whispercpp 1.3.6 → 1.3.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/.document +3 -0
- data/.rdoc_options +2 -0
- data/README.md +43 -9
- data/Rakefile +18 -3
- data/ext/dependencies.rb +10 -4
- data/ext/dependencies_for_windows.rb +17 -0
- data/ext/extconf.rb +20 -8
- data/ext/options.rb +54 -14
- data/ext/options_for_windows.rb +51 -0
- data/ext/ruby_whisper.c +35 -42
- data/ext/ruby_whisper.h +141 -0
- data/ext/ruby_whisper_context.c +157 -29
- data/ext/ruby_whisper_log_queue.c +180 -0
- data/ext/ruby_whisper_log_settable.h +46 -0
- data/ext/ruby_whisper_parakeet.c +49 -0
- data/ext/ruby_whisper_parakeet_context.c +304 -0
- data/ext/ruby_whisper_parakeet_context_params.c +117 -0
- data/ext/ruby_whisper_parakeet_model.c +84 -0
- data/ext/ruby_whisper_parakeet_params.c +548 -0
- data/ext/ruby_whisper_parakeet_segment.c +157 -0
- data/ext/ruby_whisper_parakeet_token.c +188 -0
- data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
- data/ext/ruby_whisper_params.c +265 -73
- data/ext/ruby_whisper_segment.c +6 -6
- data/ext/ruby_whisper_transcribe.cpp +23 -15
- data/ext/ruby_whisper_vad_context.c +30 -10
- data/ext/ruby_whisper_vad_context_detect.cpp +8 -9
- data/ext/ruby_whisper_vad_params.c +4 -4
- data/ext/ruby_whisper_vad_segment.c +2 -2
- data/ext/sources/CMakeLists.txt +42 -3
- data/ext/sources/CMakePresets.json +95 -0
- data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
- data/ext/sources/cmake/parakeet.pc.in +10 -0
- data/ext/sources/cmake/whisper.pc.in +2 -2
- data/ext/sources/examples/CMakeLists.txt +4 -2
- data/ext/sources/examples/bench/bench.cpp +1 -1
- data/ext/sources/examples/cli/cli.cpp +52 -10
- data/ext/sources/examples/common-ggml.cpp +4 -0
- data/ext/sources/examples/common-whisper.cpp +139 -67
- data/ext/sources/examples/common-whisper.h +11 -0
- data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
- data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
- data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
- data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
- data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
- data/ext/sources/examples/server/server.cpp +199 -163
- data/ext/sources/examples/vad-speech-segments/speech.cpp +3 -2
- data/ext/sources/ggml/CMakeLists.txt +21 -14
- data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
- data/ext/sources/ggml/include/ggml-alloc.h +1 -0
- data/ext/sources/ggml/include/ggml-backend.h +72 -10
- data/ext/sources/ggml/include/ggml-cuda.h +2 -2
- data/ext/sources/ggml/include/ggml-rpc.h +3 -3
- data/ext/sources/ggml/include/ggml-sycl.h +8 -0
- data/ext/sources/ggml/include/ggml.h +103 -9
- data/ext/sources/ggml/include/gguf.h +10 -2
- data/ext/sources/ggml/src/CMakeLists.txt +30 -6
- data/ext/sources/ggml/src/ggml-alloc.c +5 -1
- data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
- data/ext/sources/ggml/src/ggml-backend-meta.cpp +2266 -0
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +12 -0
- data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
- data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
- data/ext/sources/ggml/src/ggml-common.h +24 -2
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +59 -30
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +21 -22
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -11
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +4279 -1292
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +319 -31
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +12 -2
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +109 -5
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +107 -82
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +501 -119
- data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +106 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +91 -1
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +14 -16
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +39 -55
- data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +62 -26
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +134 -64
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
- data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +246 -28
- data/ext/sources/ggml/src/ggml-cuda/concat.cu +134 -116
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +14 -12
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +139 -34
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +88 -29
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +287 -49
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +335 -130
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +127 -24
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +169 -60
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +109 -45
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +10 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +48 -23
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +2034 -2104
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +25 -12
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +562 -97
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +66 -7
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +78 -10
- data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
- data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +45 -13
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -4
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +33 -24
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +23 -4
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +1 -5
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +2788 -1762
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +13 -4
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
- data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +17 -19
- data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +170 -127
- data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +1774 -396
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +303 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +148 -42
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +255 -62
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +25 -21
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1303 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +167 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +157 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +104 -13
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +222 -57
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +10 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +78 -26
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +48 -8
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +232 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +40 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +625 -816
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3052 -2166
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.h +650 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +337 -106
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
- data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +245 -151
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +15 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.h +8 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +390 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -5
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +27 -9
- data/ext/sources/ggml/src/ggml-impl.h +6 -1
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +207 -18
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +36 -2
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +186 -29
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +118 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +322 -21
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +4 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1226 -467
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +67 -5
- data/ext/sources/ggml/src/ggml-opencl/fa_tune.h +92 -0
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +16290 -6246
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1997 -92
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +81 -41
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +88 -39
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +1995 -96
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +1615 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1486 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +374 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +121 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1149 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +48 -64
- data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +5 -2
- data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +500 -0
- data/ext/sources/ggml/src/ggml-opencl/libdl.h +79 -0
- data/ext/sources/ggml/src/ggml-openvino/.clang-format +0 -5
- data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +2 -4
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +740 -127
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +76 -23
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +75 -14
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +29 -8
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +339 -69
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +330 -192
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +10 -4
- data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +56 -16
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +1 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +4 -4
- data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +94 -37
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +8 -16
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +14 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +108 -21
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +282 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +2 -9
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +21 -7
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +41 -8
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +44 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +226 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +19 -9
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +58 -13
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/repeat.cpp +74 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +13 -6
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +1 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +161 -39
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +3 -3
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +126 -49
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +59 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +32 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +1 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +90 -25
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +41 -22
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +18 -4
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +1 -1
- data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +70 -43
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +5 -4
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +612 -36
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +29 -26
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +460 -114
- data/ext/sources/ggml/src/ggml-openvino/utils.h +32 -9
- data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
- data/ext/sources/ggml/src/ggml-quants.c +365 -114
- data/ext/sources/ggml/src/ggml-quants.h +6 -0
- data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
- data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
- data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
- data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +50 -4
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +5 -1
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +12 -0
- data/ext/sources/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
- data/ext/sources/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/common.cpp +72 -2
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +59 -2
- data/ext/sources/ggml/src/ggml-sycl/concat.cpp +21 -1
- data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.cpp +158 -0
- data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
- data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
- data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
- data/ext/sources/ggml/src/ggml-sycl/conv2d.cpp +150 -0
- data/ext/sources/ggml/src/ggml-sycl/conv2d.hpp +10 -0
- data/ext/sources/ggml/src/ggml-sycl/conv3d.cpp +224 -0
- data/ext/sources/ggml/src/ggml-sycl/conv3d.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +121 -13
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +706 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +281 -0
- data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
- data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
- data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +678 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +997 -244
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +15 -7
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +215 -204
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +7 -5
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
- data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
- data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +69 -31
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1758 -455
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
- data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
- data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1542 -39
- data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +33 -0
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +103 -49
- data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +45 -9
- data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
- data/ext/sources/ggml/src/ggml-sycl/pool.cpp +185 -0
- data/ext/sources/ggml/src/ggml-sycl/pool.hpp +22 -0
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -1
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +17 -3
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +9 -10
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +242 -45
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +16 -0
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2843 -700
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +431 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +3 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +198 -48
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +116 -113
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +125 -64
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +21 -19
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +29 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +17 -11
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +4 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +22 -24
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +42 -40
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +49 -15
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +222 -171
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +10 -10
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +3 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +3 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +5 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +11 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +3 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +282 -211
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2209 -283
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2618 -1416
- data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +8 -7
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +90 -95
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -50
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +204 -78
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +805 -526
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +8 -6
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +90 -413
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1553 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +6 -4
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +2 -3
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +68 -48
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
- data/ext/sources/ggml/src/ggml.c +146 -42
- data/ext/sources/ggml/src/gguf.cpp +173 -28
- data/ext/sources/include/parakeet.h +342 -0
- data/ext/sources/include/whisper.h +31 -0
- data/ext/sources/media/matmul.png +0 -0
- data/ext/sources/src/CMakeLists.txt +23 -0
- data/ext/sources/src/parakeet-arch.h +188 -0
- data/ext/sources/src/parakeet.cpp +3838 -0
- data/ext/sources/src/whisper.cpp +220 -26
- data/extsources.rb +26 -10
- data/lib/whisper/log_settable.rb +33 -0
- data/lib/whisper/model/uri.rb +13 -8
- data/lib/whisper/output.rb +74 -0
- data/sig/whisper.rbs +417 -62
- data/test/helper.rb +2 -0
- data/test/jfk_reader/jfk_reader.c +50 -7
- data/test/test_callback.rb +1 -0
- data/test/test_package.rb +6 -5
- data/test/test_parakeet.rb +28 -0
- data/test/test_parakeet_callback.rb +107 -0
- data/test/test_parakeet_context.rb +116 -0
- data/test/test_parakeet_context_params.rb +24 -0
- data/test/test_parakeet_model.rb +21 -0
- data/test/test_parakeet_params.rb +78 -0
- data/test/test_parakeet_segment.rb +42 -0
- data/test/test_parakeet_token.rb +73 -0
- data/test/test_params.rb +2 -0
- data/test/test_vad.rb +9 -0
- data/test/test_vad_context.rb +2 -2
- data/test/test_vad_segment.rb +1 -1
- data/test/test_whisper.rb +24 -6
- data/whispercpp.gemspec +2 -2
- metadata +263 -304
- data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
- data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
- data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
- data/ext/sources/bindings/javascript/package.json +0 -26
- data/ext/sources/bindings/javascript/whisper.js +0 -19
- data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
- data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
- data/ext/sources/examples/addon.node/addon.cpp +0 -557
- data/ext/sources/examples/addon.node/index.js +0 -59
- data/ext/sources/examples/addon.node/package.json +0 -16
- data/ext/sources/examples/addon.node/vad-example.js +0 -132
- data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
- data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
- data/ext/sources/examples/coi-serviceworker.js +0 -146
- data/ext/sources/examples/command/CMakeLists.txt +0 -10
- data/ext/sources/examples/command/command.cpp +0 -802
- data/ext/sources/examples/command/commands.txt +0 -9
- data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
- data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
- data/ext/sources/examples/generate-karaoke.sh +0 -57
- data/ext/sources/examples/helpers.js +0 -191
- data/ext/sources/examples/livestream.sh +0 -112
- data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
- data/ext/sources/examples/lsp/lsp.cpp +0 -471
- data/ext/sources/examples/lsp/whisper.vim +0 -362
- data/ext/sources/examples/python/test_whisper_processor.py +0 -7
- data/ext/sources/examples/python/whisper_processor.py +0 -54
- data/ext/sources/examples/server/bench.js +0 -29
- data/ext/sources/examples/server.py +0 -120
- data/ext/sources/examples/stream/CMakeLists.txt +0 -10
- data/ext/sources/examples/stream/stream.cpp +0 -437
- data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
- data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
- data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
- data/ext/sources/examples/sycl/build.sh +0 -22
- data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
- data/ext/sources/examples/sycl/run-whisper.sh +0 -17
- data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -48
- data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
- data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
- data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
- data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
- data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
- data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
- data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
- data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
- data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
- data/ext/sources/examples/talk-llama/llama-context.h +0 -359
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
- data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
- data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
- data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
- data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2735
- data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
- data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
- data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
- data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
- data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
- data/ext/sources/examples/talk-llama/llama-io.h +0 -35
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -330
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1165
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
- data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
- data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -752
- data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
- data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
- data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
- data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
- data/ext/sources/examples/talk-llama/llama-model.h +0 -597
- data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
- data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
- data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
- data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
- data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
- data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
- data/ext/sources/examples/talk-llama/llama.h +0 -1573
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
- data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
- data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -137
- data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
- data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
- data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
- data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
- data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
- data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
- data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
- data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
- data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
- data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
- data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
- data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -111
- data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
- data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
- data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
- data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
- data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
- data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -157
- data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -195
- data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
- data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
- data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/jais2.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
- data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
- data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
- data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/llama.cpp +0 -175
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
- data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/models.h +0 -704
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
- data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
- data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
- data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
- data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/paddleocr.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
- data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -320
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
- data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
- data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
- data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
- data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -164
- data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -137
- data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
- data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
- data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
- data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +0 -165
- data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
- data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
- data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
- data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
- data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
- data/ext/sources/examples/talk-llama/speak +0 -40
- data/ext/sources/examples/talk-llama/speak.bat +0 -1
- data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
- data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
- data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
- data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
- data/ext/sources/examples/talk-llama/unicode.cpp +0 -1103
- data/ext/sources/examples/talk-llama/unicode.h +0 -111
- data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
- data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
- data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
- data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
- data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
- data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
- data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
- data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
- data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
- data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
- data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +0 -123
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
- data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp +0 -21
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +0 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +0 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +0 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -21
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +0 -25
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +0 -39
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +0 -23
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp +0 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +0 -21
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/round.comp +0 -29
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +0 -21
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +0 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +0 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp +0 -23
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +0 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +0 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/step.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +0 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp +0 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/xielu.comp +0 -35
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -718
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
- data/ext/sources/tests/CMakeLists.txt +0 -112
- data/ext/sources/tests/earnings21/eval.mk +0 -58
- data/ext/sources/tests/earnings21/eval.py +0 -68
- data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
- data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
- data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
- data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
- data/ext/sources/tests/earnings21/requirements.txt +0 -6
- data/ext/sources/tests/en-0-ref.txt +0 -1
- data/ext/sources/tests/en-1-ref.txt +0 -1
- data/ext/sources/tests/en-2-ref.txt +0 -1
- data/ext/sources/tests/es-0-ref.txt +0 -1
- data/ext/sources/tests/librispeech/eval.mk +0 -39
- data/ext/sources/tests/librispeech/eval.py +0 -47
- data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
- data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
- data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
- data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
- data/ext/sources/tests/librispeech/requirements.txt +0 -6
- data/ext/sources/tests/run-tests.sh +0 -130
- data/ext/sources/tests/test-c.c +0 -3
- data/ext/sources/tests/test-vad-full.cpp +0 -56
- data/ext/sources/tests/test-vad.cpp +0 -83
- data/ext/sources/tests/test-whisper.js +0 -58
- data/lib/whisper/context.rb +0 -15
- data/lib/whisper/segment.rb +0 -58
|
@@ -34,10 +34,22 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
|
|
|
34
34
|
GGML_ASSERT(V->ne[0] == K->ne[0]);
|
|
35
35
|
ggml_cuda_flash_attn_ext_tile_case<128, 128>(ctx, dst);
|
|
36
36
|
} break;
|
|
37
|
+
case 192: {
|
|
38
|
+
GGML_ASSERT(V->ne[0] == 128);
|
|
39
|
+
ggml_cuda_flash_attn_ext_tile_case<192, 128>(ctx, dst);
|
|
40
|
+
} break;
|
|
37
41
|
case 256: {
|
|
38
42
|
GGML_ASSERT(V->ne[0] == K->ne[0]);
|
|
39
43
|
ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
|
|
40
44
|
} break;
|
|
45
|
+
case 320: {
|
|
46
|
+
GGML_ASSERT(V->ne[0] == 256);
|
|
47
|
+
ggml_cuda_flash_attn_ext_tile_case<320, 256>(ctx, dst);
|
|
48
|
+
} break;
|
|
49
|
+
case 512: {
|
|
50
|
+
GGML_ASSERT(V->ne[0] == K->ne[0]);
|
|
51
|
+
ggml_cuda_flash_attn_ext_tile_case<512, 512>(ctx, dst);
|
|
52
|
+
} break;
|
|
41
53
|
case 576: {
|
|
42
54
|
GGML_ASSERT(V->ne[0] == 512);
|
|
43
55
|
ggml_cuda_flash_attn_ext_tile_case<576, 512>(ctx, dst);
|
|
@@ -62,12 +62,25 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
|
|
|
62
62
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2, 64, 64)
|
|
63
63
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
|
|
64
64
|
|
|
65
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 64, 2, 64, 64)
|
|
66
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 2, 64, 64)
|
|
67
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 64, 64)
|
|
68
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 64, 64)
|
|
69
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 64, 64)
|
|
70
|
+
|
|
65
71
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 64, 2, 64, 64)
|
|
66
72
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 2, 64, 64)
|
|
67
73
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 64, 64)
|
|
68
74
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 64, 64)
|
|
69
75
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 64, 64)
|
|
70
76
|
|
|
77
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 64, 64)
|
|
78
|
+
|
|
79
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
|
|
80
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
|
|
81
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
82
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
|
|
83
|
+
|
|
71
84
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
72
85
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
73
86
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
|
|
@@ -118,12 +131,25 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
|
|
|
118
131
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 128, 3, 32, 128)
|
|
119
132
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
|
|
120
133
|
|
|
134
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 128, 3, 64, 64)
|
|
135
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 3, 32, 64)
|
|
136
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 32, 64)
|
|
137
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
|
|
138
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
|
|
139
|
+
|
|
121
140
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 128, 3, 64, 64)
|
|
122
141
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 3, 32, 64)
|
|
123
142
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 32, 256)
|
|
124
143
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
|
|
125
144
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64)
|
|
126
145
|
|
|
146
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 32, 64)
|
|
147
|
+
|
|
148
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
|
|
149
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 32, 64)
|
|
150
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 32, 64)
|
|
151
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 32, 64)
|
|
152
|
+
|
|
127
153
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 32, 64)
|
|
128
154
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 32, 64)
|
|
129
155
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 32, 64)
|
|
@@ -181,12 +207,26 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
|
|
|
181
207
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
|
|
182
208
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2, 64, 32)
|
|
183
209
|
|
|
210
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 256, 2, 128, 64)
|
|
211
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 256, 2, 64, 64)
|
|
212
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 64, 64)
|
|
213
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
|
|
214
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
|
|
215
|
+
|
|
184
216
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 256, 2, 128, 64)
|
|
185
217
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 256, 2, 64, 128)
|
|
186
218
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 64, 128)
|
|
187
219
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
|
|
188
220
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128)
|
|
189
221
|
|
|
222
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 512, 1, 128, 64)
|
|
223
|
+
|
|
224
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
|
|
225
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
|
|
226
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
227
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
|
|
228
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 512, 1, 128, 64)
|
|
229
|
+
|
|
190
230
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
191
231
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
192
232
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
|
|
@@ -245,12 +285,26 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
|
|
|
245
285
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128, 64)
|
|
246
286
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3, 64, 64)
|
|
247
287
|
|
|
288
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 64, 8, 32, 64)
|
|
289
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 6, 32, 64)
|
|
290
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 128, 6, 32, 64)
|
|
291
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 5, 32, 64)
|
|
292
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 3, 64, 64)
|
|
293
|
+
|
|
248
294
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 64, 8, 32, 64)
|
|
249
295
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 6, 32, 256)
|
|
250
296
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 128, 6, 32, 256)
|
|
251
297
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256)
|
|
252
298
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128)
|
|
253
299
|
|
|
300
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 256, 2, 128, 64)
|
|
301
|
+
|
|
302
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
|
|
303
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
|
|
304
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
305
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4, 64, 64)
|
|
306
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2, 128, 64)
|
|
307
|
+
|
|
254
308
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
255
309
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
256
310
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4, 64, 64)
|
|
@@ -738,14 +792,14 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
738
792
|
template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap> // D == head size
|
|
739
793
|
__launch_bounds__(ggml_cuda_fattn_tile_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_tile_get_occupancy(DKQ, DV, ncols1*ncols2))
|
|
740
794
|
static __global__ void flash_attn_tile(
|
|
741
|
-
const char *
|
|
742
|
-
const char *
|
|
743
|
-
const char *
|
|
744
|
-
const char *
|
|
745
|
-
const char *
|
|
746
|
-
const int *
|
|
747
|
-
float *
|
|
748
|
-
float2 *
|
|
795
|
+
const char * Q_ptr,
|
|
796
|
+
const char * K_ptr,
|
|
797
|
+
const char * V_ptr,
|
|
798
|
+
const char * mask_ptr,
|
|
799
|
+
const char * sinks_ptr,
|
|
800
|
+
const int * KV_max_ptr,
|
|
801
|
+
float * dst_ptr,
|
|
802
|
+
float2 * dst_meta_ptr,
|
|
749
803
|
const float scale,
|
|
750
804
|
const float max_bias,
|
|
751
805
|
const float m0,
|
|
@@ -760,6 +814,14 @@ static __global__ void flash_attn_tile(
|
|
|
760
814
|
const int32_t ne31, const int32_t ne32, const int32_t ne33,
|
|
761
815
|
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
|
|
762
816
|
#ifdef FLASH_ATTN_AVAILABLE
|
|
817
|
+
const char * GGML_CUDA_RESTRICT Q = Q_ptr;
|
|
818
|
+
const char * GGML_CUDA_RESTRICT K = K_ptr;
|
|
819
|
+
const char * GGML_CUDA_RESTRICT V = V_ptr;
|
|
820
|
+
const char * GGML_CUDA_RESTRICT mask = mask_ptr;
|
|
821
|
+
const char * GGML_CUDA_RESTRICT sinks = sinks_ptr;
|
|
822
|
+
const int * GGML_CUDA_RESTRICT KV_max = KV_max_ptr;
|
|
823
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
824
|
+
float2 * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
|
|
763
825
|
|
|
764
826
|
// Skip unused kernel variants for faster compilation:
|
|
765
827
|
|
|
@@ -767,7 +829,7 @@ static __global__ void flash_attn_tile(
|
|
|
767
829
|
#ifdef GGML_USE_WMMA_FATTN
|
|
768
830
|
(ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
|
|
769
831
|
#endif // GGML_USE_WMMA_FATTN
|
|
770
|
-
(use_logit_softcap && !(DV == 128 || DV == 256))
|
|
832
|
+
(use_logit_softcap && !(DV == 128 || DV == 256 || DV == 512))
|
|
771
833
|
) {
|
|
772
834
|
GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
|
|
773
835
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
@@ -844,6 +906,8 @@ static __global__ void flash_attn_tile(
|
|
|
844
906
|
}
|
|
845
907
|
float KQ_sum[cpw] = {0.0f};
|
|
846
908
|
|
|
909
|
+
ggml_cuda_pdl_sync();
|
|
910
|
+
|
|
847
911
|
// Load Q data, convert to FP16 if fast:
|
|
848
912
|
#pragma unroll
|
|
849
913
|
for (int jc0 = 0; jc0 < cpw; ++jc0) {
|
|
@@ -1074,7 +1138,7 @@ static __global__ void flash_attn_tile(
|
|
|
1074
1138
|
}
|
|
1075
1139
|
}
|
|
1076
1140
|
#else
|
|
1077
|
-
GGML_UNUSED_VARS(
|
|
1141
|
+
GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
|
|
1078
1142
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
1079
1143
|
ne00, ne01, ne02, ne03,
|
|
1080
1144
|
nb01, nb02, nb03,
|
|
@@ -1098,7 +1162,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1098
1162
|
constexpr size_t nbytes_shared = 0;
|
|
1099
1163
|
|
|
1100
1164
|
#ifdef GGML_USE_HIP
|
|
1101
|
-
if constexpr (
|
|
1165
|
+
if constexpr (DKQ <= 128) {
|
|
1102
1166
|
if (Q->ne[1] > 32/ncols2) {
|
|
1103
1167
|
constexpr int cols_per_block = 64;
|
|
1104
1168
|
const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
|
|
@@ -1112,7 +1176,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1112
1176
|
#endif // GGML_USE_HIP
|
|
1113
1177
|
|
|
1114
1178
|
#ifndef GGML_USE_HIP
|
|
1115
|
-
if constexpr (
|
|
1179
|
+
if constexpr (DKQ <= 256)
|
|
1116
1180
|
#endif // GGML_USE_HIP
|
|
1117
1181
|
{
|
|
1118
1182
|
if (Q->ne[1] > 16/ncols2) {
|
|
@@ -1126,14 +1190,16 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1126
1190
|
}
|
|
1127
1191
|
}
|
|
1128
1192
|
|
|
1129
|
-
if (
|
|
1130
|
-
|
|
1131
|
-
|
|
1132
|
-
|
|
1133
|
-
|
|
1134
|
-
|
|
1135
|
-
|
|
1136
|
-
|
|
1193
|
+
if constexpr (ncols2 <= 16) {
|
|
1194
|
+
if (Q->ne[1] > 8/ncols2) {
|
|
1195
|
+
constexpr int cols_per_block = 16;
|
|
1196
|
+
const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
|
|
1197
|
+
const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
|
|
1198
|
+
fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
|
|
1199
|
+
launch_fattn<DV, cols_per_block/ncols2, ncols2>
|
|
1200
|
+
(ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
|
|
1201
|
+
return;
|
|
1202
|
+
}
|
|
1137
1203
|
}
|
|
1138
1204
|
|
|
1139
1205
|
if constexpr (ncols2 <= 8) {
|
|
@@ -1192,7 +1258,26 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
|
|
|
1192
1258
|
const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
|
|
1193
1259
|
const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
|
|
1194
1260
|
|
|
1195
|
-
if constexpr (
|
|
1261
|
+
if constexpr (DKQ == 320) {
|
|
1262
|
+
// This branch is only used for Mistral Small 4 which has a GQA ratio of 32.
|
|
1263
|
+
// On AMD, simply use that GQA ratio with 32 columns / block since we always have enough SRAM.
|
|
1264
|
+
// On NVIDIA however, the tile kernel is only used for GPUs that can't use the mma kernel (Pascal and older).
|
|
1265
|
+
// Therefore, use a GQA ratio of 16 with 16 columns / block to stay below 48 kiB of SRAM / block.
|
|
1266
|
+
#ifdef GGML_USE_HIP
|
|
1267
|
+
if (use_gqa_opt && gqa_ratio % 32 == 0) {
|
|
1268
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 32, use_logit_softcap>(ctx, dst);
|
|
1269
|
+
return;
|
|
1270
|
+
}
|
|
1271
|
+
#else
|
|
1272
|
+
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1273
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1274
|
+
return;
|
|
1275
|
+
}
|
|
1276
|
+
#endif // GGML_USE_HIP
|
|
1277
|
+
GGML_ABORT("flash-attn tile (320/256): expected GQA ratio multiple of 32");
|
|
1278
|
+
}
|
|
1279
|
+
|
|
1280
|
+
if constexpr (DKQ == 576) {
|
|
1196
1281
|
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1197
1282
|
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1198
1283
|
return;
|
|
@@ -1203,7 +1288,20 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
|
|
|
1203
1288
|
}
|
|
1204
1289
|
}
|
|
1205
1290
|
|
|
1206
|
-
if constexpr (
|
|
1291
|
+
if constexpr (DKQ == 192) {
|
|
1292
|
+
// MiMo-V2.5 / V2.5-Pro / V2-Flash: gqa_ratio is 8 (SWA) or 16 (full attn)
|
|
1293
|
+
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1294
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1295
|
+
return;
|
|
1296
|
+
}
|
|
1297
|
+
if (use_gqa_opt && gqa_ratio % 8 == 0) {
|
|
1298
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
|
|
1299
|
+
return;
|
|
1300
|
+
}
|
|
1301
|
+
GGML_ABORT("flash-attn tile (192/128): expected GQA ratio multiple of 8");
|
|
1302
|
+
}
|
|
1303
|
+
|
|
1304
|
+
if constexpr (DKQ <= 512 && DKQ != 320 && DKQ != 192) {
|
|
1207
1305
|
if (use_gqa_opt && gqa_ratio % 8 == 0) {
|
|
1208
1306
|
launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
|
|
1209
1307
|
return;
|
|
@@ -1219,8 +1317,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
|
|
|
1219
1317
|
return;
|
|
1220
1318
|
}
|
|
1221
1319
|
|
|
1222
|
-
|
|
1223
|
-
|
|
1320
|
+
if constexpr (DV <= 256) {
|
|
1321
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
|
|
1322
|
+
return;
|
|
1323
|
+
}
|
|
1224
1324
|
}
|
|
1225
1325
|
GGML_ABORT("fatal error");
|
|
1226
1326
|
}
|
|
@@ -1254,5 +1354,8 @@ extern DECL_FATTN_TILE_CASE( 80, 80);
|
|
|
1254
1354
|
extern DECL_FATTN_TILE_CASE( 96, 96);
|
|
1255
1355
|
extern DECL_FATTN_TILE_CASE(112, 112);
|
|
1256
1356
|
extern DECL_FATTN_TILE_CASE(128, 128);
|
|
1357
|
+
extern DECL_FATTN_TILE_CASE(192, 128);
|
|
1257
1358
|
extern DECL_FATTN_TILE_CASE(256, 256);
|
|
1359
|
+
extern DECL_FATTN_TILE_CASE(320, 256);
|
|
1360
|
+
extern DECL_FATTN_TILE_CASE(512, 512);
|
|
1258
1361
|
extern DECL_FATTN_TILE_CASE(576, 512);
|
|
@@ -19,14 +19,14 @@ static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
|
|
|
19
19
|
template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
|
|
20
20
|
__launch_bounds__(ggml_cuda_fattn_vec_get_nthreads_device(), 1)
|
|
21
21
|
static __global__ void flash_attn_ext_vec(
|
|
22
|
-
const char *
|
|
23
|
-
const char *
|
|
24
|
-
const char *
|
|
25
|
-
const char *
|
|
26
|
-
const char *
|
|
27
|
-
const int *
|
|
28
|
-
float *
|
|
29
|
-
float2 *
|
|
22
|
+
const char * Q_ptr,
|
|
23
|
+
const char * K_ptr,
|
|
24
|
+
const char * V_ptr,
|
|
25
|
+
const char * mask_ptr,
|
|
26
|
+
const char * sinks_ptr,
|
|
27
|
+
const int * KV_max_ptr,
|
|
28
|
+
float * dst_ptr,
|
|
29
|
+
float2 * dst_meta_ptr,
|
|
30
30
|
const float scale,
|
|
31
31
|
const float max_bias,
|
|
32
32
|
const float m0,
|
|
@@ -40,7 +40,16 @@ static __global__ void flash_attn_ext_vec(
|
|
|
40
40
|
const int32_t nb21, const int32_t nb22, const int64_t nb23,
|
|
41
41
|
const int32_t ne31, const int32_t ne32, const int32_t ne33,
|
|
42
42
|
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
|
|
43
|
+
ggml_cuda_pdl_lc();
|
|
43
44
|
#ifdef FLASH_ATTN_AVAILABLE
|
|
45
|
+
const char * GGML_CUDA_RESTRICT Q = Q_ptr;
|
|
46
|
+
const char * GGML_CUDA_RESTRICT K = K_ptr;
|
|
47
|
+
const char * GGML_CUDA_RESTRICT V = V_ptr;
|
|
48
|
+
const char * GGML_CUDA_RESTRICT mask = mask_ptr;
|
|
49
|
+
const char * GGML_CUDA_RESTRICT sinks = sinks_ptr;
|
|
50
|
+
const int * GGML_CUDA_RESTRICT KV_max = KV_max_ptr;
|
|
51
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
52
|
+
float2 * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
|
|
44
53
|
|
|
45
54
|
// Skip unused kernel variants for faster compilation:
|
|
46
55
|
if (use_logit_softcap && !(D == 128 || D == 256)) {
|
|
@@ -75,17 +84,17 @@ static __global__ void flash_attn_ext_vec(
|
|
|
75
84
|
#endif // GGML_USE_HIP
|
|
76
85
|
|
|
77
86
|
constexpr int nthreads = ggml_cuda_fattn_vec_get_nthreads_device();
|
|
78
|
-
constexpr int nthreads_KQ = type_K == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_KQ_q;
|
|
79
|
-
constexpr int nthreads_V = type_V == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_V_q;
|
|
87
|
+
constexpr int nthreads_KQ = (type_K == GGML_TYPE_F16 || type_K == GGML_TYPE_BF16) ? 128 / cpy_nb : nthreads_KQ_q;
|
|
88
|
+
constexpr int nthreads_V = (type_V == GGML_TYPE_F16 || type_V == GGML_TYPE_BF16) ? 128 / cpy_nb : nthreads_V_q;
|
|
80
89
|
|
|
81
90
|
static_assert(WARP_SIZE % nthreads_KQ == 0, "bad nthreads_K");
|
|
82
91
|
static_assert(WARP_SIZE % nthreads_V == 0, "bad nthreads_V");
|
|
83
92
|
|
|
84
|
-
constexpr int V_rows_per_thread = type_V == GGML_TYPE_F16 ? 2*cpy_ne : 4;
|
|
93
|
+
constexpr int V_rows_per_thread = (type_V == GGML_TYPE_F16 || type_V == GGML_TYPE_BF16) ? 2*cpy_ne : 4;
|
|
85
94
|
constexpr int V_cols_per_iter = WARP_SIZE / nthreads_V;
|
|
86
95
|
|
|
87
96
|
constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
|
|
88
|
-
constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
|
|
97
|
+
constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16 && type_K != GGML_TYPE_BF16;
|
|
89
98
|
#ifdef V_DOT2_F32_F16_AVAILABLE
|
|
90
99
|
constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half, V_rows_per_thread>();
|
|
91
100
|
#else
|
|
@@ -136,6 +145,8 @@ static __global__ void flash_attn_ext_vec(
|
|
|
136
145
|
#endif // V_DOT2_F32_F16_AVAILABLE
|
|
137
146
|
int Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
|
|
138
147
|
float2 Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
|
|
148
|
+
|
|
149
|
+
ggml_cuda_pdl_sync();
|
|
139
150
|
if constexpr (Q_q8_1) {
|
|
140
151
|
#pragma unroll
|
|
141
152
|
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
|
|
@@ -323,8 +334,18 @@ static __global__ void flash_attn_ext_vec(
|
|
|
323
334
|
#pragma unroll
|
|
324
335
|
for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
|
|
325
336
|
half2 tmp[V_rows_per_thread/2];
|
|
326
|
-
|
|
327
|
-
|
|
337
|
+
if constexpr (type_V == GGML_TYPE_BF16) {
|
|
338
|
+
float2 tmp_f[V_rows_per_thread/2];
|
|
339
|
+
dequantize_V(V + k*nb21, tmp_f,
|
|
340
|
+
2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
|
|
341
|
+
#pragma unroll
|
|
342
|
+
for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
|
|
343
|
+
tmp[i_VKQ_1] = __float22half2_rn(tmp_f[i_VKQ_1]);
|
|
344
|
+
}
|
|
345
|
+
} else {
|
|
346
|
+
dequantize_V(V + k*nb21, tmp,
|
|
347
|
+
2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
|
|
348
|
+
}
|
|
328
349
|
#pragma unroll
|
|
329
350
|
for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
|
|
330
351
|
#pragma unroll
|
|
@@ -493,7 +514,7 @@ static __global__ void flash_attn_ext_vec(
|
|
|
493
514
|
dst_meta[((sequence*int(ne01.z) + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
|
|
494
515
|
}
|
|
495
516
|
#else
|
|
496
|
-
GGML_UNUSED_VARS(
|
|
517
|
+
GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
|
|
497
518
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
498
519
|
ne00, ne01, ne02, ne03,
|
|
499
520
|
nb01, nb02, nb03,
|
|
@@ -563,6 +584,7 @@ void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_ten
|
|
|
563
584
|
extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_0); \
|
|
564
585
|
extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_1); \
|
|
565
586
|
extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q8_0); \
|
|
587
|
+
extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_BF16); \
|
|
566
588
|
|
|
567
589
|
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_F16)
|
|
568
590
|
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_0)
|
|
@@ -570,6 +592,7 @@ EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_1)
|
|
|
570
592
|
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_0)
|
|
571
593
|
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_1)
|
|
572
594
|
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q8_0)
|
|
595
|
+
EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_BF16)
|
|
573
596
|
|
|
574
597
|
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_F16)
|
|
575
598
|
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_0)
|
|
@@ -577,6 +600,7 @@ EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_1)
|
|
|
577
600
|
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_0)
|
|
578
601
|
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_1)
|
|
579
602
|
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q8_0)
|
|
603
|
+
EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_BF16)
|
|
580
604
|
|
|
581
605
|
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_F16)
|
|
582
606
|
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_0)
|
|
@@ -584,3 +608,4 @@ EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_1)
|
|
|
584
608
|
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
|
|
585
609
|
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
|
|
586
610
|
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
|
|
611
|
+
EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_BF16)
|
|
@@ -24,14 +24,14 @@ namespace wmma = rocwmma;
|
|
|
24
24
|
template<int D, int ncols, int nwarps, int VKQ_stride, typename KQ_acc_t, bool use_logit_softcap>
|
|
25
25
|
__launch_bounds__(nwarps*ggml_cuda_get_physical_warp_size(), 1)
|
|
26
26
|
static __global__ void flash_attn_ext_f16(
|
|
27
|
-
const char *
|
|
28
|
-
const char *
|
|
29
|
-
const char *
|
|
30
|
-
const char *
|
|
31
|
-
const char *
|
|
32
|
-
const int *
|
|
33
|
-
float *
|
|
34
|
-
float2 *
|
|
27
|
+
const char * Q_ptr,
|
|
28
|
+
const char * K_ptr,
|
|
29
|
+
const char * V_ptr,
|
|
30
|
+
const char * mask_ptr,
|
|
31
|
+
const char * sinks_ptr,
|
|
32
|
+
const int * KV_max_ptr,
|
|
33
|
+
float * dst_ptr,
|
|
34
|
+
float2 * dst_meta_ptr,
|
|
35
35
|
const float scale,
|
|
36
36
|
const float max_bias,
|
|
37
37
|
const float m0,
|
|
@@ -46,6 +46,14 @@ static __global__ void flash_attn_ext_f16(
|
|
|
46
46
|
const int32_t ne31, const int32_t ne32, const int32_t ne33,
|
|
47
47
|
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
|
|
48
48
|
#if defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
|
|
49
|
+
const char * GGML_CUDA_RESTRICT Q = Q_ptr;
|
|
50
|
+
const char * GGML_CUDA_RESTRICT K = K_ptr;
|
|
51
|
+
const char * GGML_CUDA_RESTRICT V = V_ptr;
|
|
52
|
+
const char * GGML_CUDA_RESTRICT mask = mask_ptr;
|
|
53
|
+
const char * GGML_CUDA_RESTRICT sinks = sinks_ptr;
|
|
54
|
+
const int * GGML_CUDA_RESTRICT KV_max = KV_max_ptr;
|
|
55
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
56
|
+
float2 * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
|
|
49
57
|
// Skip unused kernel variants for faster compilation:
|
|
50
58
|
if (use_logit_softcap && !(D == 128 || D == 256)) {
|
|
51
59
|
NO_DEVICE_CODE;
|
|
@@ -86,6 +94,7 @@ static __global__ void flash_attn_ext_f16(
|
|
|
86
94
|
constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
|
|
87
95
|
constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
|
|
88
96
|
|
|
97
|
+
ggml_cuda_pdl_sync();
|
|
89
98
|
const int sequence = blockIdx.z / ne02;
|
|
90
99
|
const int head = blockIdx.z - sequence*ne02;
|
|
91
100
|
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
|
|
@@ -493,7 +502,7 @@ static __global__ void flash_attn_ext_f16(
|
|
|
493
502
|
dst_meta[j_dst_unrolled] = dst_meta_val;
|
|
494
503
|
}
|
|
495
504
|
#else
|
|
496
|
-
GGML_UNUSED_VARS(
|
|
505
|
+
GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
|
|
497
506
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
498
507
|
ne00, ne01, ne02, ne03,
|
|
499
508
|
nb01, nb02, nb03,
|