefetch, bool numa) {
|
|
412
|
-
size = file->size();
|
|
413
|
-
int fd = file->file_id();
|
|
414
|
-
int flags = MAP_SHARED;
|
|
415
|
-
if (numa) { prefetch = 0; }
|
|
416
|
-
#ifdef __linux__
|
|
417
|
-
if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
|
|
418
|
-
LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
|
|
419
|
-
strerror(errno));
|
|
420
|
-
}
|
|
421
|
-
if (prefetch) { flags |= MAP_POPULATE; }
|
|
422
|
-
#endif
|
|
423
|
-
addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
|
|
424
|
-
if (addr == MAP_FAILED) {
|
|
425
|
-
throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
|
|
426
|
-
}
|
|
427
|
-
|
|
428
|
-
if (prefetch > 0) {
|
|
429
|
-
if (posix_madvise(addr, std::min(file->size(), prefetch), POSIX_MADV_WILLNEED)) {
|
|
430
|
-
LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
|
|
431
|
-
strerror(errno));
|
|
432
|
-
}
|
|
433
|
-
}
|
|
434
|
-
if (numa) {
|
|
435
|
-
if (posix_madvise(addr, file->size(), POSIX_MADV_RANDOM)) {
|
|
436
|
-
LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
|
|
437
|
-
strerror(errno));
|
|
438
|
-
}
|
|
439
|
-
}
|
|
440
|
-
|
|
441
|
-
mapped_fragments.emplace_back(0, file->size());
|
|
442
|
-
}
|
|
443
|
-
|
|
444
|
-
static void align_range(size_t * first, size_t * last, size_t page_size) {
|
|
445
|
-
size_t offset_in_page = *first & (page_size - 1);
|
|
446
|
-
size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
|
|
447
|
-
*first += offset_to_page;
|
|
448
|
-
|
|
449
|
-
*last = *last & ~(page_size - 1);
|
|
450
|
-
|
|
451
|
-
if (*last <= *first) {
|
|
452
|
-
*last = *first;
|
|
453
|
-
}
|
|
454
|
-
}
|
|
455
|
-
|
|
456
|
-
void unmap_fragment(size_t first, size_t last) {
|
|
457
|
-
int page_size = sysconf(_SC_PAGESIZE);
|
|
458
|
-
align_range(&first, &last, page_size);
|
|
459
|
-
size_t len = last - first;
|
|
460
|
-
|
|
461
|
-
if (len == 0) {
|
|
462
|
-
return;
|
|
463
|
-
}
|
|
464
|
-
|
|
465
|
-
GGML_ASSERT(first % page_size == 0);
|
|
466
|
-
GGML_ASSERT(last % page_size == 0);
|
|
467
|
-
GGML_ASSERT(last > first);
|
|
468
|
-
|
|
469
|
-
void * next_page_start = (uint8_t *) addr + first;
|
|
470
|
-
|
|
471
|
-
if (munmap(next_page_start, len)) {
|
|
472
|
-
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
|
|
473
|
-
}
|
|
474
|
-
|
|
475
|
-
std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
|
|
476
|
-
for (const auto & frag : mapped_fragments) {
|
|
477
|
-
if (frag.first < first && frag.second > last) {
|
|
478
|
-
new_mapped_fragments.emplace_back(frag.first, first);
|
|
479
|
-
new_mapped_fragments.emplace_back(last, frag.second);
|
|
480
|
-
} else if (frag.first < first && frag.second > first) {
|
|
481
|
-
new_mapped_fragments.emplace_back(frag.first, first);
|
|
482
|
-
} else if (frag.first < last && frag.second > last) {
|
|
483
|
-
new_mapped_fragments.emplace_back(last, frag.second);
|
|
484
|
-
} else if (frag.first >= first && frag.second <= last) {
|
|
485
|
-
} else {
|
|
486
|
-
new_mapped_fragments.push_back(frag);
|
|
487
|
-
}
|
|
488
|
-
}
|
|
489
|
-
mapped_fragments = std::move(new_mapped_fragments);
|
|
490
|
-
}
|
|
491
|
-
|
|
492
|
-
~impl() {
|
|
493
|
-
for (const auto & frag : mapped_fragments) {
|
|
494
|
-
if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
|
|
495
|
-
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
|
|
496
|
-
}
|
|
497
|
-
}
|
|
498
|
-
}
|
|
499
|
-
#elif defined(_WIN32)
|
|
500
|
-
impl(struct llama_file * file, size_t prefetch, bool numa) {
|
|
501
|
-
GGML_UNUSED(numa);
|
|
502
|
-
|
|
503
|
-
size = file->size();
|
|
504
|
-
|
|
505
|
-
HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
|
|
506
|
-
|
|
507
|
-
HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
|
|
508
|
-
|
|
509
|
-
if (hMapping == NULL) {
|
|
510
|
-
DWORD error = GetLastError();
|
|
511
|
-
throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
|
|
512
|
-
}
|
|
513
|
-
|
|
514
|
-
addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
|
|
515
|
-
DWORD error = GetLastError();
|
|
516
|
-
CloseHandle(hMapping);
|
|
517
|
-
|
|
518
|
-
if (addr == NULL) {
|
|
519
|
-
throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
|
|
520
|
-
}
|
|
521
|
-
|
|
522
|
-
if (prefetch > 0) {
|
|
523
|
-
#if _WIN32_WINNT >= 0x602
|
|
524
|
-
BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
|
|
525
|
-
HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
|
|
526
|
-
|
|
527
|
-
pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
|
|
528
|
-
|
|
529
|
-
if (pPrefetchVirtualMemory) {
|
|
530
|
-
WIN32_MEMORY_RANGE_ENTRY range;
|
|
531
|
-
range.VirtualAddress = addr;
|
|
532
|
-
range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
|
|
533
|
-
if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
|
|
534
|
-
LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
|
|
535
|
-
llama_format_win_err(GetLastError()).c_str());
|
|
536
|
-
}
|
|
537
|
-
}
|
|
538
|
-
#else
|
|
539
|
-
LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
|
|
540
|
-
#endif
|
|
541
|
-
}
|
|
542
|
-
}
|
|
543
|
-
|
|
544
|
-
void unmap_fragment(size_t first, size_t last) {
|
|
545
|
-
GGML_UNUSED(first);
|
|
546
|
-
GGML_UNUSED(last);
|
|
547
|
-
}
|
|
548
|
-
|
|
549
|
-
~impl() {
|
|
550
|
-
if (!UnmapViewOfFile(addr)) {
|
|
551
|
-
LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
|
|
552
|
-
llama_format_win_err(GetLastError()).c_str());
|
|
553
|
-
}
|
|
554
|
-
}
|
|
555
|
-
#else
|
|
556
|
-
impl(struct llama_file * file, size_t prefetch, bool numa) {
|
|
557
|
-
GGML_UNUSED(file);
|
|
558
|
-
GGML_UNUSED(prefetch);
|
|
559
|
-
GGML_UNUSED(numa);
|
|
560
|
-
|
|
561
|
-
throw std::runtime_error("mmap not supported");
|
|
562
|
-
}
|
|
563
|
-
|
|
564
|
-
void unmap_fragment(size_t first, size_t last) {
|
|
565
|
-
GGML_UNUSED(first);
|
|
566
|
-
GGML_UNUSED(last);
|
|
567
|
-
|
|
568
|
-
throw std::runtime_error("mmap not supported");
|
|
569
|
-
}
|
|
570
|
-
#endif
|
|
571
|
-
|
|
572
|
-
void * addr;
|
|
573
|
-
size_t size;
|
|
574
|
-
};
|
|
575
|
-
|
|
576
|
-
llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique<impl>(file, prefetch, numa)) {}
|
|
577
|
-
llama_mmap::~llama_mmap() = default;
|
|
578
|
-
|
|
579
|
-
size_t llama_mmap::size() const { return pimpl->size; }
|
|
580
|
-
void * llama_mmap::addr() const { return pimpl->addr; }
|
|
581
|
-
|
|
582
|
-
void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
|
|
583
|
-
|
|
584
|
-
#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
|
|
585
|
-
const bool llama_mmap::SUPPORTED = true;
|
|
586
|
-
#else
|
|
587
|
-
const bool llama_mmap::SUPPORTED = false;
|
|
588
|
-
#endif
|
|
589
|
-
|
|
590
|
-
// llama_mlock
|
|
591
|
-
|
|
592
|
-
struct llama_mlock::impl {
|
|
593
|
-
#ifdef _POSIX_MEMLOCK_RANGE
|
|
594
|
-
static size_t lock_granularity() {
|
|
595
|
-
return (size_t) sysconf(_SC_PAGESIZE);
|
|
596
|
-
}
|
|
597
|
-
|
|
598
|
-
bool raw_lock(const void * addr, size_t size) const {
|
|
599
|
-
if (!mlock(addr, size)) {
|
|
600
|
-
return true;
|
|
601
|
-
}
|
|
602
|
-
|
|
603
|
-
#ifdef __APPLE__
|
|
604
|
-
#define MLOCK_SUGGESTION \
|
|
605
|
-
"Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
|
|
606
|
-
"decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
|
|
607
|
-
#else
|
|
608
|
-
#define MLOCK_SUGGESTION \
|
|
609
|
-
"Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
|
|
610
|
-
#endif
|
|
611
|
-
|
|
612
|
-
char* errmsg = std::strerror(errno);
|
|
613
|
-
bool suggest = (errno == ENOMEM);
|
|
614
|
-
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
|
|
615
|
-
// visionOS/tvOS dont't support RLIMIT_MEMLOCK
|
|
616
|
-
// Skip resource limit checks on visionOS/tvOS
|
|
617
|
-
suggest = false;
|
|
618
|
-
#else
|
|
619
|
-
struct rlimit lock_limit;
|
|
620
|
-
if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
|
|
621
|
-
suggest = false;
|
|
622
|
-
}
|
|
623
|
-
if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
|
|
624
|
-
suggest = false;
|
|
625
|
-
}
|
|
626
|
-
#endif
|
|
627
|
-
|
|
628
|
-
LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
|
|
629
|
-
size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
|
|
630
|
-
return false;
|
|
631
|
-
}
|
|
632
|
-
|
|
633
|
-
static void raw_unlock(void * addr, size_t size) {
|
|
634
|
-
if (munlock(addr, size)) {
|
|
635
|
-
LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno));
|
|
636
|
-
}
|
|
637
|
-
}
|
|
638
|
-
#elif defined(_WIN32)
|
|
639
|
-
static size_t lock_granularity() {
|
|
640
|
-
SYSTEM_INFO si;
|
|
641
|
-
GetSystemInfo(&si);
|
|
642
|
-
return (size_t) si.dwPageSize;
|
|
643
|
-
}
|
|
644
|
-
|
|
645
|
-
bool raw_lock(void * ptr, size_t len) const {
|
|
646
|
-
for (int tries = 1; ; tries++) {
|
|
647
|
-
if (VirtualLock(ptr, len)) {
|
|
648
|
-
return true;
|
|
649
|
-
}
|
|
650
|
-
if (tries == 2) {
|
|
651
|
-
LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
|
|
652
|
-
len, size, llama_format_win_err(GetLastError()).c_str());
|
|
653
|
-
return false;
|
|
654
|
-
}
|
|
655
|
-
|
|
656
|
-
SIZE_T min_ws_size, max_ws_size;
|
|
657
|
-
if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
|
|
658
|
-
LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n",
|
|
659
|
-
llama_format_win_err(GetLastError()).c_str());
|
|
660
|
-
return false;
|
|
661
|
-
}
|
|
662
|
-
size_t increment = len + 1048576;
|
|
663
|
-
min_ws_size += increment;
|
|
664
|
-
max_ws_size += increment;
|
|
665
|
-
if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
|
|
666
|
-
LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n",
|
|
667
|
-
llama_format_win_err(GetLastError()).c_str());
|
|
668
|
-
return false;
|
|
669
|
-
}
|
|
670
|
-
}
|
|
671
|
-
}
|
|
672
|
-
|
|
673
|
-
static void raw_unlock(void * ptr, size_t len) {
|
|
674
|
-
if (!VirtualUnlock(ptr, len)) {
|
|
675
|
-
LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n",
|
|
676
|
-
llama_format_win_err(GetLastError()).c_str());
|
|
677
|
-
}
|
|
678
|
-
}
|
|
679
|
-
#else
|
|
680
|
-
static size_t lock_granularity() {
|
|
681
|
-
return (size_t) 65536;
|
|
682
|
-
}
|
|
683
|
-
|
|
684
|
-
bool raw_lock(const void * addr, size_t len) const {
|
|
685
|
-
LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
|
|
686
|
-
return false;
|
|
687
|
-
}
|
|
688
|
-
|
|
689
|
-
static void raw_unlock(const void * addr, size_t len) {}
|
|
690
|
-
#endif
|
|
691
|
-
|
|
692
|
-
impl() : addr(NULL), size(0), failed_already(false) {}
|
|
693
|
-
|
|
694
|
-
void init(void * ptr) {
|
|
695
|
-
GGML_ASSERT(addr == NULL && size == 0);
|
|
696
|
-
addr = ptr;
|
|
697
|
-
}
|
|
698
|
-
|
|
699
|
-
void grow_to(size_t target_size) {
|
|
700
|
-
GGML_ASSERT(addr);
|
|
701
|
-
if (failed_already) {
|
|
702
|
-
return;
|
|
703
|
-
}
|
|
704
|
-
size_t granularity = lock_granularity();
|
|
705
|
-
target_size = (target_size + granularity - 1) & ~(granularity - 1);
|
|
706
|
-
if (target_size > size) {
|
|
707
|
-
if (raw_lock((uint8_t *) addr + size, target_size - size)) {
|
|
708
|
-
size = target_size;
|
|
709
|
-
} else {
|
|
710
|
-
failed_already = true;
|
|
711
|
-
}
|
|
712
|
-
}
|
|
713
|
-
}
|
|
714
|
-
|
|
715
|
-
void * addr;
|
|
716
|
-
size_t size;
|
|
717
|
-
|
|
718
|
-
bool failed_already;
|
|
719
|
-
};
|
|
720
|
-
|
|
721
|
-
llama_mlock::llama_mlock() : pimpl(std::make_unique<impl>()) {}
|
|
722
|
-
llama_mlock::~llama_mlock() = default;
|
|
723
|
-
|
|
724
|
-
void llama_mlock::init(void * ptr) { pimpl->init(ptr); }
|
|
725
|
-
void llama_mlock::grow_to(size_t target_size) { pimpl->grow_to(target_size); }
|
|
726
|
-
|
|
727
|
-
#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
|
|
728
|
-
const bool llama_mlock::SUPPORTED = true;
|
|
729
|
-
#else
|
|
730
|
-
const bool llama_mlock::SUPPORTED = false;
|
|
731
|
-
#endif
|
|
732
|
-
|
|
733
|
-
size_t llama_path_max() {
|
|
734
|
-
return PATH_MAX;
|
|
735
|
-
}
|