whispercpp 1.3.1 → 1.3.2
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/.gitignore +4 -3
- data/README.md +92 -31
- data/Rakefile +26 -7
- data/ext/.gitignore +5 -7
- data/ext/dependencies.rb +61 -0
- data/ext/extconf.rb +21 -198
- data/ext/options.rb +221 -0
- data/ext/ruby_whisper.c +159 -0
- data/ext/ruby_whisper.h +17 -2
- data/ext/ruby_whisper_context.c +641 -0
- data/ext/ruby_whisper_error.c +52 -0
- data/ext/ruby_whisper_model.c +232 -0
- data/ext/ruby_whisper_params.c +1301 -0
- data/ext/ruby_whisper_segment.c +143 -0
- data/ext/ruby_whisper_transcribe.cpp +87 -0
- data/ext/ruby_whisper_vad_params.c +288 -0
- data/ext/sources/.dockerignore +3 -0
- data/ext/sources/.github/workflows/bindings-ruby.yml +21 -0
- data/ext/sources/CMakeGraphVizOptions.cmake +8 -0
- data/ext/sources/CMakeLists.txt +251 -0
- data/ext/sources/bindings/javascript/CMakeLists.txt +41 -0
- data/ext/sources/bindings/javascript/emscripten.cpp +93 -0
- data/ext/sources/bindings/javascript/libwhisper.worker.js +1 -0
- data/ext/sources/bindings/javascript/package-tmpl.json +26 -0
- data/ext/sources/bindings/javascript/package.json +26 -0
- data/ext/sources/bindings/javascript/whisper.js +19 -0
- data/ext/sources/build-xcframework.sh +547 -0
- data/ext/sources/ci/run.sh +336 -0
- data/ext/sources/close-issue.yml +28 -0
- data/ext/sources/cmake/DefaultTargetOptions.cmake +16 -0
- data/ext/sources/cmake/FindFFmpeg.cmake +163 -0
- data/ext/sources/cmake/build-info.cmake +60 -0
- data/ext/sources/cmake/git-vars.cmake +22 -0
- data/ext/sources/cmake/whisper-config.cmake.in +65 -0
- data/ext/sources/cmake/whisper.pc.in +10 -0
- data/ext/sources/examples/CMakeLists.txt +124 -0
- data/ext/sources/examples/addon.node/CMakeLists.txt +31 -0
- data/ext/sources/examples/addon.node/__test__/whisper.spec.js +37 -0
- data/ext/sources/examples/addon.node/addon.cpp +438 -0
- data/ext/sources/examples/addon.node/index.js +54 -0
- data/ext/sources/examples/addon.node/package.json +16 -0
- data/ext/sources/examples/bench/CMakeLists.txt +8 -0
- data/ext/sources/examples/bench/bench.cpp +175 -0
- data/ext/sources/examples/bench.wasm/CMakeLists.txt +49 -0
- data/ext/sources/examples/bench.wasm/emscripten.cpp +87 -0
- data/ext/sources/examples/bench.wasm/index-tmpl.html +284 -0
- data/ext/sources/examples/cli/CMakeLists.txt +8 -0
- data/ext/sources/examples/cli/cli.cpp +1294 -0
- data/ext/sources/examples/coi-serviceworker.js +146 -0
- data/ext/sources/examples/command/CMakeLists.txt +10 -0
- data/ext/sources/examples/command/command.cpp +776 -0
- data/ext/sources/examples/command/commands.txt +9 -0
- data/ext/sources/examples/command.wasm/CMakeLists.txt +50 -0
- data/ext/sources/examples/command.wasm/emscripten.cpp +327 -0
- data/ext/sources/examples/command.wasm/index-tmpl.html +414 -0
- data/ext/sources/examples/common-ggml.cpp +238 -0
- data/ext/sources/examples/common-ggml.h +18 -0
- data/ext/sources/examples/common-sdl.cpp +227 -0
- data/ext/sources/examples/common-sdl.h +49 -0
- data/ext/sources/examples/common-whisper.cpp +168 -0
- data/ext/sources/examples/common-whisper.h +24 -0
- data/ext/sources/examples/common.cpp +675 -0
- data/ext/sources/examples/common.h +322 -0
- data/ext/sources/examples/deprecation-warning/CMakeLists.txt +6 -0
- data/ext/sources/examples/deprecation-warning/deprecation-warning.cpp +38 -0
- data/ext/sources/examples/ffmpeg-transcode.cpp +368 -0
- data/ext/sources/examples/generate-karaoke.sh +57 -0
- data/ext/sources/examples/grammar-parser.cpp +423 -0
- data/ext/sources/examples/grammar-parser.h +29 -0
- data/ext/sources/examples/helpers.js +191 -0
- data/ext/sources/examples/json.hpp +24596 -0
- data/ext/sources/examples/livestream.sh +112 -0
- data/ext/sources/examples/lsp/CMakeLists.txt +9 -0
- data/ext/sources/examples/lsp/lsp.cpp +467 -0
- data/ext/sources/examples/lsp/whisper.vim +362 -0
- data/ext/sources/examples/miniaudio.h +93468 -0
- data/ext/sources/examples/python/test_whisper_processor.py +7 -0
- data/ext/sources/examples/python/whisper_processor.py +54 -0
- data/ext/sources/examples/quantize/CMakeLists.txt +6 -0
- data/ext/sources/examples/quantize/quantize.cpp +223 -0
- data/ext/sources/examples/server/CMakeLists.txt +12 -0
- data/ext/sources/examples/server/bench.js +29 -0
- data/ext/sources/examples/server/httplib.h +10497 -0
- data/ext/sources/examples/server/server.cpp +1091 -0
- data/ext/sources/examples/server.py +115 -0
- data/ext/sources/examples/stb_vorbis.c +5584 -0
- data/ext/sources/examples/stream/CMakeLists.txt +10 -0
- data/ext/sources/examples/stream/stream.cpp +429 -0
- data/ext/sources/examples/stream.wasm/CMakeLists.txt +49 -0
- data/ext/sources/examples/stream.wasm/emscripten.cpp +216 -0
- data/ext/sources/examples/stream.wasm/index-tmpl.html +414 -0
- data/ext/sources/examples/sycl/CMakeLists.txt +9 -0
- data/ext/sources/examples/sycl/build.sh +22 -0
- data/ext/sources/examples/sycl/ls-sycl-device.cpp +11 -0
- data/ext/sources/examples/sycl/run-whisper.sh +17 -0
- data/ext/sources/examples/talk-llama/CMakeLists.txt +40 -0
- data/ext/sources/examples/talk-llama/eleven-labs.py +80 -0
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +388 -0
- data/ext/sources/examples/talk-llama/llama-adapter.h +76 -0
- data/ext/sources/examples/talk-llama/llama-arch.cpp +1746 -0
- data/ext/sources/examples/talk-llama/llama-arch.h +437 -0
- data/ext/sources/examples/talk-llama/llama-batch.cpp +374 -0
- data/ext/sources/examples/talk-llama/llama-batch.h +89 -0
- data/ext/sources/examples/talk-llama/llama-chat.cpp +663 -0
- data/ext/sources/examples/talk-llama/llama-chat.h +58 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +2676 -0
- data/ext/sources/examples/talk-llama/llama-context.h +276 -0
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +5 -0
- data/ext/sources/examples/talk-llama/llama-cparams.h +41 -0
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +1229 -0
- data/ext/sources/examples/talk-llama/llama-grammar.h +173 -0
- data/ext/sources/examples/talk-llama/llama-graph.cpp +1618 -0
- data/ext/sources/examples/talk-llama/llama-graph.h +640 -0
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +95 -0
- data/ext/sources/examples/talk-llama/llama-hparams.h +190 -0
- data/ext/sources/examples/talk-llama/llama-impl.cpp +167 -0
- data/ext/sources/examples/talk-llama/llama-impl.h +61 -0
- data/ext/sources/examples/talk-llama/llama-io.cpp +15 -0
- data/ext/sources/examples/talk-llama/llama-io.h +35 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2739 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +502 -0
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +379 -0
- data/ext/sources/examples/talk-llama/llama-memory.cpp +1 -0
- data/ext/sources/examples/talk-llama/llama-memory.h +32 -0
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +600 -0
- data/ext/sources/examples/talk-llama/llama-mmap.h +68 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +1138 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.h +169 -0
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +281 -0
- data/ext/sources/examples/talk-llama/llama-model-saver.h +37 -0
- data/ext/sources/examples/talk-llama/llama-model.cpp +13814 -0
- data/ext/sources/examples/talk-llama/llama-model.h +425 -0
- data/ext/sources/examples/talk-llama/llama-quant.cpp +966 -0
- data/ext/sources/examples/talk-llama/llama-quant.h +1 -0
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +2575 -0
- data/ext/sources/examples/talk-llama/llama-sampling.h +32 -0
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +3340 -0
- data/ext/sources/examples/talk-llama/llama-vocab.h +131 -0
- data/ext/sources/examples/talk-llama/llama.cpp +354 -0
- data/ext/sources/examples/talk-llama/llama.h +1377 -0
- data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +23 -0
- data/ext/sources/examples/talk-llama/speak +40 -0
- data/ext/sources/examples/talk-llama/speak.bat +1 -0
- data/ext/sources/examples/talk-llama/speak.ps1 +14 -0
- data/ext/sources/examples/talk-llama/talk-llama.cpp +808 -0
- data/ext/sources/examples/talk-llama/unicode-data.cpp +7034 -0
- data/ext/sources/examples/talk-llama/unicode-data.h +20 -0
- data/ext/sources/examples/talk-llama/unicode.cpp +849 -0
- data/ext/sources/examples/talk-llama/unicode.h +66 -0
- data/ext/sources/examples/vad-speech-segments/CMakeLists.txt +8 -0
- data/ext/sources/examples/vad-speech-segments/speech.cpp +143 -0
- data/ext/sources/examples/wchess/CMakeLists.txt +10 -0
- data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +19 -0
- data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +803 -0
- data/ext/sources/examples/wchess/libwchess/Chessboard.h +33 -0
- data/ext/sources/examples/wchess/libwchess/WChess.cpp +193 -0
- data/ext/sources/examples/wchess/libwchess/WChess.h +63 -0
- data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +117 -0
- data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +8 -0
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +249 -0
- data/ext/sources/examples/whisper.wasm/CMakeLists.txt +50 -0
- data/ext/sources/examples/whisper.wasm/emscripten.cpp +118 -0
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +658 -0
- data/ext/sources/ggml/CMakeLists.txt +390 -0
- data/ext/sources/ggml/cmake/BuildTypes.cmake +54 -0
- data/ext/sources/ggml/cmake/GitVars.cmake +22 -0
- data/ext/sources/ggml/cmake/common.cmake +26 -0
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +152 -0
- data/ext/{ggml → sources/ggml}/include/ggml-alloc.h +1 -1
- data/ext/{ggml → sources/ggml}/include/ggml-backend.h +9 -7
- data/ext/{ggml → sources/ggml}/include/ggml-cpp.h +2 -1
- data/ext/{ggml → sources/ggml}/include/ggml-cpu.h +9 -1
- data/ext/{ggml → sources/ggml}/include/ggml-metal.h +1 -1
- data/ext/{ggml → sources/ggml}/include/ggml-opt.h +49 -28
- data/ext/{ggml → sources/ggml}/include/ggml-rpc.h +6 -1
- data/ext/{ggml → sources/ggml}/include/ggml-vulkan.h +0 -2
- data/ext/{ggml → sources/ggml}/include/ggml.h +182 -265
- data/ext/sources/ggml/include/gguf.h +202 -0
- data/ext/sources/ggml/src/CMakeLists.txt +346 -0
- data/ext/{ggml → sources/ggml}/src/ggml-alloc.c +34 -29
- data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +107 -0
- data/ext/{ggml → sources/ggml}/src/ggml-backend-impl.h +1 -2
- data/ext/{ggml → sources/ggml}/src/ggml-backend-reg.cpp +87 -53
- data/ext/{ggml → sources/ggml}/src/ggml-backend.cpp +26 -14
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +87 -0
- data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +74 -0
- data/ext/sources/ggml/src/ggml-cann/Doxyfile +2579 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.cpp +10 -4
- data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.h +5 -5
- data/ext/{ggml → sources/ggml}/src/ggml-cann/aclnn_ops.cpp +1272 -1506
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +1125 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/common.h +135 -1
- data/ext/{ggml → sources/ggml}/src/ggml-cann/ggml-cann.cpp +564 -146
- data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +30 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/dup.cpp +3 -5
- data/ext/{ggml → sources/ggml}/src/ggml-common.h +12 -8
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +504 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.cpp +2 -1
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +158 -0
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.h +16 -0
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
- data/ext/sources/ggml/src/ggml-cpu/common.h +72 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/cpu-feats-x86.cpp +5 -1
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +6431 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-impl.h +163 -41
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-quants.c +4029 -1117
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +3510 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu.cpp +67 -18
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +337 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +95 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +482 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +3544 -0
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +8903 -0
- data/ext/sources/ggml/src/ggml-cpu/ops.h +110 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +892 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +186 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.h +28 -0
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +252 -0
- data/ext/sources/ggml/src/ggml-cpu/vec.h +818 -0
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +184 -0
- data/ext/sources/ggml/src/ggml-cuda/acc.cu +61 -0
- data/ext/sources/ggml/src/ggml-cuda/acc.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/arange.cu +34 -0
- data/ext/sources/ggml/src/ggml-cuda/arange.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/argmax.cu +91 -0
- data/ext/sources/ggml/src/ggml-cuda/argmax.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +104 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +363 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +9 -0
- data/ext/sources/ggml/src/ggml-cuda/clamp.cu +45 -0
- data/ext/sources/ggml/src/ggml-cuda/clamp.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +828 -0
- data/ext/sources/ggml/src/ggml-cuda/concat.cu +221 -0
- data/ext/sources/ggml/src/ggml-cuda/concat.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +89 -0
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +730 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +26 -0
- data/ext/sources/ggml/src/ggml-cuda/count-equal.cu +64 -0
- data/ext/sources/ggml/src/ggml-cuda/count-equal.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/cp-async.cuh +57 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +705 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cuh +11 -0
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +189 -0
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +103 -0
- data/ext/sources/ggml/src/ggml-cuda/diagmask.cu +40 -0
- data/ext/sources/ggml/src/ggml-cuda/diagmask.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +881 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +1471 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +357 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +365 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +482 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +472 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +634 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +346 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +275 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cuh +15 -0
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +3505 -0
- data/ext/sources/ggml/src/ggml-cuda/gla.cu +93 -0
- data/ext/sources/ggml/src/ggml-cuda/gla.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +103 -0
- data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +396 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +324 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +3217 -0
- data/ext/sources/ggml/src/ggml-cuda/mmv.cu +336 -0
- data/ext/sources/ggml/src/ggml-cuda/mmv.cuh +12 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +595 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +12 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +458 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cuh +11 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +68 -0
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +49 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/pool2d.cu +94 -0
- data/ext/sources/ggml/src/ggml-cuda/pool2d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +190 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +27 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +456 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +31 -0
- data/ext/sources/ggml/src/ggml-cuda/scale.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +283 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +148 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +153 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/sum.cu +45 -0
- data/ext/sources/ggml/src/ggml-cuda/sum.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +39 -0
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +78 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +47 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +289 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +59 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cu +51 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +1135 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/cuda.h +1 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/hip.h +57 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/musa.h +7 -1
- data/ext/sources/ggml/src/ggml-cuda/wkv.cu +199 -0
- data/ext/sources/ggml/src/ggml-cuda/wkv.cuh +7 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +131 -0
- data/ext/{ggml → sources/ggml}/src/ggml-impl.h +64 -19
- data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +166 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +112 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +58 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +25 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +30 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +22 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +17 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +31 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +31 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +38 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +39 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +44 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +69 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +51 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +33 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +35 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +140 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +106 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +73 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +28 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +84 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +21 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +53 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +19 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +23 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +22 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +72 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +71 -0
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +120 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +622 -0
- data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.m +2178 -1064
- data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.metal +1575 -1218
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +113 -0
- data/ext/sources/ggml/src/ggml-musa/mudnn.cu +112 -0
- data/ext/sources/ggml/src/ggml-musa/mudnn.cuh +12 -0
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +96 -0
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +5124 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +83 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +62 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl +268 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl +274 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +79 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl +190 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +81 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +96 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rope.cl +721 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +16 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +87 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +87 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +86 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +86 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +84 -0
- data/ext/{ggml → sources/ggml}/src/ggml-opt.cpp +373 -190
- data/ext/{ggml → sources/ggml}/src/ggml-quants.c +114 -120
- data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
- data/ext/{ggml → sources/ggml}/src/ggml-rpc/ggml-rpc.cpp +480 -73
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +189 -0
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +37 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +345 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +39 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/common.cpp +20 -32
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +589 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/concat.cpp +32 -33
- data/ext/sources/ggml/src/ggml-sycl/concat.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/conv.cpp +4 -2
- data/ext/sources/ggml/src/ggml-sycl/conv.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/convert.cpp +104 -28
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +34 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +700 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +11 -0
- data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +791 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/dmmv.cpp +156 -17
- data/ext/sources/ggml/src/ggml-sycl/dmmv.hpp +27 -0
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +2957 -0
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1511 -0
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +75 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +99 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +309 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/ggml-sycl.cpp +1004 -1240
- data/ext/sources/ggml/src/ggml-sycl/gla.cpp +106 -0
- data/ext/sources/ggml/src/ggml-sycl/gla.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +136 -0
- data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +21 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/mmq.cpp +0 -1
- data/ext/sources/ggml/src/ggml-sycl/mmq.hpp +33 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/mmvq.cpp +261 -166
- data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +27 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/norm.cpp +204 -81
- data/ext/sources/ggml/src/ggml-sycl/norm.hpp +26 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/outprod.cpp +8 -17
- data/ext/sources/ggml/src/ggml-sycl/outprod.hpp +10 -0
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +74 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +83 -0
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +361 -0
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/softmax.cpp +35 -25
- data/ext/sources/ggml/src/ggml-sycl/softmax.hpp +20 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/tsembd.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/tsembd.hpp +20 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +1215 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +293 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.hpp +10 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +196 -0
- data/ext/sources/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
- data/ext/{ggml → sources/ggml}/src/ggml-vulkan/ggml-vulkan.cpp +3130 -1087
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +51 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +69 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +23 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +242 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +462 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +699 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp +13 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +48 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +31 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +337 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +360 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +267 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +59 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +25 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +23 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +64 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp +9 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp +76 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +33 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +100 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +169 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +118 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +82 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +79 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +90 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +118 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +154 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +130 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +136 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +167 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +868 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +441 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +442 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +99 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +28 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +77 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +52 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +58 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +60 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +47 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +173 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +50 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +37 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +1373 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +36 -0
- data/ext/{ggml → sources/ggml}/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -35
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
- data/ext/{ggml → sources/ggml}/src/ggml.c +676 -1820
- data/ext/sources/ggml/src/gguf.cpp +1330 -0
- data/ext/{include → sources/include}/whisper.h +68 -2
- data/ext/sources/src/CMakeLists.txt +143 -0
- data/ext/{src → sources/src}/coreml/whisper-decoder-impl.h +27 -15
- data/ext/{src → sources/src}/coreml/whisper-decoder-impl.m +35 -10
- data/ext/{src → sources/src}/coreml/whisper-encoder-impl.h +21 -9
- data/ext/{src → sources/src}/coreml/whisper-encoder-impl.m +28 -3
- data/ext/sources/src/coreml/whisper-encoder.mm +73 -0
- data/ext/sources/src/whisper-arch.h +197 -0
- data/ext/{src → sources/src}/whisper.cpp +1905 -374
- data/ext/sources/tests/CMakeLists.txt +105 -0
- data/ext/sources/tests/earnings21/eval.mk +58 -0
- data/ext/sources/tests/earnings21/eval.py +68 -0
- data/ext/sources/tests/earnings21/normalizers/__init__.py +2 -0
- data/ext/sources/tests/earnings21/normalizers/basic.py +80 -0
- data/ext/sources/tests/earnings21/normalizers/english.json +1741 -0
- data/ext/sources/tests/earnings21/normalizers/english.py +550 -0
- data/ext/sources/tests/earnings21/requirements.txt +6 -0
- data/ext/sources/tests/en-0-ref.txt +1 -0
- data/ext/sources/tests/en-1-ref.txt +1 -0
- data/ext/sources/tests/en-2-ref.txt +1 -0
- data/ext/sources/tests/es-0-ref.txt +1 -0
- data/ext/sources/tests/librispeech/eval.mk +39 -0
- data/ext/sources/tests/librispeech/eval.py +47 -0
- data/ext/sources/tests/librispeech/normalizers/__init__.py +2 -0
- data/ext/sources/tests/librispeech/normalizers/basic.py +80 -0
- data/ext/sources/tests/librispeech/normalizers/english.json +1741 -0
- data/ext/sources/tests/librispeech/normalizers/english.py +550 -0
- data/ext/sources/tests/librispeech/requirements.txt +6 -0
- data/ext/sources/tests/run-tests.sh +130 -0
- data/ext/sources/tests/test-c.c +3 -0
- data/ext/sources/tests/test-vad-full.cpp +54 -0
- data/ext/sources/tests/test-vad.cpp +83 -0
- data/ext/sources/tests/test-whisper.js +58 -0
- data/extsources.rb +33 -5
- data/lib/whisper/model/uri.rb +149 -128
- data/sig/whisper.rbs +480 -0
- data/tests/helper.rb +28 -0
- data/tests/test_callback.rb +45 -3
- data/tests/test_error.rb +2 -2
- data/tests/test_model.rb +38 -0
- data/tests/test_package.rb +18 -3
- data/tests/test_params.rb +145 -8
- data/tests/test_segment.rb +10 -19
- data/tests/test_vad.rb +19 -0
- data/tests/test_vad_params.rb +103 -0
- data/tests/test_whisper.rb +37 -37
- data/whispercpp.gemspec +5 -4
- metadata +766 -111
- data/ext/cpu.mk +0 -9
- data/ext/examples/dr_wav.h +0 -8815
- data/ext/ggml/src/ggml-cann/aclnn_ops.h +0 -592
- data/ext/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +0 -4262
- data/ext/ggml/src/ggml-cpu/ggml-cpu.c +0 -14123
- data/ext/ggml/src/ggml-cpu/llamafile/sgemm.cpp +0 -1884
- data/ext/ggml/src/ggml-cpu/llamafile/sgemm.h +0 -14
- data/ext/ggml/src/ggml-metal/ggml-metal-impl.h +0 -288
- data/ext/ggml/src/ggml-sycl/element_wise.cpp +0 -1030
- data/ext/ggml/src/ggml-sycl/im2col.cpp +0 -126
- data/ext/ggml/src/ggml-sycl/rope.cpp +0 -276
- data/ext/ggml/src/ggml-sycl/wkv6.cpp +0 -141
- data/ext/metal-embed.mk +0 -17
- data/ext/metal.mk +0 -6
- data/ext/ruby_whisper.cpp +0 -1909
- data/ext/scripts/get-flags.mk +0 -38
- data/lib/whisper.rb +0 -2
- /data/ext/{ggml → sources/ggml}/include/ggml-blas.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-cann.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-cuda.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-kompute.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-opencl.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-sycl.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/common.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/ggml-amx.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-blas/ggml-blas.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/ascendc_kernels.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f16.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f32.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/common.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-aarch64.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-hbm.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-hbm.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-quants.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-traits.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-traits.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-kompute/ggml-kompute.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-quants.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-threading.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-threading.h +0 -0
- /data/ext/{src → sources/src}/coreml/whisper-encoder.h +0 -0
- /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.cpp +0 -0
- /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.h +0 -0
@@ -0,0 +1,458 @@
|
|
1
|
+
#include "norm.cuh"
|
2
|
+
#include <cstdint>
|
3
|
+
|
4
|
+
template <int block_size>
|
5
|
+
static __global__ void norm_f32(
|
6
|
+
const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
|
7
|
+
const int64_t stride_sample, const float eps) {
|
8
|
+
const int nrows = gridDim.x;
|
9
|
+
const int nchannels = gridDim.y;
|
10
|
+
|
11
|
+
const int row = blockIdx.x;
|
12
|
+
const int channel = blockIdx.y;
|
13
|
+
const int sample = blockIdx.z;
|
14
|
+
const int tid = threadIdx.x;
|
15
|
+
|
16
|
+
x += sample*stride_sample + channel*stride_channel + row*stride_row;
|
17
|
+
dst += ((sample*nchannels + channel)*nrows + row)*ncols;
|
18
|
+
|
19
|
+
float2 mean_var = make_float2(0.0f, 0.0f);
|
20
|
+
|
21
|
+
for (int col = tid; col < ncols; col += block_size) {
|
22
|
+
const float xi = x[col];
|
23
|
+
mean_var.x += xi;
|
24
|
+
mean_var.y += xi * xi;
|
25
|
+
}
|
26
|
+
|
27
|
+
// sum up partial sums
|
28
|
+
mean_var = warp_reduce_sum(mean_var);
|
29
|
+
if constexpr (block_size > WARP_SIZE) {
|
30
|
+
static_assert(block_size == 1024, "unexpected block_size");
|
31
|
+
__shared__ float2 s_sum[32];
|
32
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
33
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
34
|
+
if (lane_id == 0) {
|
35
|
+
s_sum[warp_id] = mean_var;
|
36
|
+
}
|
37
|
+
__syncthreads();
|
38
|
+
mean_var = s_sum[lane_id];
|
39
|
+
mean_var = warp_reduce_sum(mean_var);
|
40
|
+
}
|
41
|
+
|
42
|
+
const float mean = mean_var.x / ncols;
|
43
|
+
const float var = mean_var.y / ncols - mean * mean;
|
44
|
+
const float inv_std = rsqrtf(var + eps);
|
45
|
+
|
46
|
+
for (int col = tid; col < ncols; col += block_size) {
|
47
|
+
dst[col] = (x[col] - mean) * inv_std;
|
48
|
+
}
|
49
|
+
}
|
50
|
+
|
51
|
+
template <int block_size>
|
52
|
+
static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
|
53
|
+
// blockIdx.x: num_groups idx
|
54
|
+
// threadIdx.x: block_size idx
|
55
|
+
const int start = blockIdx.x*group_size + threadIdx.x;
|
56
|
+
const int end = min(blockIdx.x*group_size + group_size, ne_elements);
|
57
|
+
|
58
|
+
float tmp = 0.0f; // partial sum for thread in warp
|
59
|
+
|
60
|
+
for (int j = start; j < end; j += block_size) {
|
61
|
+
tmp += x[j];
|
62
|
+
}
|
63
|
+
|
64
|
+
tmp = warp_reduce_sum(tmp);
|
65
|
+
if constexpr (block_size > WARP_SIZE) {
|
66
|
+
static_assert(block_size == 1024, "unexpected block_size");
|
67
|
+
__shared__ float s_sum[32];
|
68
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
69
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
70
|
+
if (lane_id == 0) {
|
71
|
+
s_sum[warp_id] = tmp;
|
72
|
+
}
|
73
|
+
__syncthreads();
|
74
|
+
tmp = s_sum[lane_id];
|
75
|
+
tmp = warp_reduce_sum(tmp);
|
76
|
+
}
|
77
|
+
|
78
|
+
const float mean = tmp / group_size;
|
79
|
+
tmp = 0.0f;
|
80
|
+
|
81
|
+
for (int j = start; j < end; j += block_size) {
|
82
|
+
const float xi = x[j] - mean;
|
83
|
+
dst[j] = xi;
|
84
|
+
tmp += xi * xi;
|
85
|
+
}
|
86
|
+
|
87
|
+
tmp = warp_reduce_sum(tmp);
|
88
|
+
if (block_size > WARP_SIZE) {
|
89
|
+
__shared__ float s_sum[32];
|
90
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
91
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
92
|
+
if (lane_id == 0) {
|
93
|
+
s_sum[warp_id] = tmp;
|
94
|
+
}
|
95
|
+
__syncthreads();
|
96
|
+
tmp = s_sum[lane_id];
|
97
|
+
tmp = warp_reduce_sum(tmp);
|
98
|
+
}
|
99
|
+
|
100
|
+
const float variance = tmp / group_size;
|
101
|
+
const float scale = rsqrtf(variance + eps);
|
102
|
+
for (int j = start; j < end; j += block_size) {
|
103
|
+
dst[j] *= scale;
|
104
|
+
}
|
105
|
+
}
|
106
|
+
|
107
|
+
template <int block_size>
|
108
|
+
static __global__ void rms_norm_f32(
|
109
|
+
const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
|
110
|
+
const int64_t stride_sample, const float eps) {
|
111
|
+
const int nrows = gridDim.x;
|
112
|
+
const int nchannels = gridDim.y;
|
113
|
+
|
114
|
+
const int row = blockIdx.x;
|
115
|
+
const int channel = blockIdx.y;
|
116
|
+
const int sample = blockIdx.z;
|
117
|
+
const int tid = threadIdx.x;
|
118
|
+
|
119
|
+
x += sample*stride_sample + channel*stride_channel + row*stride_row;
|
120
|
+
dst += ((sample*nchannels + channel)*nrows + row)*ncols;
|
121
|
+
|
122
|
+
float tmp = 0.0f; // partial sum for thread in warp
|
123
|
+
|
124
|
+
for (int col = tid; col < ncols; col += block_size) {
|
125
|
+
const float xi = x[col];
|
126
|
+
tmp += xi * xi;
|
127
|
+
}
|
128
|
+
|
129
|
+
// sum up partial sums
|
130
|
+
tmp = warp_reduce_sum(tmp);
|
131
|
+
if constexpr (block_size > WARP_SIZE) {
|
132
|
+
static_assert(block_size == 1024, "unexpected block_size");
|
133
|
+
__shared__ float s_sum[32];
|
134
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
135
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
136
|
+
if (lane_id == 0) {
|
137
|
+
s_sum[warp_id] = tmp;
|
138
|
+
}
|
139
|
+
__syncthreads();
|
140
|
+
tmp = s_sum[lane_id];
|
141
|
+
tmp = warp_reduce_sum(tmp);
|
142
|
+
}
|
143
|
+
|
144
|
+
const float mean = tmp / ncols;
|
145
|
+
const float scale = rsqrtf(mean + eps);
|
146
|
+
|
147
|
+
for (int col = tid; col < ncols; col += block_size) {
|
148
|
+
dst[col] = scale * x[col];
|
149
|
+
}
|
150
|
+
}
|
151
|
+
|
152
|
+
template <int block_size>
|
153
|
+
static __global__ void rms_norm_back_f32(
|
154
|
+
const float * grad, const float * xf, float * dst, const int ncols, const float eps) {
|
155
|
+
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
156
|
+
const int tid = threadIdx.x;
|
157
|
+
|
158
|
+
grad += int64_t(row)*ncols;
|
159
|
+
xf += int64_t(row)*ncols;
|
160
|
+
dst += int64_t(row)*ncols;
|
161
|
+
|
162
|
+
float sum_xx = 0.0f; // sum for squares of x, equivalent to forward pass
|
163
|
+
float sum_xg = 0.0f; // sum for x * gradient, needed because RMS norm mixes inputs
|
164
|
+
|
165
|
+
for (int col = tid; col < ncols; col += block_size) {
|
166
|
+
const float xfi = xf[col];
|
167
|
+
sum_xx += xfi * xfi;
|
168
|
+
sum_xg += xfi * grad[col];
|
169
|
+
}
|
170
|
+
|
171
|
+
// sum up partial sums
|
172
|
+
sum_xx = warp_reduce_sum(sum_xx);
|
173
|
+
sum_xg = warp_reduce_sum(sum_xg);
|
174
|
+
if constexpr (block_size > WARP_SIZE) {
|
175
|
+
static_assert(block_size == 1024, "unexpected block_size");
|
176
|
+
__shared__ float s_sum_xx[32];
|
177
|
+
__shared__ float s_sum_xg[32];
|
178
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
179
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
180
|
+
if (lane_id == 0) {
|
181
|
+
s_sum_xx[warp_id] = sum_xx;
|
182
|
+
s_sum_xg[warp_id] = sum_xg;
|
183
|
+
}
|
184
|
+
__syncthreads();
|
185
|
+
|
186
|
+
sum_xx = s_sum_xx[lane_id];
|
187
|
+
sum_xx = warp_reduce_sum(sum_xx);
|
188
|
+
|
189
|
+
sum_xg = s_sum_xg[lane_id];
|
190
|
+
sum_xg = warp_reduce_sum(sum_xg);
|
191
|
+
}
|
192
|
+
|
193
|
+
const float mean_eps = sum_xx / ncols + eps;
|
194
|
+
const float sum_eps = sum_xx + ncols*eps;
|
195
|
+
|
196
|
+
const float scale_grad = rsqrtf(mean_eps);
|
197
|
+
const float scale_x = -scale_grad * sum_xg/sum_eps;
|
198
|
+
|
199
|
+
for (int col = tid; col < ncols; col += block_size) {
|
200
|
+
dst[col] = scale_grad*grad[col] + scale_x*xf[col];
|
201
|
+
}
|
202
|
+
}
|
203
|
+
|
204
|
+
// template <int block_size>
|
205
|
+
// static __global__ void l2_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
|
206
|
+
// const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
207
|
+
// const int tid = threadIdx.x;
|
208
|
+
|
209
|
+
// float tmp = 0.0f; // partial sum for thread in warp
|
210
|
+
|
211
|
+
// for (int col = tid; col < ncols; col += block_size) {
|
212
|
+
// const float xi = x[row*ncols + col];
|
213
|
+
// tmp += xi * xi;
|
214
|
+
// }
|
215
|
+
|
216
|
+
// // sum up partial sums
|
217
|
+
// tmp = warp_reduce_sum(tmp);
|
218
|
+
// if (block_size > WARP_SIZE) {
|
219
|
+
// __shared__ float s_sum[32];
|
220
|
+
// int warp_id = threadIdx.x / WARP_SIZE;
|
221
|
+
// int lane_id = threadIdx.x % WARP_SIZE;
|
222
|
+
// if (lane_id == 0) {
|
223
|
+
// s_sum[warp_id] = tmp;
|
224
|
+
// }
|
225
|
+
// __syncthreads();
|
226
|
+
// tmp = s_sum[lane_id];
|
227
|
+
// tmp = warp_reduce_sum(tmp);
|
228
|
+
// }
|
229
|
+
|
230
|
+
// // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html
|
231
|
+
// const float scale = rsqrtf(fmaxf(tmp, eps * eps));
|
232
|
+
|
233
|
+
// for (int col = tid; col < ncols; col += block_size) {
|
234
|
+
// dst[row*ncols + col] = scale * x[row*ncols + col];
|
235
|
+
// }
|
236
|
+
// }
|
237
|
+
|
238
|
+
template <int block_size>
|
239
|
+
static __global__ void l2_norm_f32(
|
240
|
+
const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
|
241
|
+
const int64_t stride_sample, const float eps) {
|
242
|
+
const int nrows = gridDim.x;
|
243
|
+
const int nchannels = gridDim.y;
|
244
|
+
|
245
|
+
const int row = blockIdx.x;
|
246
|
+
const int channel = blockIdx.y;
|
247
|
+
const int sample = blockIdx.z;
|
248
|
+
const int tid = threadIdx.x;
|
249
|
+
|
250
|
+
x += sample*stride_sample + channel*stride_channel + row*stride_row;
|
251
|
+
dst += ((sample*nchannels + channel)*nrows + row)*ncols;
|
252
|
+
|
253
|
+
float tmp = 0.0f; // partial sum for thread in warp
|
254
|
+
|
255
|
+
for (int col = tid; col < ncols; col += block_size) {
|
256
|
+
const float xi = x[col];
|
257
|
+
tmp += xi * xi;
|
258
|
+
}
|
259
|
+
|
260
|
+
// sum up partial sums
|
261
|
+
tmp = warp_reduce_sum(tmp);
|
262
|
+
if constexpr (block_size > WARP_SIZE) {
|
263
|
+
static_assert(block_size == 1024, "unexpected block_size");
|
264
|
+
__shared__ float s_sum[32];
|
265
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
266
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
267
|
+
if (lane_id == 0) {
|
268
|
+
s_sum[warp_id] = tmp;
|
269
|
+
}
|
270
|
+
__syncthreads();
|
271
|
+
tmp = s_sum[lane_id];
|
272
|
+
tmp = warp_reduce_sum(tmp);
|
273
|
+
}
|
274
|
+
|
275
|
+
// from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html
|
276
|
+
const float scale = rsqrtf(fmaxf(tmp, eps * eps));
|
277
|
+
|
278
|
+
for (int col = tid; col < ncols; col += block_size) {
|
279
|
+
dst[col] = scale * x[col];
|
280
|
+
}
|
281
|
+
}
|
282
|
+
|
283
|
+
static void norm_f32_cuda(
|
284
|
+
const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
|
285
|
+
const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
|
286
|
+
const dim3 blocks_num(nrows, nchannels, nsamples);
|
287
|
+
if (ncols < 1024) {
|
288
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
289
|
+
norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
290
|
+
} else {
|
291
|
+
const dim3 block_dims(1024, 1, 1);
|
292
|
+
norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
293
|
+
}
|
294
|
+
}
|
295
|
+
|
296
|
+
static void group_norm_f32_cuda(
|
297
|
+
const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
|
298
|
+
if (group_size < 1024) {
|
299
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
300
|
+
group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
|
301
|
+
} else {
|
302
|
+
const dim3 block_dims(1024, 1, 1);
|
303
|
+
group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
|
304
|
+
}
|
305
|
+
}
|
306
|
+
|
307
|
+
static void rms_norm_f32_cuda(
|
308
|
+
const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
|
309
|
+
const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
|
310
|
+
const dim3 blocks_num(nrows, nchannels, nsamples);
|
311
|
+
if (ncols < 1024) {
|
312
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
313
|
+
rms_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
314
|
+
} else {
|
315
|
+
const dim3 block_dims(1024, 1, 1);
|
316
|
+
rms_norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
317
|
+
}
|
318
|
+
}
|
319
|
+
|
320
|
+
static void rms_norm_back_f32_cuda(const float * grad, const float * xf, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
|
321
|
+
if (ncols < 1024) {
|
322
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
323
|
+
rms_norm_back_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(grad, xf, dst, ncols, eps);
|
324
|
+
} else {
|
325
|
+
const dim3 block_dims(1024, 1, 1);
|
326
|
+
rms_norm_back_f32<1024><<<nrows, block_dims, 0, stream>>>(grad, xf, dst, ncols, eps);
|
327
|
+
}
|
328
|
+
}
|
329
|
+
|
330
|
+
static void l2_norm_f32_cuda(
|
331
|
+
const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
|
332
|
+
const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
|
333
|
+
const dim3 blocks_num(nrows, nchannels, nsamples);
|
334
|
+
if (ncols < 1024) {
|
335
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
336
|
+
l2_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
337
|
+
} else {
|
338
|
+
const dim3 block_dims(1024, 1, 1);
|
339
|
+
l2_norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
|
340
|
+
}
|
341
|
+
}
|
342
|
+
|
343
|
+
void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
344
|
+
const ggml_tensor * src0 = dst->src[0];
|
345
|
+
const float * src0_d = (const float *) src0->data;
|
346
|
+
float * dst_d = (float *) dst->data;
|
347
|
+
cudaStream_t stream = ctx.stream();
|
348
|
+
|
349
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
350
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
351
|
+
|
352
|
+
GGML_TENSOR_UNARY_OP_LOCALS;
|
353
|
+
|
354
|
+
float eps;
|
355
|
+
memcpy(&eps, dst->op_params, sizeof(float));
|
356
|
+
GGML_ASSERT(eps >= 0.0f);
|
357
|
+
|
358
|
+
const size_t ts0 = ggml_type_size(src0->type);
|
359
|
+
GGML_ASSERT(nb00 == ts0);
|
360
|
+
const int64_t s01 = nb01 / ts0;
|
361
|
+
const int64_t s02 = nb02 / ts0;
|
362
|
+
const int64_t s03 = nb03 / ts0;
|
363
|
+
|
364
|
+
norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
|
365
|
+
}
|
366
|
+
|
367
|
+
void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
368
|
+
const ggml_tensor * src0 = dst->src[0];
|
369
|
+
const float * src0_d = (const float *)src0->data;
|
370
|
+
float * dst_d = (float *)dst->data;
|
371
|
+
cudaStream_t stream = ctx.stream();
|
372
|
+
|
373
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
374
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
375
|
+
|
376
|
+
int num_groups = dst->op_params[0];
|
377
|
+
|
378
|
+
float eps;
|
379
|
+
memcpy(&eps, dst->op_params + 1, sizeof(float));
|
380
|
+
GGML_ASSERT(eps >= 0.0f);
|
381
|
+
|
382
|
+
int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
|
383
|
+
group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
|
384
|
+
}
|
385
|
+
|
386
|
+
void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
387
|
+
const ggml_tensor * src0 = dst->src[0];
|
388
|
+
const float * src0_d = (const float *) src0->data;
|
389
|
+
float * dst_d = (float *) dst->data;
|
390
|
+
cudaStream_t stream = ctx.stream();
|
391
|
+
|
392
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
393
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
394
|
+
|
395
|
+
GGML_TENSOR_UNARY_OP_LOCALS;
|
396
|
+
|
397
|
+
float eps;
|
398
|
+
memcpy(&eps, dst->op_params, sizeof(float));
|
399
|
+
GGML_ASSERT(eps >= 0.0f);
|
400
|
+
|
401
|
+
const size_t ts0 = ggml_type_size(src0->type);
|
402
|
+
GGML_ASSERT(nb00 == ts0);
|
403
|
+
const int64_t s01 = nb01 / ts0;
|
404
|
+
const int64_t s02 = nb02 / ts0;
|
405
|
+
const int64_t s03 = nb03 / ts0;
|
406
|
+
|
407
|
+
rms_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
|
408
|
+
}
|
409
|
+
|
410
|
+
void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
411
|
+
const ggml_tensor * grad = dst->src[0]; // gradients
|
412
|
+
const ggml_tensor * src0f = dst->src[1]; // src0 from forward pass
|
413
|
+
|
414
|
+
const float * grad_d = (const float *) grad->data;
|
415
|
+
const float * src0f_d = (const float *) src0f->data;
|
416
|
+
float * dst_d = (float *) dst->data;
|
417
|
+
|
418
|
+
cudaStream_t stream = ctx.stream();
|
419
|
+
|
420
|
+
GGML_ASSERT(ggml_is_contiguous(grad));
|
421
|
+
|
422
|
+
GGML_ASSERT( grad->type == GGML_TYPE_F32);
|
423
|
+
GGML_ASSERT(src0f->type == GGML_TYPE_F32);
|
424
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
425
|
+
|
426
|
+
const int64_t ne00 = src0f->ne[0];
|
427
|
+
const int64_t nrows = ggml_nrows(src0f);
|
428
|
+
|
429
|
+
float eps;
|
430
|
+
memcpy(&eps, dst->op_params, sizeof(float));
|
431
|
+
GGML_ASSERT(eps >= 0.0f);
|
432
|
+
|
433
|
+
rms_norm_back_f32_cuda(grad_d, src0f_d, dst_d, ne00, nrows, eps, stream);
|
434
|
+
}
|
435
|
+
|
436
|
+
void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
437
|
+
const ggml_tensor * src0 = dst->src[0];
|
438
|
+
const float * src0_d = (const float *) src0->data;
|
439
|
+
float * dst_d = (float *) dst->data;
|
440
|
+
cudaStream_t stream = ctx.stream();
|
441
|
+
|
442
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
443
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
444
|
+
|
445
|
+
GGML_TENSOR_UNARY_OP_LOCALS;
|
446
|
+
|
447
|
+
float eps;
|
448
|
+
memcpy(&eps, dst->op_params, sizeof(float));
|
449
|
+
GGML_ASSERT(eps >= 0.0f);
|
450
|
+
|
451
|
+
const size_t ts0 = ggml_type_size(src0->type);
|
452
|
+
GGML_ASSERT(nb00 == ts0);
|
453
|
+
const int64_t s01 = nb01 / ts0;
|
454
|
+
const int64_t s02 = nb02 / ts0;
|
455
|
+
const int64_t s03 = nb03 / ts0;
|
456
|
+
|
457
|
+
l2_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
|
458
|
+
}
|
@@ -0,0 +1,11 @@
|
|
1
|
+
#include "common.cuh"
|
2
|
+
|
3
|
+
void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
4
|
+
|
5
|
+
void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
6
|
+
|
7
|
+
void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
8
|
+
|
9
|
+
void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
10
|
+
|
11
|
+
void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
@@ -0,0 +1,78 @@
|
|
1
|
+
#include "ggml-impl.h"
|
2
|
+
#include "opt-step-adamw.cuh"
|
3
|
+
|
4
|
+
#include <cstdint>
|
5
|
+
|
6
|
+
static __global__ void opt_step_adamw_f32(
|
7
|
+
float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
|
8
|
+
const float * __restrict__ pars, const int64_t k) {
|
9
|
+
|
10
|
+
const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
|
11
|
+
|
12
|
+
if (i >= k) {
|
13
|
+
return;
|
14
|
+
}
|
15
|
+
|
16
|
+
const float alpha = pars[0];
|
17
|
+
const float beta1 = pars[1];
|
18
|
+
const float beta2 = pars[2];
|
19
|
+
const float eps = pars[3];
|
20
|
+
const float wd = pars[4];
|
21
|
+
const float beta1h = pars[5];
|
22
|
+
const float beta2h = pars[6];
|
23
|
+
|
24
|
+
const float gi = g[i];
|
25
|
+
const float gmi = g_m[i]*beta1 + gi*(1.0f - beta1);
|
26
|
+
const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
|
27
|
+
|
28
|
+
g_m[i] = gmi;
|
29
|
+
g_v[i] = gvi;
|
30
|
+
|
31
|
+
const float mh = gmi*beta1h;
|
32
|
+
const float vh = sqrtf(gvi*beta2h) + eps;
|
33
|
+
|
34
|
+
x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
|
35
|
+
}
|
36
|
+
|
37
|
+
static void opt_step_adamw_f32_cuda(
|
38
|
+
float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
|
39
|
+
|
40
|
+
const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
|
41
|
+
const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
|
42
|
+
opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
|
43
|
+
}
|
44
|
+
|
45
|
+
void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
46
|
+
const ggml_tensor * src0 = dst->src[0];
|
47
|
+
const ggml_tensor * src0_grad = dst->src[1];
|
48
|
+
const ggml_tensor * src0_grad_m = dst->src[2];
|
49
|
+
const ggml_tensor * src0_grad_v = dst->src[3];
|
50
|
+
const ggml_tensor * adamw_params = dst->src[4];
|
51
|
+
|
52
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
53
|
+
GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
|
54
|
+
GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
|
55
|
+
GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
|
56
|
+
GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
|
57
|
+
GGML_ASSERT(ggml_is_contiguous(src0));
|
58
|
+
GGML_ASSERT(ggml_is_contiguous(src0_grad));
|
59
|
+
GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
|
60
|
+
GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
|
61
|
+
GGML_ASSERT(ggml_is_contiguous(adamw_params));
|
62
|
+
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
|
63
|
+
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
|
64
|
+
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
|
65
|
+
GGML_ASSERT(ggml_nelements(adamw_params) == 7);
|
66
|
+
|
67
|
+
float * src0_d = (float *) src0->data;
|
68
|
+
const float * src0_grad_d = (const float *) src0_grad->data;
|
69
|
+
float * src0_grad_m_d = (float *) src0_grad_m->data;
|
70
|
+
float * src0_grad_v_d = (float *) src0_grad_v->data;
|
71
|
+
const float * adamw_params_d = (const float *) adamw_params->data;
|
72
|
+
|
73
|
+
cudaStream_t stream = ctx.stream();
|
74
|
+
|
75
|
+
const int64_t ne = ggml_nelements(src0);
|
76
|
+
|
77
|
+
opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
|
78
|
+
}
|
@@ -0,0 +1,68 @@
|
|
1
|
+
#include "out-prod.cuh"
|
2
|
+
|
3
|
+
#include <cstdint>
|
4
|
+
|
5
|
+
void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
6
|
+
const ggml_tensor * src0 = dst->src[0];
|
7
|
+
const ggml_tensor * src1 = dst->src[1];
|
8
|
+
|
9
|
+
GGML_TENSOR_BINARY_OP_LOCALS
|
10
|
+
|
11
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
12
|
+
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
13
|
+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
14
|
+
|
15
|
+
GGML_ASSERT(ne01 == ne11);
|
16
|
+
GGML_ASSERT(ne0 == ne00);
|
17
|
+
GGML_ASSERT(ne1 == ne10);
|
18
|
+
|
19
|
+
GGML_ASSERT(ne2 % src0->ne[2] == 0);
|
20
|
+
GGML_ASSERT(ne3 % src0->ne[3] == 0);
|
21
|
+
|
22
|
+
GGML_ASSERT(ne2 == src1->ne[2]);
|
23
|
+
GGML_ASSERT(ne3 == src1->ne[3]);
|
24
|
+
|
25
|
+
const float * src0_d = (const float *) src0->data;
|
26
|
+
const float * src1_d = (const float *) src1->data;
|
27
|
+
float * dst_d = (float *) dst->data;
|
28
|
+
|
29
|
+
cudaStream_t stream = ctx.stream();
|
30
|
+
cublasHandle_t handle = ctx.cublas_handle();
|
31
|
+
|
32
|
+
const float alpha = 1.0f;
|
33
|
+
const float beta = 0.0f;
|
34
|
+
|
35
|
+
CUBLAS_CHECK(cublasSetStream(handle, stream));
|
36
|
+
|
37
|
+
const int64_t lda = nb01 / sizeof(float);
|
38
|
+
const int64_t ldc = nb1 / sizeof(float);
|
39
|
+
|
40
|
+
const bool src1_T = ggml_is_transposed(src1);
|
41
|
+
const cublasOperation_t src1_cublas_op = src1_T ? CUBLAS_OP_N : CUBLAS_OP_T;
|
42
|
+
const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float);
|
43
|
+
GGML_ASSERT( (src1_T ? nb11 : nb10) == sizeof(float));
|
44
|
+
|
45
|
+
// data strides in dimensions 2/3
|
46
|
+
const size_t s02 = nb02 / sizeof(float);
|
47
|
+
const size_t s03 = nb03 / sizeof(float);
|
48
|
+
const size_t s12 = nb12 / sizeof(float);
|
49
|
+
const size_t s13 = nb13 / sizeof(float);
|
50
|
+
const size_t s2 = nb2 / sizeof(float);
|
51
|
+
const size_t s3 = nb3 / sizeof(float);
|
52
|
+
|
53
|
+
// dps == dst per src0, used for group query attention
|
54
|
+
const int64_t dps2 = ne2 / ne02;
|
55
|
+
const int64_t dps3 = ne3 / ne03;
|
56
|
+
|
57
|
+
// TODO batched matrix multiplication
|
58
|
+
for (int64_t i3 = 0; i3 < ne3; ++i3) {
|
59
|
+
for (int64_t i2 = 0; i2 < ne2; ++i2) {
|
60
|
+
CUBLAS_CHECK(
|
61
|
+
cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
|
62
|
+
ne0, ne1, ne01,
|
63
|
+
&alpha, src0_d + (i3/dps3)*s03 + (i2/dps2)*s02, lda,
|
64
|
+
src1_d + i3 *s13 + i2 *s12, ldb,
|
65
|
+
&beta, dst_d + i3 *s3 + i2 *s2, ldc));
|
66
|
+
}
|
67
|
+
}
|
68
|
+
}
|
@@ -0,0 +1,49 @@
|
|
1
|
+
#include "pad.cuh"
|
2
|
+
|
3
|
+
static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
|
4
|
+
// blockIdx.z: idx of ne2*ne3, aka ne02*ne03
|
5
|
+
// blockIdx.y: idx of ne1
|
6
|
+
// blockIDx.x: idx of ne0 / BLOCK_SIZE
|
7
|
+
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
|
8
|
+
if (nidx >= ne0) {
|
9
|
+
return;
|
10
|
+
}
|
11
|
+
|
12
|
+
// operation
|
13
|
+
int offset_dst =
|
14
|
+
nidx +
|
15
|
+
blockIdx.y * ne0 +
|
16
|
+
blockIdx.z * ne0 * gridDim.y;
|
17
|
+
if (nidx < ne00 && blockIdx.y < (unsigned)ne01 && blockIdx.z < (unsigned)(ne02*ne03)) {
|
18
|
+
int offset_src =
|
19
|
+
nidx +
|
20
|
+
blockIdx.y * ne00 +
|
21
|
+
blockIdx.z * ne00 * ne01;
|
22
|
+
dst[offset_dst] = x[offset_src];
|
23
|
+
} else {
|
24
|
+
dst[offset_dst] = 0.0f;
|
25
|
+
}
|
26
|
+
}
|
27
|
+
|
28
|
+
static void pad_f32_cuda(const float * x, float * dst,
|
29
|
+
const int ne00, const int ne01, const int ne02, const int ne03,
|
30
|
+
const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
|
31
|
+
int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
|
32
|
+
dim3 gridDim(num_blocks, ne1, ne2*ne3);
|
33
|
+
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
|
34
|
+
}
|
35
|
+
|
36
|
+
void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
37
|
+
const ggml_tensor * src0 = dst->src[0];
|
38
|
+
const float * src0_d = (const float *)src0->data;
|
39
|
+
float * dst_d = (float *)dst->data;
|
40
|
+
cudaStream_t stream = ctx.stream();
|
41
|
+
|
42
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
43
|
+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
44
|
+
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
|
45
|
+
|
46
|
+
pad_f32_cuda(src0_d, dst_d,
|
47
|
+
src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
|
48
|
+
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
|
49
|
+
}
|