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 +1,283 @@
|
|
1
|
+
#include "common.cuh"
|
2
|
+
#include "ggml.h"
|
3
|
+
#include "softmax.cuh"
|
4
|
+
#include <cstdint>
|
5
|
+
|
6
|
+
template <typename T>
|
7
|
+
static __device__ __forceinline__ float t2f32(T val) {
|
8
|
+
return (float) val;
|
9
|
+
}
|
10
|
+
|
11
|
+
template <>
|
12
|
+
__device__ float __forceinline__ t2f32<half>(half val) {
|
13
|
+
return __half2float(val);
|
14
|
+
}
|
15
|
+
|
16
|
+
// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
|
17
|
+
// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
|
18
|
+
#ifdef __clang__
|
19
|
+
#pragma clang diagnostic push
|
20
|
+
#pragma clang diagnostic ignored "-Wpass-failed"
|
21
|
+
#endif // __clang__
|
22
|
+
template <bool use_shared, int ncols_template, int block_size_template, typename T>
|
23
|
+
static __global__ void soft_max_f32(
|
24
|
+
const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
|
25
|
+
const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
|
26
|
+
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
|
27
|
+
|
28
|
+
const int tid = threadIdx.x;
|
29
|
+
const int rowx = blockIdx.x;
|
30
|
+
const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
|
31
|
+
|
32
|
+
x += int64_t(rowx)*ncols;
|
33
|
+
mask += int64_t(rowy)*ncols * (mask != nullptr);
|
34
|
+
dst += int64_t(rowx)*ncols;
|
35
|
+
|
36
|
+
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
|
37
|
+
|
38
|
+
const int warp_id = threadIdx.x / WARP_SIZE;
|
39
|
+
const int lane_id = threadIdx.x % WARP_SIZE;
|
40
|
+
|
41
|
+
const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
|
42
|
+
|
43
|
+
extern __shared__ float data_soft_max_f32[];
|
44
|
+
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
|
45
|
+
// shared memory buffer to cache values between iterations:
|
46
|
+
float * vals = use_shared ? buf_iw + WARP_SIZE : dst;
|
47
|
+
|
48
|
+
float max_val = -INFINITY;
|
49
|
+
|
50
|
+
#pragma unroll
|
51
|
+
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
52
|
+
const int col = col0 + tid;
|
53
|
+
|
54
|
+
if (ncols_template == 0 && col >= ncols) {
|
55
|
+
break;
|
56
|
+
}
|
57
|
+
|
58
|
+
const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
|
59
|
+
|
60
|
+
vals[col] = val;
|
61
|
+
max_val = max(max_val, val);
|
62
|
+
}
|
63
|
+
|
64
|
+
// find the max value in the block
|
65
|
+
max_val = warp_reduce_max(max_val);
|
66
|
+
if (block_size > WARP_SIZE) {
|
67
|
+
if (warp_id == 0) {
|
68
|
+
buf_iw[lane_id] = -INFINITY;
|
69
|
+
}
|
70
|
+
__syncthreads();
|
71
|
+
|
72
|
+
if (lane_id == 0) {
|
73
|
+
buf_iw[warp_id] = max_val;
|
74
|
+
}
|
75
|
+
__syncthreads();
|
76
|
+
|
77
|
+
max_val = buf_iw[lane_id];
|
78
|
+
max_val = warp_reduce_max(max_val);
|
79
|
+
}
|
80
|
+
|
81
|
+
float tmp = 0.0f; // partial sum
|
82
|
+
|
83
|
+
#pragma unroll
|
84
|
+
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
85
|
+
const int col = col0 + tid;
|
86
|
+
|
87
|
+
if (ncols_template == 0 && col >= ncols) {
|
88
|
+
break;
|
89
|
+
}
|
90
|
+
|
91
|
+
const float val = expf(vals[col] - max_val);
|
92
|
+
tmp += val;
|
93
|
+
vals[col] = val;
|
94
|
+
}
|
95
|
+
|
96
|
+
// find the sum of exps in the block
|
97
|
+
tmp = warp_reduce_sum(tmp);
|
98
|
+
if (block_size > WARP_SIZE) {
|
99
|
+
__syncthreads();
|
100
|
+
if (warp_id == 0) {
|
101
|
+
buf_iw[lane_id] = 0.0f;
|
102
|
+
}
|
103
|
+
__syncthreads();
|
104
|
+
|
105
|
+
if (lane_id == 0) {
|
106
|
+
buf_iw[warp_id] = tmp;
|
107
|
+
}
|
108
|
+
__syncthreads();
|
109
|
+
|
110
|
+
tmp = buf_iw[lane_id];
|
111
|
+
tmp = warp_reduce_sum(tmp);
|
112
|
+
}
|
113
|
+
|
114
|
+
const float inv_sum = 1.0f / tmp;
|
115
|
+
|
116
|
+
#pragma unroll
|
117
|
+
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
118
|
+
const int col = col0 + tid;
|
119
|
+
|
120
|
+
if (ncols_template == 0 && col >= ncols) {
|
121
|
+
return;
|
122
|
+
}
|
123
|
+
|
124
|
+
dst[col] = vals[col] * inv_sum;
|
125
|
+
}
|
126
|
+
}
|
127
|
+
#ifdef __clang__
|
128
|
+
#pragma clang diagnostic pop
|
129
|
+
#endif // __clang__
|
130
|
+
|
131
|
+
static __global__ void soft_max_back_f32(
|
132
|
+
const float * grad, const float * dstf, float * dst, const int ncols, const float scale) {
|
133
|
+
const int tid = threadIdx.x;
|
134
|
+
const int rowx = blockIdx.x;
|
135
|
+
|
136
|
+
grad += int64_t(rowx)*ncols;
|
137
|
+
dstf += int64_t(rowx)*ncols;
|
138
|
+
dst += int64_t(rowx)*ncols;
|
139
|
+
|
140
|
+
float dgf_dot = 0.0f; // dot product of dst from forward pass and gradients
|
141
|
+
|
142
|
+
for (int col = tid; col < ncols; col += WARP_SIZE) {
|
143
|
+
dgf_dot += dstf[col]*grad[col];
|
144
|
+
}
|
145
|
+
|
146
|
+
dgf_dot = warp_reduce_sum(dgf_dot);
|
147
|
+
|
148
|
+
for (int col = tid; col < ncols; col += WARP_SIZE) {
|
149
|
+
dst[col] = scale * (grad[col] - dgf_dot) * dstf[col];
|
150
|
+
}
|
151
|
+
}
|
152
|
+
|
153
|
+
template<typename T>
|
154
|
+
static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
|
155
|
+
int nth = WARP_SIZE;
|
156
|
+
while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
|
157
|
+
const dim3 block_dims(nth, 1, 1);
|
158
|
+
const dim3 block_nums(nrows_x, 1, 1);
|
159
|
+
const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
|
160
|
+
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
|
161
|
+
|
162
|
+
const uint32_t n_head = nrows_x/nrows_y;
|
163
|
+
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
|
164
|
+
|
165
|
+
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
|
166
|
+
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
|
167
|
+
|
168
|
+
// FIXME: this limit could be raised by ~2-4x on Ampere or newer
|
169
|
+
if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
|
170
|
+
switch (ncols_x) {
|
171
|
+
case 32:
|
172
|
+
soft_max_f32<true, 32, 32><<<block_nums, block_dims, nbytes_shared, stream>>>
|
173
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
174
|
+
break;
|
175
|
+
case 64:
|
176
|
+
soft_max_f32<true, 64, 64><<<block_nums, block_dims, nbytes_shared, stream>>>
|
177
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
178
|
+
break;
|
179
|
+
case 128:
|
180
|
+
soft_max_f32<true, 128, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
|
181
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
182
|
+
break;
|
183
|
+
case 256:
|
184
|
+
soft_max_f32<true, 256, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
|
185
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
186
|
+
break;
|
187
|
+
case 512:
|
188
|
+
soft_max_f32<true, 512, 512><<<block_nums, block_dims, nbytes_shared, stream>>>
|
189
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
190
|
+
break;
|
191
|
+
case 1024:
|
192
|
+
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
|
193
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
194
|
+
break;
|
195
|
+
case 2048:
|
196
|
+
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
|
197
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
198
|
+
break;
|
199
|
+
case 4096:
|
200
|
+
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
|
201
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
202
|
+
break;
|
203
|
+
default:
|
204
|
+
soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>
|
205
|
+
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
206
|
+
break;
|
207
|
+
}
|
208
|
+
} else {
|
209
|
+
const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
|
210
|
+
soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
211
|
+
}
|
212
|
+
}
|
213
|
+
|
214
|
+
static void soft_max_back_f32_cuda(
|
215
|
+
const float * grad, const float * dstf, float * dst,
|
216
|
+
const int ncols, const int nrows, const float scale, cudaStream_t stream) {
|
217
|
+
const dim3 block_dims(WARP_SIZE, 1, 1);
|
218
|
+
const dim3 block_nums(nrows, 1, 1);
|
219
|
+
|
220
|
+
soft_max_back_f32<<<block_nums, block_dims, 0, stream>>>(grad, dstf, dst, ncols, scale);
|
221
|
+
}
|
222
|
+
|
223
|
+
void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
224
|
+
const ggml_tensor * src0 = dst->src[0];
|
225
|
+
const ggml_tensor * src1 = dst->src[1];
|
226
|
+
|
227
|
+
const float * src0_d = (const float *) src0->data;
|
228
|
+
const void * src1_d = src1 ? (const void *) src1->data : nullptr;
|
229
|
+
float * dst_d = (float *) dst->data;
|
230
|
+
|
231
|
+
cudaStream_t stream = ctx.stream();
|
232
|
+
|
233
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
234
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
235
|
+
|
236
|
+
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
|
237
|
+
|
238
|
+
const int64_t ne00 = src0->ne[0];
|
239
|
+
const int64_t nrows_x = ggml_nrows(src0);
|
240
|
+
const int64_t nrows_y = src0->ne[1];
|
241
|
+
|
242
|
+
float scale = 1.0f;
|
243
|
+
float max_bias = 0.0f;
|
244
|
+
|
245
|
+
memcpy(&scale, (const float *) dst->op_params + 0, sizeof(float));
|
246
|
+
memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
|
247
|
+
|
248
|
+
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
|
249
|
+
|
250
|
+
if (use_f16) {
|
251
|
+
soft_max_f32_cuda(src0_d, (const half *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
|
252
|
+
} else {
|
253
|
+
soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
|
254
|
+
}
|
255
|
+
}
|
256
|
+
|
257
|
+
void ggml_cuda_op_soft_max_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
258
|
+
const ggml_tensor * src0 = dst->src[0]; // grad
|
259
|
+
const ggml_tensor * src1 = dst->src[1]; // forward pass output
|
260
|
+
|
261
|
+
const float * src0_d = (const float *) src0->data;
|
262
|
+
const float * src1_d = (const float *) src1->data;
|
263
|
+
float * dst_d = (float *) dst->data;
|
264
|
+
|
265
|
+
cudaStream_t stream = ctx.stream();
|
266
|
+
|
267
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
268
|
+
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
269
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
270
|
+
|
271
|
+
const int64_t ncols = src0->ne[0];
|
272
|
+
const int64_t nrows = ggml_nrows(src0);
|
273
|
+
|
274
|
+
float scale = 1.0f;
|
275
|
+
float max_bias = 0.0f;
|
276
|
+
|
277
|
+
memcpy(&scale, (const float *) dst->op_params + 0, sizeof(float));
|
278
|
+
memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
|
279
|
+
|
280
|
+
GGML_ASSERT(max_bias == 0.0f);
|
281
|
+
|
282
|
+
soft_max_back_f32_cuda(src0_d, src1_d, dst_d, ncols, nrows, scale, stream);
|
283
|
+
}
|
@@ -0,0 +1,148 @@
|
|
1
|
+
#include "ssm-conv.cuh"
|
2
|
+
|
3
|
+
template <size_t split_d_inner, size_t d_conv>
|
4
|
+
static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
|
5
|
+
const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
|
6
|
+
float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
|
7
|
+
const int64_t n_t) {
|
8
|
+
GGML_UNUSED(src0_nb0);
|
9
|
+
const int tid = threadIdx.x;
|
10
|
+
const int bidx = blockIdx.x;
|
11
|
+
const int bidy = blockIdx.y;
|
12
|
+
|
13
|
+
const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1);
|
14
|
+
const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1);
|
15
|
+
float * y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0);
|
16
|
+
|
17
|
+
const int stride_x = src0_nb1 / sizeof(float);
|
18
|
+
const int stride_w = src1_nb1 / sizeof(float);
|
19
|
+
const int stride_y = dst_nb1 / sizeof(float);
|
20
|
+
|
21
|
+
float x[d_conv] = { 0.0f };
|
22
|
+
float w[d_conv] = { 0.0f };
|
23
|
+
|
24
|
+
#pragma unroll
|
25
|
+
for (size_t j = 0; j < d_conv; j++) {
|
26
|
+
w[j] = w_block[tid * stride_w + j];
|
27
|
+
}
|
28
|
+
|
29
|
+
for (int64_t i = 0; i < n_t; i++) {
|
30
|
+
float sumf = 0.0f;
|
31
|
+
|
32
|
+
if (i == 0) {
|
33
|
+
for (size_t j = 0; j < d_conv; j++) {
|
34
|
+
x[j] = x_block[tid * stride_x + j];
|
35
|
+
}
|
36
|
+
} else {
|
37
|
+
x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
|
38
|
+
}
|
39
|
+
|
40
|
+
#pragma unroll
|
41
|
+
for (size_t j = 0; j < d_conv; j++) {
|
42
|
+
sumf += x[(i + j) % d_conv] * w[j];
|
43
|
+
}
|
44
|
+
y_block[i * stride_y + tid] = sumf;
|
45
|
+
}
|
46
|
+
}
|
47
|
+
|
48
|
+
template <size_t split_d_inner, size_t d_conv, int64_t split_n_t>
|
49
|
+
static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
|
50
|
+
const int src0_nb0, const int src0_nb1, const int src0_nb2,
|
51
|
+
const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
|
52
|
+
const int dst_nb1, const int dst_nb2, const int64_t n_t) {
|
53
|
+
const int tid = threadIdx.x;
|
54
|
+
const int bidx = blockIdx.x;
|
55
|
+
const int bidy = blockIdx.y;
|
56
|
+
const int bidz = blockIdx.z;
|
57
|
+
|
58
|
+
const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 +
|
59
|
+
bidz * split_n_t * src0_nb0);
|
60
|
+
const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1);
|
61
|
+
float * y_block =
|
62
|
+
(float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0);
|
63
|
+
|
64
|
+
const int stride_x = src0_nb1 / sizeof(float);
|
65
|
+
const int stride_w = src1_nb1 / sizeof(float);
|
66
|
+
const int stride_y = dst_nb1 / sizeof(float);
|
67
|
+
|
68
|
+
float x[d_conv] = { 0.0f };
|
69
|
+
float w[d_conv] = { 0.0f };
|
70
|
+
|
71
|
+
#pragma unroll
|
72
|
+
for (size_t j = 0; j < d_conv; j++) {
|
73
|
+
w[j] = w_block[tid * stride_w + j];
|
74
|
+
}
|
75
|
+
|
76
|
+
#pragma unroll
|
77
|
+
for (int64_t i = 0; i < split_n_t; i++) {
|
78
|
+
if (bidz * split_n_t + i < n_t) {
|
79
|
+
float sumf = 0.0f;
|
80
|
+
|
81
|
+
if (i == 0) {
|
82
|
+
for (size_t j = 0; j < d_conv; j++) {
|
83
|
+
x[j] = x_block[tid * stride_x + j];
|
84
|
+
}
|
85
|
+
} else {
|
86
|
+
x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
|
87
|
+
}
|
88
|
+
|
89
|
+
#pragma unroll
|
90
|
+
for (size_t j = 0; j < d_conv; j++) {
|
91
|
+
sumf += x[(i + j) % d_conv] * w[j];
|
92
|
+
}
|
93
|
+
y_block[i * stride_y + tid] = sumf;
|
94
|
+
}
|
95
|
+
}
|
96
|
+
}
|
97
|
+
|
98
|
+
static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
|
99
|
+
const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
|
100
|
+
const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t,
|
101
|
+
const int64_t n_s, cudaStream_t stream) {
|
102
|
+
const int threads = 128;
|
103
|
+
GGML_ASSERT(nr % threads == 0);
|
104
|
+
|
105
|
+
if (n_t <= 32) {
|
106
|
+
const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
|
107
|
+
if (nc == 4) {
|
108
|
+
ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
|
109
|
+
dst, dst_nb0, dst_nb1, dst_nb2, n_t);
|
110
|
+
} else {
|
111
|
+
GGML_ABORT("Only support kernel size = 4 now.");
|
112
|
+
}
|
113
|
+
} else {
|
114
|
+
if (nc == 4) {
|
115
|
+
const int64_t split_n_t = 32;
|
116
|
+
dim3 blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
|
117
|
+
ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
|
118
|
+
src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
|
119
|
+
} else {
|
120
|
+
GGML_ABORT("Only support kernel size = 4 right now.");
|
121
|
+
}
|
122
|
+
}
|
123
|
+
}
|
124
|
+
|
125
|
+
void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
126
|
+
const struct ggml_tensor * src0 = dst->src[0]; // conv_x
|
127
|
+
const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight
|
128
|
+
|
129
|
+
const int64_t nc = src1->ne[0]; // d_conv
|
130
|
+
const int64_t nr = src0->ne[1]; // d_inner
|
131
|
+
const int64_t n_t = dst->ne[1]; // tokens per sequence
|
132
|
+
const int64_t n_s = dst->ne[2]; // number of sequences in the batch
|
133
|
+
|
134
|
+
GGML_ASSERT(dst->ne[0] == nr);
|
135
|
+
GGML_ASSERT(src0->nb[0] == sizeof(float));
|
136
|
+
GGML_ASSERT(src1->nb[0] == sizeof(float));
|
137
|
+
GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
|
138
|
+
|
139
|
+
const float * src0_d = (const float *) src0->data;
|
140
|
+
const float * src1_d = (const float *) src1->data;
|
141
|
+
float * dst_d = (float *) dst->data;
|
142
|
+
cudaStream_t stream = ctx.stream();
|
143
|
+
|
144
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
145
|
+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
146
|
+
ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1],
|
147
|
+
dst->nb[2], nc, nr, n_t, n_s, stream);
|
148
|
+
}
|
@@ -0,0 +1,153 @@
|
|
1
|
+
#include "ssm-scan.cuh"
|
2
|
+
|
3
|
+
template <size_t splitD, size_t N>
|
4
|
+
__global__ void __launch_bounds__(splitD, 2)
|
5
|
+
ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
|
6
|
+
const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
|
7
|
+
const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
|
8
|
+
const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
|
9
|
+
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
|
10
|
+
float * __restrict__ dst, const int64_t L) {
|
11
|
+
GGML_UNUSED(src1_nb0);
|
12
|
+
GGML_UNUSED(src2_nb0);
|
13
|
+
const int bidx = blockIdx.x; // split along B
|
14
|
+
const int bidy = blockIdx.y; // split along D
|
15
|
+
const int tid = threadIdx.x;
|
16
|
+
const int wid = tid / 32;
|
17
|
+
const int wtid = tid % 32;
|
18
|
+
|
19
|
+
extern __shared__ float smem[];
|
20
|
+
const int stride_sA = N + 1;
|
21
|
+
const int stride_ss0 = N + 1;
|
22
|
+
float * smem_A = smem;
|
23
|
+
float * smem_s0 = smem_A + splitD * stride_sA;
|
24
|
+
|
25
|
+
const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
|
26
|
+
const float * x_block = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
|
27
|
+
const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
|
28
|
+
const float * A_block = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
|
29
|
+
const float * B_block = (const float *) ((const char *) src4 + (bidx * src4_nb2));
|
30
|
+
const float * C_block = (const float *) ((const char *) src5 + (bidx * src5_nb2));
|
31
|
+
float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
|
32
|
+
float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
|
33
|
+
|
34
|
+
const int stride_s0 = src0_nb1 / sizeof(float);
|
35
|
+
const int stride_x = src1_nb1 / sizeof(float);
|
36
|
+
const int stride_dt = src2_nb1 / sizeof(float);
|
37
|
+
const int stride_A = src3_nb1 / sizeof(float);
|
38
|
+
const int stride_B = src4_nb1 / sizeof(float);
|
39
|
+
const int stride_C = src5_nb1 / sizeof(float);
|
40
|
+
const int stride_s = stride_s0;
|
41
|
+
const int stride_y = stride_x;
|
42
|
+
|
43
|
+
// can N not be 16? for example 32?
|
44
|
+
if (N == 16) {
|
45
|
+
#pragma unroll
|
46
|
+
for (size_t i = 0; i < splitD / 4; i += 2) {
|
47
|
+
float value = A_block[(wid * warpSize + i) * stride_A + wtid];
|
48
|
+
// todo: bank conflict
|
49
|
+
// I am always confused with how to use the swizzling method to solve
|
50
|
+
// bank conflit. Hoping somebody can tell me.
|
51
|
+
smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
|
52
|
+
}
|
53
|
+
#pragma unroll
|
54
|
+
for (size_t i = 0; i < splitD / 4; i += 2) {
|
55
|
+
float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
|
56
|
+
smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
|
57
|
+
}
|
58
|
+
}
|
59
|
+
|
60
|
+
__syncthreads();
|
61
|
+
|
62
|
+
for (int64_t i = 0; i < L; i++) {
|
63
|
+
float dt_soft_plus = dt_block[i * stride_dt + tid];
|
64
|
+
if (dt_soft_plus <= 20.0f) {
|
65
|
+
dt_soft_plus = log1pf(exp(dt_soft_plus));
|
66
|
+
}
|
67
|
+
float x_dt = x_block[i * stride_x + tid] * dt_soft_plus;
|
68
|
+
float sumf = 0.0f;
|
69
|
+
#pragma unroll
|
70
|
+
for (size_t j = 0; j < N; j++) {
|
71
|
+
float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) +
|
72
|
+
(B_block[i * stride_B + j] * x_dt);
|
73
|
+
sumf += state * C_block[i * stride_C + j];
|
74
|
+
if (i == L - 1) {
|
75
|
+
s_block[tid * stride_s + j] = state;
|
76
|
+
} else {
|
77
|
+
smem_s0[tid * stride_ss0 + j] = state;
|
78
|
+
}
|
79
|
+
}
|
80
|
+
__syncthreads();
|
81
|
+
y_block[i * stride_y + tid] = sumf;
|
82
|
+
}
|
83
|
+
}
|
84
|
+
|
85
|
+
static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
|
86
|
+
const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
|
87
|
+
const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
|
88
|
+
const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
|
89
|
+
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
|
90
|
+
float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
|
91
|
+
cudaStream_t stream) {
|
92
|
+
const int threads = 128;
|
93
|
+
// todo: consider D cannot be divided,does this situation exist?
|
94
|
+
GGML_ASSERT(D % threads == 0);
|
95
|
+
const dim3 blocks(B, (D + threads - 1) / threads, 1);
|
96
|
+
const int smem_size = (threads * (N + 1) * 2) * sizeof(float);
|
97
|
+
if (N == 16) {
|
98
|
+
ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
|
99
|
+
src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
|
100
|
+
src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
|
101
|
+
} else {
|
102
|
+
GGML_ABORT("doesn't support N!=16.");
|
103
|
+
}
|
104
|
+
}
|
105
|
+
|
106
|
+
void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
107
|
+
const struct ggml_tensor * src0 = dst->src[0]; // s
|
108
|
+
const struct ggml_tensor * src1 = dst->src[1]; // x
|
109
|
+
const struct ggml_tensor * src2 = dst->src[2]; // dt
|
110
|
+
const struct ggml_tensor * src3 = dst->src[3]; // A
|
111
|
+
const struct ggml_tensor * src4 = dst->src[4]; // B
|
112
|
+
const struct ggml_tensor * src5 = dst->src[5]; // C
|
113
|
+
|
114
|
+
// const int64_t d_state = src0->ne[0];
|
115
|
+
// const int64_t d_inner = src0->ne[1];
|
116
|
+
// const int64_t l = src1->ne[1];
|
117
|
+
// const int64_t b = src0->ne[2];
|
118
|
+
|
119
|
+
const int64_t nc = src0->ne[0]; // d_state
|
120
|
+
const int64_t nr = src0->ne[1]; // d_inner
|
121
|
+
const int64_t n_t = src1->ne[1]; // number of tokens per sequence
|
122
|
+
const int64_t n_s = src0->ne[2]; // number of sequences in the batch
|
123
|
+
|
124
|
+
GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
|
125
|
+
GGML_ASSERT(src0->nb[0] == sizeof(float));
|
126
|
+
GGML_ASSERT(src1->nb[0] == sizeof(float));
|
127
|
+
GGML_ASSERT(src2->nb[0] == sizeof(float));
|
128
|
+
GGML_ASSERT(src3->nb[0] == sizeof(float));
|
129
|
+
GGML_ASSERT(src4->nb[0] == sizeof(float));
|
130
|
+
GGML_ASSERT(src5->nb[0] == sizeof(float));
|
131
|
+
// required for the dot product between s and C
|
132
|
+
GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
|
133
|
+
// required for per-sequence offsets for states
|
134
|
+
GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
|
135
|
+
// required to get correct offset for state destination (i.e. src1->nb[3])
|
136
|
+
GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
|
137
|
+
|
138
|
+
const float * src0_d = (const float *) src0->data;
|
139
|
+
const float * src1_d = (const float *) src1->data;
|
140
|
+
const float * src2_d = (const float *) src2->data;
|
141
|
+
const float * src3_d = (const float *) src3->data;
|
142
|
+
const float * src4_d = (const float *) src4->data;
|
143
|
+
const float * src5_d = (const float *) src5->data;
|
144
|
+
float * dst_d = (float *) dst->data;
|
145
|
+
cudaStream_t stream = ctx.stream();
|
146
|
+
|
147
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
148
|
+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
149
|
+
|
150
|
+
ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
|
151
|
+
src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
|
152
|
+
src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
|
153
|
+
}
|
@@ -0,0 +1,45 @@
|
|
1
|
+
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
|
2
|
+
#define USE_CUB
|
3
|
+
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
|
4
|
+
|
5
|
+
#ifdef USE_CUB
|
6
|
+
#include <cub/cub.cuh>
|
7
|
+
using namespace cub;
|
8
|
+
#endif // USE_CUB
|
9
|
+
|
10
|
+
#include "sumrows.cuh"
|
11
|
+
#include "sum.cuh"
|
12
|
+
|
13
|
+
#include <cstdint>
|
14
|
+
|
15
|
+
void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
|
16
|
+
#ifdef USE_CUB
|
17
|
+
size_t tmp_size = 0;
|
18
|
+
DeviceReduce::Sum(nullptr, tmp_size, x, dst, ne, stream);
|
19
|
+
ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
|
20
|
+
DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, x, dst, ne, stream);
|
21
|
+
#else
|
22
|
+
// Use (inefficient) sum_rows implementation as a fallback.
|
23
|
+
// For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
|
24
|
+
sum_rows_f32_cuda(x, dst, ne, 1, stream);
|
25
|
+
GGML_UNUSED(pool);
|
26
|
+
#endif // USE_CUB
|
27
|
+
}
|
28
|
+
|
29
|
+
void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
30
|
+
const ggml_tensor * src0 = dst->src[0];
|
31
|
+
|
32
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
33
|
+
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
34
|
+
GGML_ASSERT(ggml_is_contiguously_allocated(src0));
|
35
|
+
|
36
|
+
const float * src0_d = (const float *) src0->data;
|
37
|
+
float * dst_d = (float *) dst->data;
|
38
|
+
|
39
|
+
const int64_t ne = ggml_nelements(src0);
|
40
|
+
|
41
|
+
ggml_cuda_pool & pool = ctx.pool();
|
42
|
+
cudaStream_t stream = ctx.stream();
|
43
|
+
|
44
|
+
sum_f32_cuda(pool, src0_d, dst_d, ne, stream);
|
45
|
+
}
|