whispercpp 1.3.5 → 1.3.7
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/.document +3 -0
- data/.rdoc_options +2 -0
- data/LICENSE +1 -1
- data/README.md +133 -3
- data/Rakefile +18 -3
- data/ext/dependencies.rb +10 -4
- data/ext/dependencies_for_windows.rb +17 -0
- data/ext/extconf.rb +20 -7
- data/ext/options.rb +54 -14
- data/ext/options_for_windows.rb +51 -0
- data/ext/ruby_whisper.c +56 -46
- data/ext/ruby_whisper.h +165 -2
- data/ext/ruby_whisper_context.c +297 -126
- data/ext/ruby_whisper_context_params.c +163 -0
- data/ext/ruby_whisper_log_queue.c +180 -0
- data/ext/ruby_whisper_log_settable.h +47 -0
- data/ext/ruby_whisper_model.c +0 -1
- data/ext/ruby_whisper_parakeet.c +49 -0
- data/ext/ruby_whisper_parakeet_context.c +304 -0
- data/ext/ruby_whisper_parakeet_context_params.c +117 -0
- data/ext/ruby_whisper_parakeet_model.c +84 -0
- data/ext/ruby_whisper_parakeet_params.c +548 -0
- data/ext/ruby_whisper_parakeet_segment.c +157 -0
- data/ext/ruby_whisper_parakeet_token.c +188 -0
- data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
- data/ext/ruby_whisper_params.c +256 -66
- data/ext/ruby_whisper_segment.c +6 -7
- data/ext/ruby_whisper_token.c +29 -9
- data/ext/ruby_whisper_transcribe.cpp +46 -16
- data/ext/ruby_whisper_vad_context.c +48 -1
- data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
- data/ext/ruby_whisper_vad_params.c +0 -1
- data/ext/ruby_whisper_vad_segment.c +0 -1
- data/ext/ruby_whisper_vad_segments.c +0 -1
- data/ext/sources/CMakeLists.txt +41 -3
- data/ext/sources/CMakePresets.json +95 -0
- data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
- data/ext/sources/cmake/parakeet.pc.in +10 -0
- data/ext/sources/cmake/whisper-config.cmake.in +5 -40
- data/ext/sources/cmake/whisper.pc.in +1 -1
- data/ext/sources/examples/CMakeLists.txt +4 -2
- data/ext/sources/examples/bench/bench.cpp +24 -19
- data/ext/sources/examples/cli/cli.cpp +51 -9
- data/ext/sources/examples/common-ggml.cpp +4 -0
- data/ext/sources/examples/common-whisper.cpp +139 -67
- data/ext/sources/examples/common-whisper.h +11 -0
- data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
- data/ext/sources/examples/miniaudio.h +4507 -2131
- data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
- data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
- data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
- data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
- data/ext/sources/examples/server/server.cpp +213 -163
- data/ext/sources/ggml/CMakeLists.txt +29 -15
- data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
- data/ext/sources/ggml/include/ggml-alloc.h +1 -0
- data/ext/sources/ggml/include/ggml-backend.h +73 -11
- data/ext/sources/ggml/include/ggml-cann.h +1 -1
- data/ext/sources/ggml/include/ggml-cpu.h +5 -0
- data/ext/sources/ggml/include/ggml-cuda.h +3 -0
- data/ext/sources/ggml/include/ggml-openvino.h +37 -0
- data/ext/sources/ggml/include/ggml-opt.h +1 -1
- data/ext/sources/ggml/include/ggml-rpc.h +8 -3
- data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
- data/ext/sources/ggml/include/ggml.h +155 -16
- data/ext/sources/ggml/include/gguf.h +10 -2
- data/ext/sources/ggml/src/CMakeLists.txt +25 -5
- data/ext/sources/ggml/src/ggml-alloc.c +9 -10
- data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
- data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
- data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
- data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +40 -86
- data/ext/sources/ggml/src/ggml-backend.cpp +114 -10
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +10 -2
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +1016 -442
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +111 -85
- data/ext/sources/ggml/src/ggml-cann/common.h +23 -14
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +255 -92
- data/ext/sources/ggml/src/ggml-common.h +22 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +68 -34
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +44 -19
- data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +101 -101
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2874 -613
- data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +5480 -840
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1361 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -11
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +186 -36
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +119 -19
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +112 -26
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +153 -16
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +17 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +976 -251
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +671 -266
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1277 -263
- data/ext/sources/ggml/src/ggml-cpu/ops.h +4 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +95 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2893 -679
- data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +226 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +114 -19
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +54 -53
- data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +18 -8
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
- data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +73 -28
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +69 -41
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +359 -29
- data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
- data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +94 -27
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +20 -9
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +333 -85
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +632 -190
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +162 -49
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +43 -18
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +44 -14
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +241 -23
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
- data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +312 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1454 -599
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +397 -183
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +161 -88
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +522 -431
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +139 -72
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +608 -88
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +47 -79
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +134 -27
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +7 -17
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +244 -137
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
- data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
- data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
- data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +96 -40
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -5
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +202 -135
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +86 -2
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +111 -17
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +30 -2
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +84 -46
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1612 -753
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +51 -11
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +361 -261
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +294 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +753 -241
- data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +295 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +471 -296
- data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +159 -53
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +3 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +372 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +137 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +97 -14
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +163 -67
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +308 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +216 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +142 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -1348
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +547 -635
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3556 -1101
- data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +475 -269
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +94 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +222 -217
- data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +432 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +886 -117
- data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
- data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +40 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +28 -9
- data/ext/sources/ggml/src/ggml-impl.h +68 -1
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +409 -83
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +54 -5
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +254 -52
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +254 -23
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +756 -285
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +7 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +359 -133
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1867 -1123
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +71 -4
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +14127 -5314
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +97 -88
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +104 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1978 -67
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
- data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +195 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
- data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +178 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
- data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
- data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
- data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +985 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +380 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1132 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +956 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +149 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +47 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +40 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +317 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +257 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +86 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +880 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.h +143 -0
- data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
- data/ext/sources/ggml/src/ggml-quants.c +385 -119
- data/ext/sources/ggml/src/ggml-quants.h +6 -0
- data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
- data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
- data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
- data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +64 -91
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +4 -1
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
- data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +356 -11
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +184 -14
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +31 -1
- data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
- data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
- data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +77 -156
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1181 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +674 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +227 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
- data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
- data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +347 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1134 -236
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
- data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
- data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
- data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
- data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +72 -1
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
- data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
- data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
- data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +228 -53
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
- data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +12 -1
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3250 -940
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +533 -180
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +113 -68
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +412 -222
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +222 -83
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +22 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +51 -14
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +39 -63
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +13 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +27 -11
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -149
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3221 -97
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3493 -1997
- data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +115 -141
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +93 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -44
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +198 -230
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +234 -335
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +871 -42
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +149 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +36 -138
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +151 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +15 -40
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +39 -12
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +24 -15
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +253 -16
- data/ext/sources/ggml/src/ggml.c +268 -52
- data/ext/sources/ggml/src/gguf.cpp +377 -47
- data/ext/sources/include/parakeet.h +342 -0
- data/ext/sources/include/whisper.h +10 -0
- data/ext/sources/media/matmul.png +0 -0
- data/ext/sources/src/CMakeLists.txt +23 -0
- data/ext/sources/src/parakeet-arch.h +188 -0
- data/ext/sources/src/parakeet.cpp +3838 -0
- data/ext/sources/src/whisper.cpp +62 -40
- data/extsources.rb +26 -10
- data/lib/whisper/log_settable.rb +36 -0
- data/lib/whisper/model/uri.rb +13 -1
- data/lib/whisper/output.rb +74 -0
- data/sig/whisper.rbs +445 -55
- data/test/helper.rb +2 -0
- data/test/jfk_reader/jfk_reader.c +50 -7
- data/test/test_callback.rb +1 -0
- data/test/test_context_params.rb +82 -0
- data/test/test_package.rb +6 -5
- data/test/test_parakeet.rb +28 -0
- data/test/test_parakeet_callback.rb +107 -0
- data/test/test_parakeet_context.rb +116 -0
- data/test/test_parakeet_context_params.rb +24 -0
- data/test/test_parakeet_model.rb +21 -0
- data/test/test_parakeet_params.rb +78 -0
- data/test/test_parakeet_segment.rb +42 -0
- data/test/test_parakeet_token.rb +73 -0
- data/test/test_params.rb +2 -0
- data/test/test_token.rb +11 -0
- data/test/test_vad_context.rb +58 -8
- data/test/test_vad_segment.rb +1 -1
- data/test/test_whisper.rb +44 -6
- data/whispercpp.gemspec +2 -2
- metadata +426 -280
- data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
- data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
- data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
- data/ext/sources/bindings/javascript/package.json +0 -26
- data/ext/sources/bindings/javascript/whisper.js +0 -19
- data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
- data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
- data/ext/sources/examples/addon.node/addon.cpp +0 -557
- data/ext/sources/examples/addon.node/index.js +0 -59
- data/ext/sources/examples/addon.node/package.json +0 -16
- data/ext/sources/examples/addon.node/vad-example.js +0 -132
- data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
- data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
- data/ext/sources/examples/coi-serviceworker.js +0 -146
- data/ext/sources/examples/command/CMakeLists.txt +0 -10
- data/ext/sources/examples/command/command.cpp +0 -802
- data/ext/sources/examples/command/commands.txt +0 -9
- data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
- data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
- data/ext/sources/examples/generate-karaoke.sh +0 -57
- data/ext/sources/examples/helpers.js +0 -191
- data/ext/sources/examples/livestream.sh +0 -112
- data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
- data/ext/sources/examples/lsp/lsp.cpp +0 -471
- data/ext/sources/examples/lsp/whisper.vim +0 -362
- data/ext/sources/examples/python/test_whisper_processor.py +0 -7
- data/ext/sources/examples/python/whisper_processor.py +0 -54
- data/ext/sources/examples/server/bench.js +0 -29
- data/ext/sources/examples/server.py +0 -120
- data/ext/sources/examples/stream/CMakeLists.txt +0 -10
- data/ext/sources/examples/stream/stream.cpp +0 -437
- data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
- data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
- data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
- data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
- data/ext/sources/examples/sycl/build.sh +0 -22
- data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
- data/ext/sources/examples/sycl/run-whisper.sh +0 -17
- data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -47
- data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -494
- data/ext/sources/examples/talk-llama/llama-adapter.h +0 -88
- data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2559
- data/ext/sources/examples/talk-llama/llama-arch.h +0 -586
- data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -917
- data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
- data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -876
- data/ext/sources/examples/talk-llama/llama-chat.h +0 -70
- data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3645
- data/ext/sources/examples/talk-llama/llama-context.h +0 -360
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
- data/ext/sources/examples/talk-llama/llama-cparams.h +0 -42
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
- data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
- data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2282
- data/ext/sources/examples/talk-llama/llama-graph.h +0 -910
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -241
- data/ext/sources/examples/talk-llama/llama-hparams.h +0 -284
- data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
- data/ext/sources/examples/talk-llama/llama-impl.h +0 -63
- data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
- data/ext/sources/examples/talk-llama/llama-io.h +0 -35
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -328
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2100
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -390
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1167
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
- data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
- data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -735
- data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1247
- data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -176
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -285
- data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -37
- data/ext/sources/examples/talk-llama/llama-model.cpp +0 -8338
- data/ext/sources/examples/talk-llama/llama-model.h +0 -544
- data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1072
- data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +0 -3771
- data/ext/sources/examples/talk-llama/llama-sampling.h +0 -44
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3900
- data/ext/sources/examples/talk-llama/llama-vocab.h +0 -182
- data/ext/sources/examples/talk-llama/llama.cpp +0 -1140
- data/ext/sources/examples/talk-llama/llama.h +0 -1540
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -191
- data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
- data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -138
- data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/bert.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
- data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
- data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
- data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
- data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -259
- data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -134
- data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
- data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -113
- data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
- data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
- data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
- data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
- data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
- data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
- data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
- data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -196
- data/ext/sources/examples/talk-llama/models/granite.cpp +0 -211
- data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +0 -283
- data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -141
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -154
- data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
- data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -175
- data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
- data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
- data/ext/sources/examples/talk-llama/models/llama.cpp +0 -168
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -55
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -199
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
- data/ext/sources/examples/talk-llama/models/models.h +0 -569
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -116
- data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
- data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
- data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
- data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
- data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
- data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
- data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
- data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -316
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/plm.cpp +0 -168
- data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
- data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
- data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -117
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -873
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -149
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -141
- data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -162
- data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
- data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
- data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -135
- data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
- data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
- data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
- data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
- data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
- data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
- data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
- data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
- data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
- data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
- data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
- data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
- data/ext/sources/examples/talk-llama/speak +0 -40
- data/ext/sources/examples/talk-llama/speak.bat +0 -1
- data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
- data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
- data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
- data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
- data/ext/sources/examples/talk-llama/unicode.cpp +0 -1147
- data/ext/sources/examples/talk-llama/unicode.h +0 -111
- data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
- data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
- data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
- data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
- data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
- data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
- data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
- data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
- data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
- data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
- data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +0 -157
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -165
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
- data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
- data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
- data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -147
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +0 -907
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +0 -247
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
- data/ext/sources/tests/CMakeLists.txt +0 -112
- data/ext/sources/tests/earnings21/eval.mk +0 -58
- data/ext/sources/tests/earnings21/eval.py +0 -68
- data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
- data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
- data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
- data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
- data/ext/sources/tests/earnings21/requirements.txt +0 -6
- data/ext/sources/tests/en-0-ref.txt +0 -1
- data/ext/sources/tests/en-1-ref.txt +0 -1
- data/ext/sources/tests/en-2-ref.txt +0 -1
- data/ext/sources/tests/es-0-ref.txt +0 -1
- data/ext/sources/tests/librispeech/eval.mk +0 -39
- data/ext/sources/tests/librispeech/eval.py +0 -47
- data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
- data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
- data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
- data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
- data/ext/sources/tests/librispeech/requirements.txt +0 -6
- data/ext/sources/tests/run-tests.sh +0 -130
- data/ext/sources/tests/test-c.c +0 -3
- data/ext/sources/tests/test-vad-full.cpp +0 -56
- data/ext/sources/tests/test-vad.cpp +0 -83
- data/ext/sources/tests/test-whisper.js +0 -58
- data/lib/whisper/context.rb +0 -15
- data/lib/whisper/segment.rb +0 -58
|
@@ -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 +62,26 @@ 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, 4, 128, 2, 64, 64)
|
|
80
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
81
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
|
|
82
|
+
|
|
83
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
84
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
71
85
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
|
|
72
86
|
|
|
73
87
|
return 0;
|
|
@@ -116,12 +130,26 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
|
|
|
116
130
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 128, 3, 32, 128)
|
|
117
131
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
|
|
118
132
|
|
|
133
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 128, 3, 64, 64)
|
|
134
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 3, 32, 64)
|
|
135
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 32, 64)
|
|
136
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
|
|
137
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
|
|
138
|
+
|
|
119
139
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 128, 3, 64, 64)
|
|
120
140
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 3, 32, 64)
|
|
121
141
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 32, 256)
|
|
122
142
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
|
|
123
143
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64)
|
|
124
144
|
|
|
145
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 32, 64)
|
|
146
|
+
|
|
147
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 32, 64)
|
|
148
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 32, 64)
|
|
149
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 32, 64)
|
|
150
|
+
|
|
151
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 32, 64)
|
|
152
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 32, 64)
|
|
125
153
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 32, 64)
|
|
126
154
|
|
|
127
155
|
return 0;
|
|
@@ -177,12 +205,27 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
|
|
|
177
205
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
|
|
178
206
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2, 64, 32)
|
|
179
207
|
|
|
208
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 256, 2, 128, 64)
|
|
209
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 256, 2, 64, 64)
|
|
210
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 64, 64)
|
|
211
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
|
|
212
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
|
|
213
|
+
|
|
180
214
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 256, 2, 128, 64)
|
|
181
215
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 256, 2, 64, 128)
|
|
182
216
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 64, 128)
|
|
183
217
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
|
|
184
218
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128)
|
|
185
219
|
|
|
220
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 512, 1, 128, 64)
|
|
221
|
+
|
|
222
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
|
|
223
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
224
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
|
|
225
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 512, 1, 128, 64)
|
|
226
|
+
|
|
227
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
228
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
186
229
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
|
|
187
230
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128, 64)
|
|
188
231
|
|
|
@@ -239,12 +282,27 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
|
|
|
239
282
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128, 64)
|
|
240
283
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3, 64, 64)
|
|
241
284
|
|
|
285
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 64, 8, 32, 64)
|
|
286
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 6, 32, 64)
|
|
287
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 128, 6, 32, 64)
|
|
288
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 5, 32, 64)
|
|
289
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 3, 64, 64)
|
|
290
|
+
|
|
242
291
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 64, 8, 32, 64)
|
|
243
292
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 6, 32, 256)
|
|
244
293
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 128, 6, 32, 256)
|
|
245
294
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256)
|
|
246
295
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128)
|
|
247
296
|
|
|
297
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 256, 2, 128, 64)
|
|
298
|
+
|
|
299
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
|
|
300
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
|
|
301
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4, 64, 64)
|
|
302
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2, 128, 64)
|
|
303
|
+
|
|
304
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
|
|
305
|
+
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
|
|
248
306
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4, 64, 64)
|
|
249
307
|
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128, 64)
|
|
250
308
|
|
|
@@ -343,7 +401,7 @@ static __device__ __forceinline__ void flash_attn_tile_load_tile(
|
|
|
343
401
|
for (int j0 = j0_start; j0 < j0_stop; j0 += stride_j) {
|
|
344
402
|
const int j = j0*cpy_ne + (stride_j == warp_size ? threadIdx.x : threadIdx.x % stride_j)*cpy_ne;
|
|
345
403
|
|
|
346
|
-
const half2 zero[cpy_ne] = {{0.0f, 0.0f}};
|
|
404
|
+
const __align__(16) half2 zero[cpy_ne] = {{0.0f, 0.0f}};
|
|
347
405
|
ggml_cuda_memcpy_1<cpy_nb>(
|
|
348
406
|
tile_KV + i*(J/2 + J_padding) + j,
|
|
349
407
|
!oob_check || i < i_sup ? KV + i*stride_KV + j : zero);
|
|
@@ -394,11 +452,11 @@ static __device__ __forceinline__ void flash_attn_tile_load_tile(
|
|
|
394
452
|
const int j = j0*(cpy_ne/2) + (stride_j == warp_size ? threadIdx.x : threadIdx.x % stride_j)*(cpy_ne/2);
|
|
395
453
|
|
|
396
454
|
const half2 zero[cpy_ne/2] = {{0.0f, 0.0f}};
|
|
397
|
-
half2 tmp_h2[cpy_ne/2];
|
|
455
|
+
__align__(16) half2 tmp_h2[cpy_ne/2];
|
|
398
456
|
ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
|
|
399
457
|
tmp_h2, !oob_check || i < i_sup ? KV + i*stride_KV + j : zero);
|
|
400
458
|
|
|
401
|
-
float2 tmp_f2[cpy_ne/2];
|
|
459
|
+
__align__(16) float2 tmp_f2[cpy_ne/2];
|
|
402
460
|
#pragma unroll
|
|
403
461
|
for (int l = 0; l < cpy_ne/2; ++l) {
|
|
404
462
|
tmp_f2[l] = __half22float2(tmp_h2[l]);
|
|
@@ -445,14 +503,14 @@ static __device__ __forceinline__ void flash_attn_tile_iter_KQ(
|
|
|
445
503
|
static_assert((nbatch_K/2) % cpy_ne == 0, "bad nbatch_K");
|
|
446
504
|
#pragma unroll
|
|
447
505
|
for (int k_KQ_1 = 0; k_KQ_1 < nbatch_K/2; k_KQ_1 += cpy_ne) {
|
|
448
|
-
half2 K_k[nbatch_fa/(np*warp_size)][cpy_ne];
|
|
449
|
-
half2 Q_k[cpw][cpy_ne];
|
|
506
|
+
__align__(16) half2 K_k[nbatch_fa/(np*warp_size)][cpy_ne];
|
|
507
|
+
__align__(16) half2 Q_k[cpw][cpy_ne];
|
|
450
508
|
#else
|
|
451
509
|
static_assert(nbatch_K % cpy_ne == 0, "bad nbatch_K");
|
|
452
510
|
#pragma unroll
|
|
453
511
|
for (int k_KQ_1 = 0; k_KQ_1 < nbatch_K; k_KQ_1 += cpy_ne) {
|
|
454
|
-
float K_k[nbatch_fa/(np*warp_size)][cpy_ne];
|
|
455
|
-
float Q_k[cpw][cpy_ne];
|
|
512
|
+
__align__(16) float K_k[nbatch_fa/(np*warp_size)][cpy_ne];
|
|
513
|
+
__align__(16) float Q_k[cpw][cpy_ne];
|
|
456
514
|
#endif // FAST_FP16_AVAILABLE
|
|
457
515
|
|
|
458
516
|
#pragma unroll
|
|
@@ -602,9 +660,9 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
602
660
|
#pragma unroll
|
|
603
661
|
for (int jc0 = 0; jc0 < cpw; jc0 += KQ_cs) {
|
|
604
662
|
#ifdef FAST_FP16_AVAILABLE
|
|
605
|
-
half tmp[nbatch_fa/(np*warp_size)][KQ_cs];
|
|
663
|
+
__align__(16) half tmp[nbatch_fa/(np*warp_size)][KQ_cs];
|
|
606
664
|
#else
|
|
607
|
-
float tmp[nbatch_fa/(np*warp_size)][KQ_cs];
|
|
665
|
+
__align__(16) float tmp[nbatch_fa/(np*warp_size)][KQ_cs];
|
|
608
666
|
#endif // FAST_FP16_AVAILABLE
|
|
609
667
|
|
|
610
668
|
#pragma unroll
|
|
@@ -664,8 +722,8 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
664
722
|
#ifdef FAST_FP16_AVAILABLE
|
|
665
723
|
#pragma unroll
|
|
666
724
|
for (int k1 = 0; k1 < nbatch_V; k1 += np) {
|
|
667
|
-
half2 V_k[(DVp/2)/warp_size];
|
|
668
|
-
half2 KQ_k[cpw];
|
|
725
|
+
__align__(16) half2 V_k[(DVp/2)/warp_size];
|
|
726
|
+
__align__(16) half2 KQ_k[cpw];
|
|
669
727
|
|
|
670
728
|
constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;
|
|
671
729
|
#pragma unroll
|
|
@@ -676,7 +734,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
676
734
|
for (int jc_VKQ_0 = 0; jc_VKQ_0 < cpw; jc_VKQ_0 += KQ_cs) {
|
|
677
735
|
const int jc_KQ = jc_VKQ_0/KQ_cs + (threadIdx.y / np)*(cpw/KQ_cs);
|
|
678
736
|
|
|
679
|
-
half tmp[KQ_cs];
|
|
737
|
+
__align__(16) half tmp[KQ_cs];
|
|
680
738
|
ggml_cuda_memcpy_1<KQ_cs*sizeof(half)>(
|
|
681
739
|
&tmp, KQ + jc_KQ*(nbatch_fa*KQ_cs) + (k0 + k1 + threadIdx.y % np)*KQ_cs);
|
|
682
740
|
#pragma unroll
|
|
@@ -696,8 +754,8 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
696
754
|
#else
|
|
697
755
|
#pragma unroll
|
|
698
756
|
for (int k1 = 0; k1 < nbatch_V; k1 += np) {
|
|
699
|
-
float2 V_k[(DVp/2)/warp_size];
|
|
700
|
-
float KQ_k[cpw];
|
|
757
|
+
__align__(16) float2 V_k[(DVp/2)/warp_size];
|
|
758
|
+
__align__(16) float KQ_k[cpw];
|
|
701
759
|
|
|
702
760
|
constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
|
|
703
761
|
#pragma unroll
|
|
@@ -730,14 +788,14 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
|
|
|
730
788
|
template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap> // D == head size
|
|
731
789
|
__launch_bounds__(ggml_cuda_fattn_tile_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_tile_get_occupancy(DKQ, DV, ncols1*ncols2))
|
|
732
790
|
static __global__ void flash_attn_tile(
|
|
733
|
-
const char *
|
|
734
|
-
const char *
|
|
735
|
-
const char *
|
|
736
|
-
const char *
|
|
737
|
-
const char *
|
|
738
|
-
const int *
|
|
739
|
-
float *
|
|
740
|
-
float2 *
|
|
791
|
+
const char * Q_ptr,
|
|
792
|
+
const char * K_ptr,
|
|
793
|
+
const char * V_ptr,
|
|
794
|
+
const char * mask_ptr,
|
|
795
|
+
const char * sinks_ptr,
|
|
796
|
+
const int * KV_max_ptr,
|
|
797
|
+
float * dst_ptr,
|
|
798
|
+
float2 * dst_meta_ptr,
|
|
741
799
|
const float scale,
|
|
742
800
|
const float max_bias,
|
|
743
801
|
const float m0,
|
|
@@ -752,6 +810,14 @@ static __global__ void flash_attn_tile(
|
|
|
752
810
|
const int32_t ne31, const int32_t ne32, const int32_t ne33,
|
|
753
811
|
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
|
|
754
812
|
#ifdef FLASH_ATTN_AVAILABLE
|
|
813
|
+
const char * GGML_CUDA_RESTRICT Q = Q_ptr;
|
|
814
|
+
const char * GGML_CUDA_RESTRICT K = K_ptr;
|
|
815
|
+
const char * GGML_CUDA_RESTRICT V = V_ptr;
|
|
816
|
+
const char * GGML_CUDA_RESTRICT mask = mask_ptr;
|
|
817
|
+
const char * GGML_CUDA_RESTRICT sinks = sinks_ptr;
|
|
818
|
+
const int * GGML_CUDA_RESTRICT KV_max = KV_max_ptr;
|
|
819
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
820
|
+
float2 * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
|
|
755
821
|
|
|
756
822
|
// Skip unused kernel variants for faster compilation:
|
|
757
823
|
|
|
@@ -759,7 +825,7 @@ static __global__ void flash_attn_tile(
|
|
|
759
825
|
#ifdef GGML_USE_WMMA_FATTN
|
|
760
826
|
(ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
|
|
761
827
|
#endif // GGML_USE_WMMA_FATTN
|
|
762
|
-
(use_logit_softcap && !(DV == 128 || DV == 256))
|
|
828
|
+
(use_logit_softcap && !(DV == 128 || DV == 256 || DV == 512))
|
|
763
829
|
) {
|
|
764
830
|
GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
|
|
765
831
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
@@ -821,12 +887,12 @@ static __global__ void flash_attn_tile(
|
|
|
821
887
|
__shared__ half2 Q_tmp[ncols * DKQ/2];
|
|
822
888
|
__shared__ half2 KV_tmp[nbatch_fa * (nbatch_K/2 + cpy_ne) + DVp-DV];
|
|
823
889
|
__shared__ half KQ[ncols * nbatch_fa];
|
|
824
|
-
half2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
|
|
890
|
+
__align__(16) half2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
|
|
825
891
|
#else
|
|
826
892
|
__shared__ float Q_tmp[ncols * DKQ];
|
|
827
893
|
__shared__ float KV_tmp[nbatch_fa * (nbatch_K + cpy_ne) + DVp-DV];
|
|
828
894
|
__shared__ float KQ[ncols * nbatch_fa];
|
|
829
|
-
float2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
|
|
895
|
+
__align__(16) float2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
|
|
830
896
|
#endif // FAST_FP16_AVAILABLE
|
|
831
897
|
|
|
832
898
|
float KQ_max[cpw];
|
|
@@ -836,6 +902,8 @@ static __global__ void flash_attn_tile(
|
|
|
836
902
|
}
|
|
837
903
|
float KQ_sum[cpw] = {0.0f};
|
|
838
904
|
|
|
905
|
+
ggml_cuda_pdl_sync();
|
|
906
|
+
|
|
839
907
|
// Load Q data, convert to FP16 if fast:
|
|
840
908
|
#pragma unroll
|
|
841
909
|
for (int jc0 = 0; jc0 < cpw; ++jc0) {
|
|
@@ -849,7 +917,7 @@ static __global__ void flash_attn_tile(
|
|
|
849
917
|
#pragma unroll
|
|
850
918
|
for (int i0 = 0; i0 < DKQp; i0 += np*warp_size*cpy_ne_D) {
|
|
851
919
|
if (i0 + np*warp_size*cpy_ne_D <= DKQ || i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D < DKQ) {
|
|
852
|
-
float tmp_f[cpy_ne_D] = {0.0f};
|
|
920
|
+
__align__(16) float tmp_f[cpy_ne_D] = {0.0f};
|
|
853
921
|
ggml_cuda_memcpy_1<sizeof(tmp_f)>
|
|
854
922
|
(tmp_f, &Q_f[c*(nb02/sizeof(float)) + fastmodulo(col_Q_0 + j, ne01)*(nb01/sizeof(float))
|
|
855
923
|
+ i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D]);
|
|
@@ -860,7 +928,7 @@ static __global__ void flash_attn_tile(
|
|
|
860
928
|
}
|
|
861
929
|
|
|
862
930
|
#ifdef FAST_FP16_AVAILABLE
|
|
863
|
-
half2 tmp_h2[cpy_ne_D/2];
|
|
931
|
+
__align__(16) half2 tmp_h2[cpy_ne_D/2];
|
|
864
932
|
#pragma unroll
|
|
865
933
|
for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
|
|
866
934
|
tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
|
|
@@ -959,7 +1027,7 @@ static __global__ void flash_attn_tile(
|
|
|
959
1027
|
constexpr int cpy_ne_D = cpy_ne < (DVp/2)/warp_size ? cpy_ne : (DVp/2)/warp_size;
|
|
960
1028
|
#pragma unroll
|
|
961
1029
|
for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
|
|
962
|
-
half2 tmp[cpy_ne_D];
|
|
1030
|
+
__align__(16) half2 tmp[cpy_ne_D];
|
|
963
1031
|
ggml_cuda_memcpy_1<cpy_ne_D*4>(tmp, &VKQ_combine[(threadIdx.y + ip)*(DVp/2) + i0 + threadIdx.x*cpy_ne_D]);
|
|
964
1032
|
#pragma unroll
|
|
965
1033
|
for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
|
|
@@ -970,7 +1038,7 @@ static __global__ void flash_attn_tile(
|
|
|
970
1038
|
constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
|
|
971
1039
|
#pragma unroll
|
|
972
1040
|
for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
|
|
973
|
-
float tmp[cpy_ne_D];
|
|
1041
|
+
__align__(16) float tmp[cpy_ne_D];
|
|
974
1042
|
ggml_cuda_memcpy_1<cpy_ne_D*4>(tmp, &VKQ_combine[(threadIdx.y + ip)*DVp + i0 + threadIdx.x*cpy_ne_D]);
|
|
975
1043
|
#pragma unroll
|
|
976
1044
|
for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
|
|
@@ -1033,7 +1101,7 @@ static __global__ void flash_attn_tile(
|
|
|
1033
1101
|
constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;
|
|
1034
1102
|
#pragma unroll
|
|
1035
1103
|
for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
|
|
1036
|
-
float2 tmp[cpy_ne_D];
|
|
1104
|
+
__align__(16) float2 tmp[cpy_ne_D];
|
|
1037
1105
|
#pragma unroll
|
|
1038
1106
|
for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
|
|
1039
1107
|
tmp[i1] = __half22float2(VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size + i1]);
|
|
@@ -1066,7 +1134,7 @@ static __global__ void flash_attn_tile(
|
|
|
1066
1134
|
}
|
|
1067
1135
|
}
|
|
1068
1136
|
#else
|
|
1069
|
-
GGML_UNUSED_VARS(
|
|
1137
|
+
GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
|
|
1070
1138
|
max_bias, m0, m1, n_head_log2, logit_softcap,
|
|
1071
1139
|
ne00, ne01, ne02, ne03,
|
|
1072
1140
|
nb01, nb02, nb03,
|
|
@@ -1090,7 +1158,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1090
1158
|
constexpr size_t nbytes_shared = 0;
|
|
1091
1159
|
|
|
1092
1160
|
#ifdef GGML_USE_HIP
|
|
1093
|
-
if constexpr (
|
|
1161
|
+
if constexpr (DKQ <= 128) {
|
|
1094
1162
|
if (Q->ne[1] > 32/ncols2) {
|
|
1095
1163
|
constexpr int cols_per_block = 64;
|
|
1096
1164
|
const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
|
|
@@ -1104,7 +1172,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1104
1172
|
#endif // GGML_USE_HIP
|
|
1105
1173
|
|
|
1106
1174
|
#ifndef GGML_USE_HIP
|
|
1107
|
-
if constexpr (
|
|
1175
|
+
if constexpr (DKQ <= 256)
|
|
1108
1176
|
#endif // GGML_USE_HIP
|
|
1109
1177
|
{
|
|
1110
1178
|
if (Q->ne[1] > 16/ncols2) {
|
|
@@ -1118,14 +1186,16 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
|
|
|
1118
1186
|
}
|
|
1119
1187
|
}
|
|
1120
1188
|
|
|
1121
|
-
if (
|
|
1122
|
-
|
|
1123
|
-
|
|
1124
|
-
|
|
1125
|
-
|
|
1126
|
-
|
|
1127
|
-
|
|
1128
|
-
|
|
1189
|
+
if constexpr (ncols2 <= 16) {
|
|
1190
|
+
if (Q->ne[1] > 8/ncols2) {
|
|
1191
|
+
constexpr int cols_per_block = 16;
|
|
1192
|
+
const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
|
|
1193
|
+
const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
|
|
1194
|
+
fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
|
|
1195
|
+
launch_fattn<DV, cols_per_block/ncols2, ncols2>
|
|
1196
|
+
(ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
|
|
1197
|
+
return;
|
|
1198
|
+
}
|
|
1129
1199
|
}
|
|
1130
1200
|
|
|
1131
1201
|
if constexpr (ncols2 <= 8) {
|
|
@@ -1178,18 +1248,56 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
|
|
|
1178
1248
|
GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
|
|
1179
1249
|
const int gqa_ratio = Q->ne[2] / K->ne[2];
|
|
1180
1250
|
|
|
1251
|
+
// On NVIDIA (Pascal and older) the GQA optimizations seem to be detrimental in some cases.
|
|
1252
|
+
// However, for DKQ == 576, DV == 512 only the kernel variant with GQA optimizations is implemented.
|
|
1181
1253
|
const bool nvidia = GGML_CUDA_CC_IS_NVIDIA(ggml_cuda_info().devices[ggml_cuda_get_device()].cc);
|
|
1182
|
-
const int gqa_limit = nvidia && gqa_ratio <= 4 ? 16 : INT_MAX;
|
|
1254
|
+
const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
|
|
1183
1255
|
const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
|
|
1184
1256
|
|
|
1185
|
-
if constexpr (
|
|
1257
|
+
if constexpr (DKQ == 320) {
|
|
1258
|
+
// This branch is only used for Mistral Small 4 which has a GQA ratio of 32.
|
|
1259
|
+
// On AMD, simply use that GQA ratio with 32 columns / block since we always have enough SRAM.
|
|
1260
|
+
// On NVIDIA however, the tile kernel is only used for GPUs that can't use the mma kernel (Pascal and older).
|
|
1261
|
+
// Therefore, use a GQA ratio of 16 with 16 columns / block to stay below 48 kiB of SRAM / block.
|
|
1262
|
+
#ifdef GGML_USE_HIP
|
|
1263
|
+
if (use_gqa_opt && gqa_ratio % 32 == 0) {
|
|
1264
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 32, use_logit_softcap>(ctx, dst);
|
|
1265
|
+
return;
|
|
1266
|
+
}
|
|
1267
|
+
#else
|
|
1268
|
+
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1269
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1270
|
+
return;
|
|
1271
|
+
}
|
|
1272
|
+
#endif // GGML_USE_HIP
|
|
1273
|
+
GGML_ABORT("flash-attn tile (320/256): expected GQA ratio multiple of 32");
|
|
1274
|
+
}
|
|
1275
|
+
|
|
1276
|
+
if constexpr (DKQ == 576) {
|
|
1186
1277
|
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1187
1278
|
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1188
1279
|
return;
|
|
1189
1280
|
}
|
|
1281
|
+
if (use_gqa_opt && gqa_ratio % 4 == 0) {
|
|
1282
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
|
|
1283
|
+
return;
|
|
1284
|
+
}
|
|
1190
1285
|
}
|
|
1191
1286
|
|
|
1192
|
-
if constexpr (
|
|
1287
|
+
if constexpr (DKQ == 192) {
|
|
1288
|
+
// MiMo-V2.5 / V2.5-Pro / V2-Flash: gqa_ratio is 8 (SWA) or 16 (full attn)
|
|
1289
|
+
if (use_gqa_opt && gqa_ratio % 16 == 0) {
|
|
1290
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
|
|
1291
|
+
return;
|
|
1292
|
+
}
|
|
1293
|
+
if (use_gqa_opt && gqa_ratio % 8 == 0) {
|
|
1294
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
|
|
1295
|
+
return;
|
|
1296
|
+
}
|
|
1297
|
+
GGML_ABORT("flash-attn tile (192/128): expected GQA ratio multiple of 8");
|
|
1298
|
+
}
|
|
1299
|
+
|
|
1300
|
+
if constexpr (DKQ <= 512 && DKQ != 320 && DKQ != 192) {
|
|
1193
1301
|
if (use_gqa_opt && gqa_ratio % 8 == 0) {
|
|
1194
1302
|
launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
|
|
1195
1303
|
return;
|
|
@@ -1200,13 +1308,15 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
|
|
|
1200
1308
|
return;
|
|
1201
1309
|
}
|
|
1202
1310
|
|
|
1203
|
-
if (
|
|
1204
|
-
|
|
1311
|
+
if constexpr (DV <= 256) {
|
|
1312
|
+
if (use_gqa_opt && gqa_ratio % 2 == 0) {
|
|
1313
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
|
|
1314
|
+
return;
|
|
1315
|
+
}
|
|
1316
|
+
|
|
1317
|
+
launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
|
|
1205
1318
|
return;
|
|
1206
1319
|
}
|
|
1207
|
-
|
|
1208
|
-
launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
|
|
1209
|
-
return;
|
|
1210
1320
|
}
|
|
1211
1321
|
GGML_ABORT("fatal error");
|
|
1212
1322
|
}
|
|
@@ -1240,5 +1350,8 @@ extern DECL_FATTN_TILE_CASE( 80, 80);
|
|
|
1240
1350
|
extern DECL_FATTN_TILE_CASE( 96, 96);
|
|
1241
1351
|
extern DECL_FATTN_TILE_CASE(112, 112);
|
|
1242
1352
|
extern DECL_FATTN_TILE_CASE(128, 128);
|
|
1353
|
+
extern DECL_FATTN_TILE_CASE(192, 128);
|
|
1243
1354
|
extern DECL_FATTN_TILE_CASE(256, 256);
|
|
1355
|
+
extern DECL_FATTN_TILE_CASE(320, 256);
|
|
1356
|
+
extern DECL_FATTN_TILE_CASE(512, 512);
|
|
1244
1357
|
extern DECL_FATTN_TILE_CASE(576, 512);
|
|
@@ -10,7 +10,7 @@ static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
|
|
|
10
10
|
return 128;
|
|
11
11
|
}
|
|
12
12
|
|
|
13
|
-
//
|
|
13
|
+
// Currently llvm with the amdgcn target does not support unrolling loops
|
|
14
14
|
// that contain a break that can not be resolved at compile time.
|
|
15
15
|
#ifdef __clang__
|
|
16
16
|
#pragma clang diagnostic push
|
|
@@ -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
|
|
@@ -132,10 +141,12 @@ static __global__ void flash_attn_ext_vec(
|
|
|
132
141
|
#ifdef V_DOT2_F32_F16_AVAILABLE
|
|
133
142
|
half2 Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
|
|
134
143
|
#else
|
|
135
|
-
float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
|
|
144
|
+
__align__(16) float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
|
|
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) {
|
|
@@ -200,7 +211,7 @@ static __global__ void flash_attn_ext_vec(
|
|
|
200
211
|
for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
|
|
201
212
|
const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
|
|
202
213
|
|
|
203
|
-
float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
|
|
214
|
+
__align__(16) float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
|
|
204
215
|
if (ncols == 1 || ic0 + j < int(ne01.z)) {
|
|
205
216
|
ggml_cuda_memcpy_1<cpy_nb>(tmp, &Q_j[i]);
|
|
206
217
|
ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);
|
|
@@ -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)
|