whispercpp 1.3.6 → 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/README.md +38 -5
- data/Rakefile +18 -3
- data/ext/dependencies.rb +10 -4
- data/ext/dependencies_for_windows.rb +17 -0
- data/ext/extconf.rb +20 -8
- data/ext/options.rb +54 -14
- data/ext/options_for_windows.rb +51 -0
- data/ext/ruby_whisper.c +36 -42
- data/ext/ruby_whisper.h +135 -0
- data/ext/ruby_whisper_context.c +107 -28
- data/ext/ruby_whisper_log_queue.c +180 -0
- data/ext/ruby_whisper_log_settable.h +47 -0
- 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 -65
- data/ext/ruby_whisper_segment.c +6 -6
- data/ext/ruby_whisper_transcribe.cpp +42 -15
- 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.pc.in +1 -1
- data/ext/sources/examples/CMakeLists.txt +4 -2
- data/ext/sources/examples/bench/bench.cpp +1 -1
- data/ext/sources/examples/cli/cli.cpp +43 -9
- data/ext/sources/examples/common-ggml.cpp +2 -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/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 +199 -163
- data/ext/sources/ggml/CMakeLists.txt +21 -13
- 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 +72 -10
- data/ext/sources/ggml/include/ggml-cuda.h +3 -0
- data/ext/sources/ggml/include/ggml-rpc.h +3 -3
- data/ext/sources/ggml/include/ggml.h +101 -9
- data/ext/sources/ggml/include/gguf.h +10 -2
- data/ext/sources/ggml/src/CMakeLists.txt +22 -5
- data/ext/sources/ggml/src/ggml-alloc.c +5 -1
- 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 +12 -0
- data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
- data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
- data/ext/sources/ggml/src/ggml-common.h +11 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +58 -29
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +16 -16
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +116 -7
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
- 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 +4279 -1292
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +177 -27
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +5 -0
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +95 -5
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +88 -70
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +372 -73
- data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +55 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +90 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +3 -16
- 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/vec.cpp +37 -53
- data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
- 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 +62 -26
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +44 -18
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +242 -28
- 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 +53 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +14 -6
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +278 -44
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +331 -130
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +126 -27
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +152 -49
- 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 +84 -35
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1069 -609
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +485 -57
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
- 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/ssm-conv.cu +45 -13
- 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 +1 -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 +1 -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 +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -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 +1 -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 +1 -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 -4
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +26 -23
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +22 -4
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +2 -1
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1428 -743
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +45 -7
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
- 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 +170 -127
- 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 +125 -97
- 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 +148 -42
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +252 -62
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +87 -1
- 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 +96 -13
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +182 -57
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +71 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +9 -8
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -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-repl.h +74 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +1 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +529 -815
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +2522 -234
- 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 +291 -95
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
- 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 +244 -151
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
- data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -1
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +22 -9
- data/ext/sources/ggml/src/ggml-impl.h +6 -1
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +138 -13
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +32 -1
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +164 -28
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +80 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +190 -19
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +2 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +823 -322
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +54 -5
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +12248 -5907
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1819 -112
- 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_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/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
- 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_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/get_rows.cl +15 -9
- 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_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_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_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_f32_flat.cl +48 -64
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +15 -5
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +18 -11
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +35 -13
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +264 -192
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +33 -7
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +1 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +1 -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 +27 -3
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +67 -36
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +1 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +101 -44
- data/ext/sources/ggml/src/ggml-openvino/utils.h +23 -3
- data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
- data/ext/sources/ggml/src/ggml-quants.c +289 -114
- data/ext/sources/ggml/src/ggml-quants.h +3 -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 +50 -4
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +3 -1
- data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +41 -1
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +115 -13
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
- 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/element_wise.cpp +1 -90
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +0 -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 +7 -5
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
- 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 +69 -31
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -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 +823 -190
- 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/pad.cpp +27 -27
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
- 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-dkq512-dv512.cpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -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 +215 -53
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +11 -0
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2060 -535
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
- 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/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 +197 -48
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +115 -113
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
- 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_mmq_funcs.glsl +203 -0
- 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 +125 -64
- 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/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_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 +11 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +43 -10
- 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_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/rope_funcs.glsl +5 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
- 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 +11 -1
- 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 +171 -147
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2202 -283
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2610 -1403
- 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/binary.wgsl +8 -7
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +76 -95
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
- 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 -50
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
- 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.wgsl +183 -78
- 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 +655 -495
- 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 +8 -6
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +80 -409
- 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/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/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.wgsl +6 -4
- 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 +2 -3
- 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/unary.wgsl +68 -48
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
- data/ext/sources/ggml/src/ggml.c +110 -28
- data/ext/sources/ggml/src/gguf.cpp +173 -28
- 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 +56 -12
- 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 +411 -62
- 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_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_vad_segment.rb +1 -1
- data/test/test_whisper.rb +24 -6
- data/whispercpp.gemspec +2 -2
- metadata +215 -281
- 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 -48
- data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
- data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
- data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
- data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
- data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
- data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
- data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
- data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
- data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
- data/ext/sources/examples/talk-llama/llama-context.h +0 -359
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
- data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
- data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
- 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 -2735
- data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
- data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
- data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
- data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
- 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 -330
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
- 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 -1165
- 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 -752
- data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
- data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
- data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
- data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
- data/ext/sources/examples/talk-llama/llama-model.h +0 -597
- data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
- data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
- data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
- data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
- data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
- data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
- data/ext/sources/examples/talk-llama/llama.h +0 -1573
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
- 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 -137
- data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
- data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
- 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 -122
- data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
- data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
- data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
- data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
- 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 -111
- 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 -157
- 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 -195
- data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
- data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
- 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/jais2.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
- data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
- 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 -175
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
- data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/models.h +0 -704
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
- data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
- 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/paddleocr.cpp +0 -122
- 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 -320
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
- 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 -120
- data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
- data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
- 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 -164
- 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 -137
- 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/step35-iswa.cpp +0 -165
- 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 -1103
- 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/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
- 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-openvino/openvino/pass/eliminate_zp.cpp +0 -123
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
- data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
- 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.wgsl +0 -718
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
- 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
- /data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +0 -0
|
@@ -1,145 +0,0 @@
|
|
|
1
|
-
#include "models.h"
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
|
5
|
-
const int64_t n_embd_head = hparams.n_embd_head_v();
|
|
6
|
-
|
|
7
|
-
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
|
|
8
|
-
|
|
9
|
-
ggml_tensor * cur;
|
|
10
|
-
ggml_tensor * inpL;
|
|
11
|
-
|
|
12
|
-
inpL = build_inp_embd(model.tok_embd);
|
|
13
|
-
|
|
14
|
-
// inp_pos - contains the positions
|
|
15
|
-
ggml_tensor * inp_pos = build_inp_pos();
|
|
16
|
-
|
|
17
|
-
auto * inp_attn = build_attn_inp_kv();
|
|
18
|
-
|
|
19
|
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
20
|
-
|
|
21
|
-
for (int il = 0; il < n_layer; ++il) {
|
|
22
|
-
ggml_tensor * inpSA = inpL;
|
|
23
|
-
|
|
24
|
-
cur = build_norm(inpL,
|
|
25
|
-
model.layers[il].attn_norm, NULL,
|
|
26
|
-
LLM_NORM_RMS, il);
|
|
27
|
-
cb(cur, "attn_norm", il);
|
|
28
|
-
|
|
29
|
-
// self-attention
|
|
30
|
-
{
|
|
31
|
-
// compute Q and K and RoPE them
|
|
32
|
-
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
|
|
33
|
-
cb(Qcur, "Qcur", il);
|
|
34
|
-
if (model.layers[il].bq) {
|
|
35
|
-
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
|
|
36
|
-
cb(Qcur, "Qcur", il);
|
|
37
|
-
}
|
|
38
|
-
|
|
39
|
-
// B1.K
|
|
40
|
-
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
|
|
41
|
-
cb(Kcur, "Kcur", il);
|
|
42
|
-
if (model.layers[il].bk) {
|
|
43
|
-
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
|
|
44
|
-
cb(Kcur, "Kcur", il);
|
|
45
|
-
}
|
|
46
|
-
|
|
47
|
-
// B1.V
|
|
48
|
-
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
|
|
49
|
-
cb(Vcur, "Vcur", il);
|
|
50
|
-
if (model.layers[il].bv) {
|
|
51
|
-
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
|
|
52
|
-
cb(Vcur, "Vcur", il);
|
|
53
|
-
}
|
|
54
|
-
|
|
55
|
-
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
|
56
|
-
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
|
57
|
-
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
|
58
|
-
|
|
59
|
-
Qcur = ggml_rope_ext(
|
|
60
|
-
ctx0, Qcur, inp_pos, nullptr,
|
|
61
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
62
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
63
|
-
);
|
|
64
|
-
|
|
65
|
-
Kcur = ggml_rope_ext(
|
|
66
|
-
ctx0, Kcur, inp_pos, nullptr,
|
|
67
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
68
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
69
|
-
);
|
|
70
|
-
|
|
71
|
-
cb(Qcur, "Qcur", il);
|
|
72
|
-
cb(Kcur, "Kcur", il);
|
|
73
|
-
cb(Vcur, "Vcur", il);
|
|
74
|
-
|
|
75
|
-
cur = build_attn(inp_attn,
|
|
76
|
-
NULL, NULL,
|
|
77
|
-
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
|
|
78
|
-
|
|
79
|
-
cur = build_norm(cur,
|
|
80
|
-
model.layers[il].attn_sub_norm, NULL,
|
|
81
|
-
LLM_NORM_RMS, il);
|
|
82
|
-
cb(cur, "attn_sub_norm", il);
|
|
83
|
-
|
|
84
|
-
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
|
|
85
|
-
if (model.layers[il].bo) {
|
|
86
|
-
cur = ggml_add(ctx0, cur, model.layers[il].bo);
|
|
87
|
-
}
|
|
88
|
-
cb(cur, "attn_out", il);
|
|
89
|
-
}
|
|
90
|
-
|
|
91
|
-
if (il == n_layer - 1 && inp_out_ids) {
|
|
92
|
-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
93
|
-
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
94
|
-
}
|
|
95
|
-
|
|
96
|
-
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
97
|
-
cb(ffn_inp, "ffn_inp", il);
|
|
98
|
-
|
|
99
|
-
// feed-forward forward
|
|
100
|
-
cur = build_norm(ffn_inp,
|
|
101
|
-
model.layers[il].ffn_norm, NULL,
|
|
102
|
-
LLM_NORM_RMS, il);
|
|
103
|
-
cb(cur, "ffn_norm", il);
|
|
104
|
-
|
|
105
|
-
cur = build_ffn(cur,
|
|
106
|
-
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
|
|
107
|
-
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
|
|
108
|
-
NULL, NULL, NULL,
|
|
109
|
-
NULL,
|
|
110
|
-
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
|
111
|
-
cb(cur, "ffn_sub_out", il);
|
|
112
|
-
|
|
113
|
-
cur = build_norm(cur,
|
|
114
|
-
model.layers[il].ffn_sub_norm, NULL,
|
|
115
|
-
LLM_NORM_RMS, il);
|
|
116
|
-
cb(cur, "ffn_sub_norm", il);
|
|
117
|
-
|
|
118
|
-
cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
|
|
119
|
-
cb(cur, "ffn_down", il);
|
|
120
|
-
|
|
121
|
-
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
122
|
-
cb(cur, "l_out", il);
|
|
123
|
-
|
|
124
|
-
// input for next layer
|
|
125
|
-
inpL = cur;
|
|
126
|
-
}
|
|
127
|
-
|
|
128
|
-
cur = inpL;
|
|
129
|
-
|
|
130
|
-
cur = build_norm(cur,
|
|
131
|
-
model.output_norm, NULL,
|
|
132
|
-
LLM_NORM_RMS, -1);
|
|
133
|
-
|
|
134
|
-
cb(cur, "result_norm", -1);
|
|
135
|
-
res->t_embd = cur;
|
|
136
|
-
|
|
137
|
-
// lm_head
|
|
138
|
-
// FIXME: do not use model.tok_embd directly, duplicate as model.output
|
|
139
|
-
cur = build_lora_mm(model.tok_embd, cur);
|
|
140
|
-
|
|
141
|
-
cb(cur, "result_output", -1);
|
|
142
|
-
res->t_logits = cur;
|
|
143
|
-
|
|
144
|
-
ggml_build_forward_expand(gf, cur);
|
|
145
|
-
}
|
|
@@ -1,101 +0,0 @@
|
|
|
1
|
-
#include "models.h"
|
|
2
|
-
|
|
3
|
-
llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
|
4
|
-
const int64_t n_embd_head = hparams.n_embd_head_v();
|
|
5
|
-
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
|
|
6
|
-
|
|
7
|
-
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
|
|
8
|
-
|
|
9
|
-
ggml_tensor * cur;
|
|
10
|
-
ggml_tensor * inpL;
|
|
11
|
-
|
|
12
|
-
inpL = build_inp_embd(model.tok_embd);
|
|
13
|
-
|
|
14
|
-
auto * inp_attn = build_attn_inp_kv();
|
|
15
|
-
|
|
16
|
-
inpL = build_norm(inpL,
|
|
17
|
-
model.tok_norm,
|
|
18
|
-
model.tok_norm_b,
|
|
19
|
-
LLM_NORM, -1);
|
|
20
|
-
cb(inpL, "inp_norm", -1);
|
|
21
|
-
|
|
22
|
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
23
|
-
|
|
24
|
-
for (int il = 0; il < n_layer; ++il) {
|
|
25
|
-
cur = build_norm(inpL,
|
|
26
|
-
model.layers[il].attn_norm,
|
|
27
|
-
model.layers[il].attn_norm_b,
|
|
28
|
-
LLM_NORM, il);
|
|
29
|
-
cb(cur, "attn_norm", il);
|
|
30
|
-
|
|
31
|
-
// self-attention
|
|
32
|
-
{
|
|
33
|
-
cur = build_lora_mm(model.layers[il].wqkv, cur);
|
|
34
|
-
cb(cur, "wqkv", il);
|
|
35
|
-
|
|
36
|
-
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
|
|
37
|
-
cb(cur, "bqkv", il);
|
|
38
|
-
|
|
39
|
-
ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
|
|
40
|
-
ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
|
|
41
|
-
ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
|
|
42
|
-
|
|
43
|
-
cb(Qcur, "Qcur", il);
|
|
44
|
-
cb(Kcur, "Kcur", il);
|
|
45
|
-
cb(Vcur, "Vcur", il);
|
|
46
|
-
|
|
47
|
-
cur = build_attn(inp_attn,
|
|
48
|
-
model.layers[il].wo, model.layers[il].bo,
|
|
49
|
-
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
|
|
50
|
-
}
|
|
51
|
-
|
|
52
|
-
if (il == n_layer - 1 && inp_out_ids) {
|
|
53
|
-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
54
|
-
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
|
|
55
|
-
}
|
|
56
|
-
|
|
57
|
-
// Add the input
|
|
58
|
-
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
|
|
59
|
-
cb(ffn_inp, "ffn_inp", il);
|
|
60
|
-
|
|
61
|
-
// FF
|
|
62
|
-
{
|
|
63
|
-
cur = build_norm(ffn_inp,
|
|
64
|
-
model.layers[il].ffn_norm,
|
|
65
|
-
model.layers[il].ffn_norm_b,
|
|
66
|
-
LLM_NORM, il);
|
|
67
|
-
cb(cur, "ffn_norm", il);
|
|
68
|
-
|
|
69
|
-
cur = build_ffn(cur,
|
|
70
|
-
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
|
|
71
|
-
NULL, NULL, NULL,
|
|
72
|
-
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
|
|
73
|
-
NULL,
|
|
74
|
-
LLM_FFN_GELU, LLM_FFN_SEQ, il);
|
|
75
|
-
cb(cur, "ffn_out", il);
|
|
76
|
-
}
|
|
77
|
-
|
|
78
|
-
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
79
|
-
|
|
80
|
-
cur = build_cvec(cur, il);
|
|
81
|
-
cb(cur, "l_out", il);
|
|
82
|
-
|
|
83
|
-
// input for next layer
|
|
84
|
-
inpL = cur;
|
|
85
|
-
}
|
|
86
|
-
|
|
87
|
-
cur = build_norm(inpL,
|
|
88
|
-
model.output_norm,
|
|
89
|
-
model.output_norm_b,
|
|
90
|
-
LLM_NORM, -1);
|
|
91
|
-
|
|
92
|
-
cb(cur, "result_norm", -1);
|
|
93
|
-
res->t_embd = cur;
|
|
94
|
-
|
|
95
|
-
cur = build_lora_mm(model.output, cur);
|
|
96
|
-
|
|
97
|
-
cb(cur, "result_output", -1);
|
|
98
|
-
res->t_logits = cur;
|
|
99
|
-
|
|
100
|
-
ggml_build_forward_expand(gf, cur);
|
|
101
|
-
}
|
|
@@ -1,178 +0,0 @@
|
|
|
1
|
-
#include "models.h"
|
|
2
|
-
|
|
3
|
-
#include <float.h>
|
|
4
|
-
|
|
5
|
-
llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
|
6
|
-
const int64_t n_embd_head = hparams.n_embd_head_v();
|
|
7
|
-
|
|
8
|
-
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
|
|
9
|
-
GGML_ASSERT(n_embd_head == n_rot);
|
|
10
|
-
|
|
11
|
-
ggml_tensor * cur;
|
|
12
|
-
ggml_tensor * inpL;
|
|
13
|
-
|
|
14
|
-
inpL = build_inp_embd(model.tok_embd);
|
|
15
|
-
|
|
16
|
-
// inp_pos - contains the positions
|
|
17
|
-
ggml_tensor * inp_pos = build_inp_pos();
|
|
18
|
-
|
|
19
|
-
auto * inp_attn = build_attn_inp_kv();
|
|
20
|
-
|
|
21
|
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
22
|
-
|
|
23
|
-
for (int il = 0; il < n_layer; ++il) {
|
|
24
|
-
ggml_tensor * inpSA = inpL;
|
|
25
|
-
|
|
26
|
-
// norm
|
|
27
|
-
if (hparams.swin_norm) {
|
|
28
|
-
cur = inpL;
|
|
29
|
-
} else {
|
|
30
|
-
cur = build_norm(inpL,
|
|
31
|
-
model.layers[il].attn_norm, NULL,
|
|
32
|
-
LLM_NORM_RMS, il);
|
|
33
|
-
cb(cur, "attn_norm", il);
|
|
34
|
-
}
|
|
35
|
-
|
|
36
|
-
// self-attention
|
|
37
|
-
{
|
|
38
|
-
// compute Q and K and RoPE them
|
|
39
|
-
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
|
40
|
-
cb(Qcur, "Qcur", il);
|
|
41
|
-
|
|
42
|
-
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
|
|
43
|
-
cb(Kcur, "Kcur", il);
|
|
44
|
-
|
|
45
|
-
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
|
|
46
|
-
cb(Vcur, "Vcur", il);
|
|
47
|
-
|
|
48
|
-
if (model.layers[il].attn_q_norm) {
|
|
49
|
-
Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
|
|
50
|
-
ggml_element_size(Qcur) * n_embd_head,
|
|
51
|
-
ggml_element_size(Qcur) * n_embd_head * n_head,
|
|
52
|
-
0);
|
|
53
|
-
cb(Qcur, "Qcur", il);
|
|
54
|
-
|
|
55
|
-
Qcur = build_norm(Qcur,
|
|
56
|
-
model.layers[il].attn_q_norm,
|
|
57
|
-
model.layers[il].attn_q_norm_b,
|
|
58
|
-
LLM_NORM, il);
|
|
59
|
-
cb(Qcur, "Qcur", il);
|
|
60
|
-
}
|
|
61
|
-
|
|
62
|
-
if (model.layers[il].attn_k_norm) {
|
|
63
|
-
Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
|
|
64
|
-
ggml_element_size(Kcur) * n_embd_head,
|
|
65
|
-
ggml_element_size(Kcur) * n_embd_head * n_head_kv,
|
|
66
|
-
0);
|
|
67
|
-
cb(Kcur, "Kcur", il);
|
|
68
|
-
|
|
69
|
-
Kcur = build_norm(Kcur,
|
|
70
|
-
model.layers[il].attn_k_norm,
|
|
71
|
-
model.layers[il].attn_k_norm_b,
|
|
72
|
-
LLM_NORM, il);
|
|
73
|
-
cb(Kcur, "Kcur", il);
|
|
74
|
-
}
|
|
75
|
-
|
|
76
|
-
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
|
77
|
-
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
|
78
|
-
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
|
79
|
-
|
|
80
|
-
Qcur = ggml_rope_ext(
|
|
81
|
-
ctx0, Qcur, inp_pos, nullptr,
|
|
82
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
83
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
84
|
-
);
|
|
85
|
-
|
|
86
|
-
Kcur = ggml_rope_ext(
|
|
87
|
-
ctx0, Kcur, inp_pos, nullptr,
|
|
88
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
89
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
90
|
-
);
|
|
91
|
-
|
|
92
|
-
cb(Qcur, "Qcur", il);
|
|
93
|
-
cb(Kcur, "Kcur", il);
|
|
94
|
-
cb(Vcur, "Vcur", il);
|
|
95
|
-
|
|
96
|
-
cur = build_attn(inp_attn,
|
|
97
|
-
model.layers[il].wo, nullptr,
|
|
98
|
-
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
|
|
99
|
-
}
|
|
100
|
-
|
|
101
|
-
if (il == n_layer - 1 && inp_out_ids) {
|
|
102
|
-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
103
|
-
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
104
|
-
}
|
|
105
|
-
|
|
106
|
-
if (hparams.swin_norm) {
|
|
107
|
-
cur = build_norm(cur,
|
|
108
|
-
model.layers[il].attn_norm, NULL,
|
|
109
|
-
LLM_NORM_RMS, il);
|
|
110
|
-
}
|
|
111
|
-
|
|
112
|
-
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
113
|
-
cb(ffn_inp, "ffn_inp", il);
|
|
114
|
-
|
|
115
|
-
// feed-forward network
|
|
116
|
-
if (!hparams.swin_norm) {
|
|
117
|
-
cur = build_norm(ffn_inp,
|
|
118
|
-
model.layers[il].ffn_norm, NULL,
|
|
119
|
-
LLM_NORM_RMS, il);
|
|
120
|
-
cb(cur, "ffn_norm", il);
|
|
121
|
-
}
|
|
122
|
-
|
|
123
|
-
cur = build_ffn(cur,
|
|
124
|
-
model.layers[il].ffn_up, NULL, NULL,
|
|
125
|
-
model.layers[il].ffn_gate, NULL, NULL,
|
|
126
|
-
model.layers[il].ffn_down, NULL, NULL,
|
|
127
|
-
NULL,
|
|
128
|
-
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
|
129
|
-
cb(cur, "ffn_out", il);
|
|
130
|
-
|
|
131
|
-
if (hparams.swin_norm) {
|
|
132
|
-
cur = build_norm(cur,
|
|
133
|
-
model.layers[il].ffn_norm, NULL,
|
|
134
|
-
LLM_NORM_RMS, il);
|
|
135
|
-
cb(cur, "ffn_norm", il);
|
|
136
|
-
}
|
|
137
|
-
|
|
138
|
-
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
139
|
-
cb(cur, "ffn_out", il);
|
|
140
|
-
|
|
141
|
-
cur = build_cvec(cur, il);
|
|
142
|
-
cb(cur, "l_out", il);
|
|
143
|
-
|
|
144
|
-
// input for next layer
|
|
145
|
-
inpL = cur;
|
|
146
|
-
}
|
|
147
|
-
|
|
148
|
-
cur = inpL;
|
|
149
|
-
|
|
150
|
-
cur = build_norm(cur,
|
|
151
|
-
model.output_norm, NULL,
|
|
152
|
-
LLM_NORM_RMS, -1);
|
|
153
|
-
|
|
154
|
-
cb(cur, "result_norm", -1);
|
|
155
|
-
res->t_embd = cur;
|
|
156
|
-
|
|
157
|
-
// lm_head
|
|
158
|
-
cur = build_lora_mm(model.output, cur);
|
|
159
|
-
cb(cur, "result_output_with_img_logits", -1);
|
|
160
|
-
|
|
161
|
-
// TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
|
|
162
|
-
// Needs to be removed once image outputs are supported.
|
|
163
|
-
int img_token_end_idx = 8196;
|
|
164
|
-
int img_token_start_idx = 4;
|
|
165
|
-
int num_img_tokens = img_token_end_idx - img_token_start_idx;
|
|
166
|
-
// creates 1d tensor of size num_img_tokens and values -FLT_MAX,
|
|
167
|
-
// which ensures that text token values are always at least larger than image token values
|
|
168
|
-
ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
|
|
169
|
-
img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
|
|
170
|
-
cb(img_logits, "img_logits", -1);
|
|
171
|
-
|
|
172
|
-
cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
|
|
173
|
-
|
|
174
|
-
cb(cur, "result_output", -1);
|
|
175
|
-
res->t_logits = cur;
|
|
176
|
-
|
|
177
|
-
ggml_build_forward_expand(gf, cur);
|
|
178
|
-
}
|
|
@@ -1,132 +0,0 @@
|
|
|
1
|
-
#include "models.h"
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
|
5
|
-
const int64_t n_embd_head = hparams.n_embd_head_v();
|
|
6
|
-
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
|
|
7
|
-
|
|
8
|
-
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
|
|
9
|
-
|
|
10
|
-
ggml_tensor * cur;
|
|
11
|
-
ggml_tensor * inpL;
|
|
12
|
-
|
|
13
|
-
inpL = build_inp_embd(model.tok_embd);
|
|
14
|
-
|
|
15
|
-
// inp_pos - contains the positions
|
|
16
|
-
ggml_tensor * inp_pos = build_inp_pos();
|
|
17
|
-
|
|
18
|
-
auto * inp_attn = build_attn_inp_kv();
|
|
19
|
-
|
|
20
|
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
21
|
-
|
|
22
|
-
for (int il = 0; il < n_layer; ++il) {
|
|
23
|
-
ggml_tensor * inpSA = inpL;
|
|
24
|
-
|
|
25
|
-
cur = build_norm(inpL,
|
|
26
|
-
model.layers[il].attn_norm,
|
|
27
|
-
NULL,
|
|
28
|
-
LLM_NORM_RMS, il);
|
|
29
|
-
cb(cur, "attn_norm", il);
|
|
30
|
-
|
|
31
|
-
// self-attention
|
|
32
|
-
{
|
|
33
|
-
ggml_tensor * Qcur = nullptr;
|
|
34
|
-
ggml_tensor * Kcur = nullptr;
|
|
35
|
-
ggml_tensor * Vcur = nullptr;
|
|
36
|
-
|
|
37
|
-
if (model.layers[il].wqkv == nullptr) {
|
|
38
|
-
Qcur = build_lora_mm(model.layers[il].wq, cur);
|
|
39
|
-
if (model.layers[il].bq) {
|
|
40
|
-
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
|
|
41
|
-
}
|
|
42
|
-
Kcur = build_lora_mm(model.layers[il].wk, cur);
|
|
43
|
-
if (model.layers[il].bk) {
|
|
44
|
-
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
|
|
45
|
-
}
|
|
46
|
-
Vcur = build_lora_mm(model.layers[il].wv, cur);
|
|
47
|
-
if (model.layers[il].bv) {
|
|
48
|
-
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
|
|
49
|
-
}
|
|
50
|
-
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
|
51
|
-
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
|
52
|
-
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
|
53
|
-
} else {
|
|
54
|
-
cur = build_lora_mm(model.layers[il].wqkv, cur);
|
|
55
|
-
cb(cur, "wqkv", il);
|
|
56
|
-
if (model.layers[il].bqkv) {
|
|
57
|
-
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
|
|
58
|
-
cb(cur, "bqkv", il);
|
|
59
|
-
}
|
|
60
|
-
Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
|
|
61
|
-
Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
|
|
62
|
-
Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
|
|
63
|
-
}
|
|
64
|
-
|
|
65
|
-
//printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
|
|
66
|
-
Qcur = ggml_rope_ext(
|
|
67
|
-
ctx0, Qcur, inp_pos, nullptr,
|
|
68
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
69
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
70
|
-
);
|
|
71
|
-
|
|
72
|
-
Kcur = ggml_rope_ext(
|
|
73
|
-
ctx0, Kcur, inp_pos, nullptr,
|
|
74
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
75
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
76
|
-
);
|
|
77
|
-
|
|
78
|
-
cb(Qcur, "Qcur", il);
|
|
79
|
-
cb(Kcur, "Kcur", il);
|
|
80
|
-
cb(Vcur, "Vcur", il);
|
|
81
|
-
|
|
82
|
-
cur = build_attn(inp_attn,
|
|
83
|
-
model.layers[il].wo, NULL,
|
|
84
|
-
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
|
|
85
|
-
}
|
|
86
|
-
|
|
87
|
-
if (il == n_layer - 1 && inp_out_ids) {
|
|
88
|
-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
89
|
-
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
90
|
-
}
|
|
91
|
-
|
|
92
|
-
// Add the input
|
|
93
|
-
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
94
|
-
cb(ffn_inp, "ffn_inp", il);
|
|
95
|
-
|
|
96
|
-
// FF
|
|
97
|
-
{
|
|
98
|
-
cur = build_norm(ffn_inp,
|
|
99
|
-
model.layers[il].ffn_norm,
|
|
100
|
-
NULL,
|
|
101
|
-
LLM_NORM_RMS, il);
|
|
102
|
-
cb(cur, "ffn_norm", il);
|
|
103
|
-
|
|
104
|
-
cur = build_ffn(cur,
|
|
105
|
-
model.layers[il].ffn_up, NULL, NULL,
|
|
106
|
-
NULL, NULL, NULL,
|
|
107
|
-
model.layers[il].ffn_down, NULL, NULL,
|
|
108
|
-
NULL,
|
|
109
|
-
LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
|
|
110
|
-
cb(cur, "ffn_out", il);
|
|
111
|
-
|
|
112
|
-
}
|
|
113
|
-
|
|
114
|
-
inpL = ggml_add(ctx0, cur, ffn_inp);
|
|
115
|
-
cb(inpL, "l_out", il);
|
|
116
|
-
}
|
|
117
|
-
|
|
118
|
-
cur = build_norm(inpL,
|
|
119
|
-
model.output_norm,
|
|
120
|
-
NULL,
|
|
121
|
-
LLM_NORM_RMS, -1);
|
|
122
|
-
|
|
123
|
-
cb(cur, "result_norm", -1);
|
|
124
|
-
res->t_embd = cur;
|
|
125
|
-
|
|
126
|
-
cur = build_lora_mm(model.output, cur);
|
|
127
|
-
|
|
128
|
-
cb(cur, "result_output", -1);
|
|
129
|
-
res->t_logits = cur;
|
|
130
|
-
|
|
131
|
-
ggml_build_forward_expand(gf, cur);
|
|
132
|
-
}
|
|
@@ -1,111 +0,0 @@
|
|
|
1
|
-
#include "models.h"
|
|
2
|
-
|
|
3
|
-
llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
|
4
|
-
const int64_t n_embd_head = hparams.n_embd_head_v();
|
|
5
|
-
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
|
|
6
|
-
|
|
7
|
-
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
|
|
8
|
-
GGML_ASSERT(n_embd_head == n_rot);
|
|
9
|
-
|
|
10
|
-
ggml_tensor * cur;
|
|
11
|
-
ggml_tensor * inpL;
|
|
12
|
-
|
|
13
|
-
inpL = build_inp_embd(model.tok_embd);
|
|
14
|
-
|
|
15
|
-
// inp_pos - contains the positions
|
|
16
|
-
ggml_tensor * inp_pos = build_inp_pos();
|
|
17
|
-
|
|
18
|
-
auto * inp_attn = build_attn_inp_kv();
|
|
19
|
-
|
|
20
|
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
21
|
-
|
|
22
|
-
for (int il = 0; il < n_layer; ++il) {
|
|
23
|
-
cur = build_norm(inpL,
|
|
24
|
-
model.layers[il].attn_norm,
|
|
25
|
-
model.layers[il].attn_norm_b,
|
|
26
|
-
LLM_NORM, il);
|
|
27
|
-
cb(cur, "attn_norm", il);
|
|
28
|
-
|
|
29
|
-
// self-attention
|
|
30
|
-
{
|
|
31
|
-
cur = build_lora_mm(model.layers[il].wqkv, cur);
|
|
32
|
-
cb(cur, "wqkv", il);
|
|
33
|
-
|
|
34
|
-
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
|
|
35
|
-
cb(cur, "bqkv", il);
|
|
36
|
-
|
|
37
|
-
ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
|
|
38
|
-
ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
|
|
39
|
-
ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
|
|
40
|
-
|
|
41
|
-
Qcur = ggml_rope_ext(
|
|
42
|
-
ctx0, Qcur, inp_pos, nullptr,
|
|
43
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
44
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
45
|
-
);
|
|
46
|
-
|
|
47
|
-
Kcur = ggml_rope_ext(
|
|
48
|
-
ctx0, Kcur, inp_pos, nullptr,
|
|
49
|
-
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
50
|
-
ext_factor, attn_factor, beta_fast, beta_slow
|
|
51
|
-
);
|
|
52
|
-
|
|
53
|
-
cb(Qcur, "Qcur", il);
|
|
54
|
-
cb(Kcur, "Kcur", il);
|
|
55
|
-
cb(Vcur, "Vcur", il);
|
|
56
|
-
|
|
57
|
-
cur = build_attn(inp_attn,
|
|
58
|
-
model.layers[il].wo, model.layers[il].bo,
|
|
59
|
-
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
|
|
60
|
-
}
|
|
61
|
-
|
|
62
|
-
if (il == n_layer - 1 && inp_out_ids) {
|
|
63
|
-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
64
|
-
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
|
|
65
|
-
}
|
|
66
|
-
|
|
67
|
-
// add the input
|
|
68
|
-
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
|
|
69
|
-
cb(ffn_inp, "ffn_inp", il);
|
|
70
|
-
|
|
71
|
-
// FF
|
|
72
|
-
{
|
|
73
|
-
cur = build_norm(ffn_inp,
|
|
74
|
-
model.layers[il].ffn_norm,
|
|
75
|
-
model.layers[il].ffn_norm_b,
|
|
76
|
-
LLM_NORM, il);
|
|
77
|
-
cb(cur, "ffn_norm", il);
|
|
78
|
-
|
|
79
|
-
cur = build_ffn(cur,
|
|
80
|
-
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
|
|
81
|
-
NULL, NULL, NULL,
|
|
82
|
-
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
|
|
83
|
-
NULL,
|
|
84
|
-
LLM_FFN_GELU, LLM_FFN_SEQ, il);
|
|
85
|
-
cb(cur, "ffn_out", il);
|
|
86
|
-
}
|
|
87
|
-
|
|
88
|
-
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
89
|
-
|
|
90
|
-
cur = build_cvec(cur, il);
|
|
91
|
-
cb(cur, "l_out", il);
|
|
92
|
-
|
|
93
|
-
// input for next layer
|
|
94
|
-
inpL = cur;
|
|
95
|
-
}
|
|
96
|
-
|
|
97
|
-
cur = build_norm(inpL,
|
|
98
|
-
model.output_norm,
|
|
99
|
-
model.output_norm_b,
|
|
100
|
-
LLM_NORM, -1);
|
|
101
|
-
|
|
102
|
-
cb(cur, "result_norm", -1);
|
|
103
|
-
res->t_embd = cur;
|
|
104
|
-
|
|
105
|
-
cur = build_lora_mm(model.output, cur);
|
|
106
|
-
|
|
107
|
-
cb(cur, "result_output", -1);
|
|
108
|
-
res->t_logits = cur;
|
|
109
|
-
|
|
110
|
-
ggml_build_forward_expand(gf, cur);
|
|
111
|
-
}
|