whispercpp 1.3.3 → 1.3.4
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/ext/ruby_whisper_params.c +55 -25
- data/ext/sources/CMakeLists.txt +1 -1
- data/ext/sources/bindings/javascript/package.json +1 -1
- data/ext/sources/build-xcframework.sh +24 -0
- data/ext/sources/examples/CMakeLists.txt +1 -0
- data/ext/sources/examples/addon.node/addon.cpp +19 -19
- data/ext/sources/examples/addon.node/index.js +7 -5
- data/ext/sources/examples/bench/bench.cpp +26 -16
- data/ext/sources/examples/bench.wasm/index-tmpl.html +10 -9
- data/ext/sources/examples/cli/cli.cpp +4 -2
- data/ext/sources/examples/command/command.cpp +26 -24
- data/ext/sources/examples/command.wasm/index-tmpl.html +5 -4
- data/ext/sources/examples/common-ggml.cpp +2 -0
- data/ext/sources/examples/lsp/lsp.cpp +19 -17
- data/ext/sources/examples/server/server.cpp +24 -13
- data/ext/sources/examples/server.py +6 -1
- data/ext/sources/examples/stream/stream.cpp +4 -2
- data/ext/sources/examples/stream.wasm/emscripten.cpp +6 -6
- data/ext/sources/examples/stream.wasm/index-tmpl.html +82 -5
- data/ext/sources/examples/talk-llama/CMakeLists.txt +2 -2
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +101 -4
- data/ext/sources/examples/talk-llama/llama-adapter.h +6 -0
- data/ext/sources/examples/talk-llama/llama-arch.cpp +588 -15
- data/ext/sources/examples/talk-llama/llama-arch.h +58 -1
- data/ext/sources/examples/talk-llama/llama-batch.cpp +103 -71
- data/ext/sources/examples/talk-llama/llama-batch.h +31 -18
- data/ext/sources/examples/talk-llama/llama-chat.cpp +120 -5
- data/ext/sources/examples/talk-llama/llama-chat.h +7 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +460 -357
- data/ext/sources/examples/talk-llama/llama-context.h +44 -29
- data/ext/sources/examples/talk-llama/llama-cparams.h +4 -4
- data/ext/sources/examples/talk-llama/llama-graph.cpp +543 -271
- data/ext/sources/examples/talk-llama/llama-graph.h +278 -168
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +118 -4
- data/ext/sources/examples/talk-llama/llama-hparams.h +61 -15
- data/ext/sources/examples/talk-llama/llama-impl.h +2 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +326 -0
- data/ext/sources/examples/talk-llama/{llama-kv-cache-unified-iswa.h → llama-kv-cache-iswa.h} +38 -29
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2020 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +358 -27
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +80 -28
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +56 -36
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +30 -29
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +48 -19
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +13 -14
- data/ext/sources/examples/talk-llama/llama-memory.h +16 -10
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +2 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.h +3 -2
- data/ext/sources/examples/talk-llama/llama-model.cpp +7165 -2336
- data/ext/sources/examples/talk-llama/llama-model.h +60 -9
- data/ext/sources/examples/talk-llama/llama-quant.cpp +48 -10
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +226 -126
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +440 -13
- data/ext/sources/examples/talk-llama/llama-vocab.h +45 -0
- data/ext/sources/examples/talk-llama/llama.cpp +65 -10
- data/ext/sources/examples/talk-llama/llama.h +95 -177
- data/ext/sources/examples/talk-llama/talk-llama.cpp +9 -6
- data/ext/sources/examples/talk-llama/unicode.cpp +207 -0
- data/ext/sources/examples/talk-llama/unicode.h +45 -0
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +4 -2
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +17 -16
- data/ext/sources/ggml/CMakeLists.txt +59 -31
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +132 -93
- data/ext/sources/ggml/include/ggml-backend.h +17 -1
- data/ext/sources/ggml/include/ggml-cpu.h +1 -1
- data/ext/sources/ggml/include/ggml-metal.h +1 -6
- data/ext/sources/ggml/include/ggml-opt.h +25 -6
- data/ext/sources/ggml/include/ggml-webgpu.h +19 -0
- data/ext/sources/ggml/include/ggml-zdnn.h +17 -0
- data/ext/sources/ggml/include/ggml.h +221 -16
- data/ext/sources/ggml/src/CMakeLists.txt +17 -2
- data/ext/sources/ggml/src/ggml-alloc.c +265 -141
- data/ext/sources/ggml/src/ggml-backend-impl.h +4 -1
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +30 -13
- data/ext/sources/ggml/src/ggml-backend.cpp +221 -38
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +5 -4
- data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +14 -0
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +3 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +903 -717
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +143 -25
- data/ext/sources/ggml/src/ggml-cann/common.h +143 -1
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +488 -69
- data/ext/sources/ggml/src/ggml-common.h +17 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +40 -18
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +4 -2
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +132 -596
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +14 -286
- data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +103 -582
- data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +162 -589
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +265 -437
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +3 -58
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +521 -353
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +54 -314
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +184 -675
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +4679 -1657
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +32 -2
- data/ext/sources/ggml/src/ggml-cpu/common.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -6
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +70 -42
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +35 -28
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +152 -18
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +7 -1
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +227 -97
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +474 -1116
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1587 -1177
- data/ext/sources/ggml/src/ggml-cpu/ops.h +5 -8
- data/ext/sources/ggml/src/ggml-cpu/quants.c +35 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +458 -47
- data/ext/sources/ggml/src/ggml-cpu/repack.h +22 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +89 -60
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1024 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
- data/ext/sources/ggml/src/ggml-cpu/traits.cpp +2 -2
- data/ext/sources/ggml/src/ggml-cpu/traits.h +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +170 -26
- data/ext/sources/ggml/src/ggml-cpu/vec.h +506 -63
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +20 -16
- data/ext/sources/ggml/src/ggml-cuda/add-id.cu +58 -0
- data/ext/sources/ggml/src/ggml-cuda/add-id.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +330 -191
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +250 -63
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +1 -4
- data/ext/sources/ggml/src/ggml-cuda/conv2d.cu +166 -0
- data/ext/sources/ggml/src/ggml-cuda/conv2d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +95 -22
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +15 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +64 -307
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +2 -14
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +14 -40
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +498 -367
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +137 -91
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +755 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +593 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +86 -50
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +185 -198
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +50 -39
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +379 -107
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +196 -35
- data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +56 -2
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +198 -45
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +123 -0
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +496 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +206 -57
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +1262 -721
- data/ext/sources/ggml/src/ggml-cuda/{mmv.cu → mmvf.cu} +53 -53
- data/ext/sources/ggml/src/ggml-cuda/{mmv.cuh → mmvf.cuh} +3 -3
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +64 -73
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +284 -12
- data/ext/sources/ggml/src/ggml-cuda/norm.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +46 -23
- data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
- data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +12 -10
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +53 -0
- data/ext/sources/ggml/src/ggml-cuda/roll.cu +67 -0
- data/ext/sources/ggml/src/ggml-cuda/roll.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +21 -27
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +14 -11
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +276 -0
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cu +34 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +126 -59
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +10 -2
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +322 -100
- data/ext/sources/ggml/src/ggml-cuda/sum.cu +6 -10
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +21 -4
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +21 -18
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +259 -0
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +14 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +3 -3
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +90 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +8 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cu +92 -6
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +110 -22
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +58 -36
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -3
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +10 -2
- data/ext/sources/ggml/src/ggml-impl.h +119 -9
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -7
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +446 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +33 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +600 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +1376 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +226 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +1308 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +136 -63
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +3158 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +82 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +718 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +2854 -1503
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +18 -8
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +18 -0
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +2510 -242
- data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +107 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +84 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/div.cl +66 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +370 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/glu.cl +177 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/group_norm.cl +49 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +73 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +133 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +80 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +79 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +3 -2
- data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +189 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/sub.cl +66 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +20 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tsembd.cl +2 -2
- data/ext/sources/ggml/src/ggml-opencl/kernels/upscale.cl +2 -3
- data/ext/sources/ggml/src/ggml-opt.cpp +97 -41
- data/ext/sources/ggml/src/ggml-quants.c +111 -16
- data/ext/sources/ggml/src/ggml-quants.h +6 -0
- data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +67 -47
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +15 -5
- data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/concat.cpp +25 -16
- data/ext/sources/ggml/src/ggml-sycl/conv.cpp +10 -4
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +166 -99
- data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +72 -306
- data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +213 -1
- data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +67 -49
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +1 -31
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +79 -29
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +14 -26
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +9 -6
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +328 -323
- data/ext/sources/ggml/src/ggml-sycl/gla.cpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/mmq.cpp +80 -60
- data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +201 -132
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +74 -55
- data/ext/sources/ggml/src/ggml-sycl/quantize.hpp +133 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +8 -9
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +35 -42
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +234 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/tsembd.cpp +12 -6
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +2 -6
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +16 -12
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3492 -883
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +41 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +13 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +39 -29
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +349 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +2 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +66 -12
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +154 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +2 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +6 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +4 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +69 -24
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +60 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +98 -42
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +64 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +74 -13
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +4 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +19 -10
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +25 -15
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +18 -14
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +126 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +65 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +11 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +140 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +144 -531
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +206 -38
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp +556 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +12 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +15 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +111 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +24 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +53 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +55 -11
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +1 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.comp +5 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +29 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +38 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +4 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +101 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +69 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp +25 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +335 -77
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +54 -0
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1558 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl +44 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl +41 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +45 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +930 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +60 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +124 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.tmpl.wgsl +874 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl +44 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl +41 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +907 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +57 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl +48 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +81 -0
- data/ext/sources/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
- data/ext/sources/ggml/src/ggml-zdnn/common.hpp +59 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +628 -0
- data/ext/sources/ggml/src/ggml-zdnn/mmf.cpp +80 -0
- data/ext/sources/ggml/src/ggml-zdnn/mmf.hpp +12 -0
- data/ext/sources/ggml/src/ggml-zdnn/utils.cpp +79 -0
- data/ext/sources/ggml/src/ggml-zdnn/utils.hpp +19 -0
- data/ext/sources/ggml/src/ggml.c +478 -98
- data/ext/sources/ggml/src/gguf.cpp +8 -1
- data/ext/sources/src/whisper.cpp +23 -46
- data/ext/sources/tests/CMakeLists.txt +8 -1
- data/ext/sources/tests/test-vad-full.cpp +3 -3
- data/ext/sources/tests/test-vad.cpp +2 -2
- data/lib/whisper/model/uri.rb +1 -1
- data/sig/whisper.rbs +7 -0
- data/test/test_params.rb +8 -0
- data/test/test_whisper.rb +1 -1
- data/whispercpp.gemspec +1 -1
- metadata +164 -157
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.cpp +0 -279
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified.cpp +0 -1841
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified.h +0 -303
- data/ext/sources/ggml/include/ggml-kompute.h +0 -50
- data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +0 -107
- data/ext/sources/ggml/src/ggml-amx/common.h +0 -94
- data/ext/sources/ggml/src/ggml-amx/ggml-amx.cpp +0 -446
- data/ext/sources/ggml/src/ggml-amx/mmq.cpp +0 -2510
- data/ext/sources/ggml/src/ggml-amx/mmq.h +0 -17
- data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +0 -30
- data/ext/sources/ggml/src/ggml-cann/kernels/ascendc_kernels.h +0 -19
- data/ext/sources/ggml/src/ggml-cann/kernels/dup.cpp +0 -234
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f16.cpp +0 -197
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f32.cpp +0 -190
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -204
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -191
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -218
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -216
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -295
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +0 -357
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +0 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +0 -365
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +0 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +0 -482
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +0 -472
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +0 -166
- data/ext/sources/ggml/src/ggml-kompute/ggml-kompute.cpp +0 -2251
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +0 -112
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +0 -58
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +0 -25
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +0 -30
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +0 -22
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +0 -17
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +0 -31
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +0 -31
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +0 -38
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +0 -39
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +0 -44
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +0 -69
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +0 -51
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +0 -33
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +0 -35
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +0 -140
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +0 -106
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +0 -73
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +0 -28
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +0 -84
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +0 -21
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +0 -53
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +0 -19
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +0 -23
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +0 -22
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +0 -72
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +0 -71
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.m +0 -6280
|
@@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B,
|
|
|
11
11
|
const u_int n_seq_tokens = T / B;
|
|
12
12
|
sycl::range<1> block_dims((C / H));
|
|
13
13
|
sycl::range<1> grid_dims((B * H));
|
|
14
|
-
|
|
14
|
+
stream->submit([&](sycl::handler & cgh) {
|
|
15
15
|
/* local memory accessors*/
|
|
16
16
|
auto _k = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
|
|
17
17
|
auto _r = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
|
|
18
18
|
auto _td = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
|
|
19
19
|
|
|
20
|
-
|
|
20
|
+
cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
|
|
21
21
|
u_int tid = item.get_local_id(0);
|
|
22
22
|
u_int bid = item.get_group(0);
|
|
23
23
|
|
|
@@ -26,7 +26,7 @@ static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_
|
|
|
26
26
|
|
|
27
27
|
// make each work-item deal with more elements since sycl global range can not exceed max int
|
|
28
28
|
for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
|
|
29
|
-
const int64_t ksize = OW *
|
|
29
|
+
const int64_t ksize = OW * KH;
|
|
30
30
|
const int64_t kx = i / ksize;
|
|
31
31
|
const int64_t kd = kx * ksize;
|
|
32
32
|
const int64_t ky = (i - kd) / OW;
|
|
@@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I
|
|
|
70
70
|
|
|
71
71
|
const int64_t CHW = IC * KH * KW;
|
|
72
72
|
|
|
73
|
-
|
|
73
|
+
stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
|
|
74
74
|
im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
|
|
75
75
|
p0, p1, d0, d1, item_ct1);
|
|
76
76
|
});
|
|
@@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1818
1818
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
1819
1819
|
{sycl::aspect::fp16});
|
|
1820
1820
|
|
|
1821
|
-
|
|
1821
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
1822
1822
|
sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
|
|
1823
1823
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
1824
1824
|
sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
|
|
@@ -1829,8 +1829,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1829
1829
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
1830
1830
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
1831
1831
|
|
|
1832
|
-
|
|
1833
|
-
|
|
1832
|
+
cgh.parallel_for(
|
|
1833
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
1834
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
1834
1835
|
mul_mat_q4_0<need_check>(
|
|
1835
1836
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
1836
1837
|
nrows_dst, item_ct1,
|
|
@@ -1852,7 +1853,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1852
1853
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
1853
1854
|
{sycl::aspect::fp16});
|
|
1854
1855
|
|
|
1855
|
-
|
|
1856
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
1856
1857
|
sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
|
|
1857
1858
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
1858
1859
|
sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
|
|
@@ -1863,8 +1864,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1863
1864
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
1864
1865
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
1865
1866
|
|
|
1866
|
-
|
|
1867
|
-
|
|
1867
|
+
cgh.parallel_for(
|
|
1868
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
1869
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
1868
1870
|
mul_mat_q4_0<need_check>(
|
|
1869
1871
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
1870
1872
|
nrows_dst, item_ct1,
|
|
@@ -1931,7 +1933,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1931
1933
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
1932
1934
|
{sycl::aspect::fp16});
|
|
1933
1935
|
|
|
1934
|
-
|
|
1936
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
1935
1937
|
sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
|
|
1936
1938
|
sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
|
|
1937
1939
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
|
|
@@ -1942,8 +1944,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1942
1944
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
1943
1945
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
1944
1946
|
|
|
1945
|
-
|
|
1946
|
-
|
|
1947
|
+
cgh.parallel_for(
|
|
1948
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
1949
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
1947
1950
|
mul_mat_q4_1<need_check>(
|
|
1948
1951
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
1949
1952
|
nrows_dst, item_ct1,
|
|
@@ -1965,7 +1968,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1965
1968
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
1966
1969
|
{sycl::aspect::fp16});
|
|
1967
1970
|
|
|
1968
|
-
|
|
1971
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
1969
1972
|
sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
|
|
1970
1973
|
sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
|
|
1971
1974
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
|
|
@@ -1976,8 +1979,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
1976
1979
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
1977
1980
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
1978
1981
|
|
|
1979
|
-
|
|
1980
|
-
|
|
1982
|
+
cgh.parallel_for(
|
|
1983
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
1984
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
1981
1985
|
mul_mat_q4_1<need_check>(
|
|
1982
1986
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
1983
1987
|
nrows_dst, item_ct1,
|
|
@@ -2044,7 +2048,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2044
2048
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2045
2049
|
{sycl::aspect::fp16});
|
|
2046
2050
|
|
|
2047
|
-
|
|
2051
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2048
2052
|
sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
|
|
2049
2053
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2050
2054
|
sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
|
|
@@ -2055,8 +2059,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2055
2059
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2056
2060
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2057
2061
|
|
|
2058
|
-
|
|
2059
|
-
|
|
2062
|
+
cgh.parallel_for(
|
|
2063
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2064
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2060
2065
|
mul_mat_q5_0<need_check>(
|
|
2061
2066
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2062
2067
|
nrows_dst, item_ct1,
|
|
@@ -2078,7 +2083,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2078
2083
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2079
2084
|
{sycl::aspect::fp16});
|
|
2080
2085
|
|
|
2081
|
-
|
|
2086
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2082
2087
|
sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
|
|
2083
2088
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2084
2089
|
sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
|
|
@@ -2089,8 +2094,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2089
2094
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2090
2095
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2091
2096
|
|
|
2092
|
-
|
|
2093
|
-
|
|
2097
|
+
cgh.parallel_for(
|
|
2098
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2099
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2094
2100
|
mul_mat_q5_0<need_check>(
|
|
2095
2101
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2096
2102
|
nrows_dst, item_ct1,
|
|
@@ -2157,7 +2163,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2157
2163
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2158
2164
|
{sycl::aspect::fp16});
|
|
2159
2165
|
|
|
2160
|
-
|
|
2166
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2161
2167
|
sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
|
|
2162
2168
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2163
2169
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
|
|
@@ -2168,8 +2174,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2168
2174
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2169
2175
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2170
2176
|
|
|
2171
|
-
|
|
2172
|
-
|
|
2177
|
+
cgh.parallel_for(
|
|
2178
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2179
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2173
2180
|
mul_mat_q5_1<need_check>(
|
|
2174
2181
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2175
2182
|
nrows_dst, item_ct1,
|
|
@@ -2191,7 +2198,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2191
2198
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2192
2199
|
{sycl::aspect::fp16});
|
|
2193
2200
|
|
|
2194
|
-
|
|
2201
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2195
2202
|
sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
|
|
2196
2203
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2197
2204
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
|
|
@@ -2202,8 +2209,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2202
2209
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2203
2210
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2204
2211
|
|
|
2205
|
-
|
|
2206
|
-
|
|
2212
|
+
cgh.parallel_for(
|
|
2213
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2214
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2207
2215
|
mul_mat_q5_1<need_check>(
|
|
2208
2216
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2209
2217
|
nrows_dst, item_ct1,
|
|
@@ -2270,7 +2278,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2270
2278
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2271
2279
|
{sycl::aspect::fp16});
|
|
2272
2280
|
|
|
2273
|
-
|
|
2281
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2274
2282
|
sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
|
|
2275
2283
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2276
2284
|
sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
|
|
@@ -2281,8 +2289,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2281
2289
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2282
2290
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2283
2291
|
|
|
2284
|
-
|
|
2285
|
-
|
|
2292
|
+
cgh.parallel_for(
|
|
2293
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2294
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2286
2295
|
mul_mat_q8_0<need_check>(
|
|
2287
2296
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2288
2297
|
nrows_dst, item_ct1,
|
|
@@ -2304,7 +2313,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2304
2313
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2305
2314
|
{sycl::aspect::fp16});
|
|
2306
2315
|
|
|
2307
|
-
|
|
2316
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2308
2317
|
sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
|
|
2309
2318
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2310
2319
|
sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
|
|
@@ -2315,8 +2324,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2315
2324
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2316
2325
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2317
2326
|
|
|
2318
|
-
|
|
2319
|
-
|
|
2327
|
+
cgh.parallel_for(
|
|
2328
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2329
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2320
2330
|
mul_mat_q8_0<need_check>(
|
|
2321
2331
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2322
2332
|
nrows_dst, item_ct1,
|
|
@@ -2383,7 +2393,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2383
2393
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2384
2394
|
{sycl::aspect::fp16});
|
|
2385
2395
|
|
|
2386
|
-
|
|
2396
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2387
2397
|
sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
|
|
2388
2398
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2389
2399
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
|
|
@@ -2396,8 +2406,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2396
2406
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2397
2407
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2398
2408
|
|
|
2399
|
-
|
|
2400
|
-
|
|
2409
|
+
cgh.parallel_for(
|
|
2410
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2411
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2401
2412
|
mul_mat_q2_K<need_check>(
|
|
2402
2413
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2403
2414
|
nrows_dst, item_ct1,
|
|
@@ -2420,7 +2431,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2420
2431
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2421
2432
|
{sycl::aspect::fp16});
|
|
2422
2433
|
|
|
2423
|
-
|
|
2434
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2424
2435
|
sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
|
|
2425
2436
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2426
2437
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
|
|
@@ -2433,8 +2444,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2433
2444
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2434
2445
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2435
2446
|
|
|
2436
|
-
|
|
2437
|
-
|
|
2447
|
+
cgh.parallel_for(
|
|
2448
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2449
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2438
2450
|
mul_mat_q2_K<need_check>(
|
|
2439
2451
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2440
2452
|
nrows_dst, item_ct1,
|
|
@@ -2504,7 +2516,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2504
2516
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2505
2517
|
{sycl::aspect::fp16});
|
|
2506
2518
|
|
|
2507
|
-
|
|
2519
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2508
2520
|
sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
|
|
2509
2521
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2510
2522
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
|
|
@@ -2519,8 +2531,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2519
2531
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2520
2532
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2521
2533
|
|
|
2522
|
-
|
|
2523
|
-
|
|
2534
|
+
cgh.parallel_for(
|
|
2535
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2536
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2524
2537
|
mul_mat_q3_K<need_check>(
|
|
2525
2538
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2526
2539
|
nrows_dst, item_ct1,
|
|
@@ -2544,7 +2557,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2544
2557
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2545
2558
|
{sycl::aspect::fp16});
|
|
2546
2559
|
|
|
2547
|
-
|
|
2560
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2548
2561
|
sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
|
|
2549
2562
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2550
2563
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
|
|
@@ -2559,8 +2572,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2559
2572
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2560
2573
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2561
2574
|
|
|
2562
|
-
|
|
2563
|
-
|
|
2575
|
+
cgh.parallel_for(
|
|
2576
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2577
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2564
2578
|
mul_mat_q3_K<need_check>(
|
|
2565
2579
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2566
2580
|
nrows_dst, item_ct1,
|
|
@@ -2630,7 +2644,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2630
2644
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2631
2645
|
{sycl::aspect::fp16});
|
|
2632
2646
|
|
|
2633
|
-
|
|
2647
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2634
2648
|
sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
|
|
2635
2649
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2636
2650
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
|
|
@@ -2643,8 +2657,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2643
2657
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2644
2658
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2645
2659
|
|
|
2646
|
-
|
|
2647
|
-
|
|
2660
|
+
cgh.parallel_for(
|
|
2661
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2662
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2648
2663
|
mul_mat_q4_K<need_check>(
|
|
2649
2664
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2650
2665
|
nrows_dst, item_ct1,
|
|
@@ -2667,7 +2682,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2667
2682
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2668
2683
|
{sycl::aspect::fp16});
|
|
2669
2684
|
|
|
2670
|
-
|
|
2685
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2671
2686
|
sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
|
|
2672
2687
|
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
|
|
2673
2688
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
|
|
@@ -2680,8 +2695,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2680
2695
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2681
2696
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2682
2697
|
|
|
2683
|
-
|
|
2684
|
-
|
|
2698
|
+
cgh.parallel_for(
|
|
2699
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2700
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2685
2701
|
mul_mat_q4_K<need_check>(
|
|
2686
2702
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2687
2703
|
nrows_dst, item_ct1,
|
|
@@ -2749,7 +2765,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2749
2765
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2750
2766
|
{sycl::aspect::fp16});
|
|
2751
2767
|
|
|
2752
|
-
|
|
2768
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2753
2769
|
sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
|
|
2754
2770
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2755
2771
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
|
|
@@ -2762,8 +2778,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2762
2778
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2763
2779
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2764
2780
|
|
|
2765
|
-
|
|
2766
|
-
|
|
2781
|
+
cgh.parallel_for(
|
|
2782
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2783
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2767
2784
|
mul_mat_q5_K<need_check>(
|
|
2768
2785
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2769
2786
|
nrows_dst, item_ct1,
|
|
@@ -2786,7 +2803,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2786
2803
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2787
2804
|
{sycl::aspect::fp16});
|
|
2788
2805
|
|
|
2789
|
-
|
|
2806
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2790
2807
|
sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
|
|
2791
2808
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2792
2809
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
|
|
@@ -2799,8 +2816,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2799
2816
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2800
2817
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2801
2818
|
|
|
2802
|
-
|
|
2803
|
-
|
|
2819
|
+
cgh.parallel_for(
|
|
2820
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2821
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2804
2822
|
mul_mat_q5_K<need_check>(
|
|
2805
2823
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2806
2824
|
nrows_dst, item_ct1,
|
|
@@ -2868,7 +2886,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2868
2886
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2869
2887
|
{sycl::aspect::fp16});
|
|
2870
2888
|
|
|
2871
|
-
|
|
2889
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2872
2890
|
sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
|
|
2873
2891
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2874
2892
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
|
|
@@ -2881,8 +2899,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2881
2899
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2882
2900
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2883
2901
|
|
|
2884
|
-
|
|
2885
|
-
|
|
2902
|
+
cgh.parallel_for(
|
|
2903
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2904
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2886
2905
|
mul_mat_q6_K<need_check>(
|
|
2887
2906
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2888
2907
|
nrows_dst, item_ct1,
|
|
@@ -2905,7 +2924,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2905
2924
|
dpct::has_capability_or_fail(stream->get_device(),
|
|
2906
2925
|
{sycl::aspect::fp16});
|
|
2907
2926
|
|
|
2908
|
-
|
|
2927
|
+
stream->submit([&](sycl::handler &cgh) {
|
|
2909
2928
|
sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
|
|
2910
2929
|
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
|
|
2911
2930
|
sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
|
|
@@ -2918,8 +2937,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
|
|
|
2918
2937
|
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
|
|
2919
2938
|
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
|
|
2920
2939
|
|
|
2921
|
-
|
|
2922
|
-
|
|
2940
|
+
cgh.parallel_for(
|
|
2941
|
+
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
|
2942
|
+
[=](sycl::nd_item<3> item_ct1) {
|
|
2923
2943
|
mul_mat_q6_K<need_check>(
|
|
2924
2944
|
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
|
|
2925
2945
|
nrows_dst, item_ct1,
|