elelem-ggml 0.1.0
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 +7 -0
- data/LICENSE.txt +21 -0
- data/exe/elelem-ggml +8 -0
- data/ext/elelem/ggml/elelem.c +466 -0
- data/ext/elelem/ggml/extconf.rb +52 -0
- data/lib/elelem/ggml/chat_ml.rb +23 -0
- data/lib/elelem/ggml/gguf.rb +82 -0
- data/lib/elelem/ggml/provider.rb +63 -0
- data/lib/elelem/ggml/tokenizer.rb +95 -0
- data/lib/elelem/ggml/version.rb +7 -0
- data/lib/elelem/ggml.rb +37 -0
- data/vendor/ggml/AUTHORS +335 -0
- data/vendor/ggml/CMakeLists.txt +499 -0
- data/vendor/ggml/CONTRIBUTING.md +3 -0
- data/vendor/ggml/LICENSE +21 -0
- data/vendor/ggml/README.md +49 -0
- data/vendor/ggml/cmake/FindNCCL.cmake +36 -0
- data/vendor/ggml/cmake/GitVars.cmake +22 -0
- data/vendor/ggml/cmake/common.cmake +50 -0
- data/vendor/ggml/cmake/ggml-config.cmake.in +216 -0
- data/vendor/ggml/ggml.pc.in +10 -0
- data/vendor/ggml/include/ggml-alloc.h +86 -0
- data/vendor/ggml/include/ggml-backend.h +437 -0
- data/vendor/ggml/include/ggml-blas.h +25 -0
- data/vendor/ggml/include/ggml-cann.h +123 -0
- data/vendor/ggml/include/ggml-cpp.h +39 -0
- data/vendor/ggml/include/ggml-cpu.h +152 -0
- data/vendor/ggml/include/ggml-cuda.h +47 -0
- data/vendor/ggml/include/ggml-et.h +28 -0
- data/vendor/ggml/include/ggml-hexagon.h +19 -0
- data/vendor/ggml/include/ggml-metal.h +61 -0
- data/vendor/ggml/include/ggml-opencl.h +26 -0
- data/vendor/ggml/include/ggml-openvino.h +37 -0
- data/vendor/ggml/include/ggml-opt.h +256 -0
- data/vendor/ggml/include/ggml-rpc.h +35 -0
- data/vendor/ggml/include/ggml-sycl.h +57 -0
- data/vendor/ggml/include/ggml-virtgpu.h +14 -0
- data/vendor/ggml/include/ggml-vulkan.h +29 -0
- data/vendor/ggml/include/ggml-webgpu.h +19 -0
- data/vendor/ggml/include/ggml-zdnn.h +17 -0
- data/vendor/ggml/include/ggml-zendnn.h +22 -0
- data/vendor/ggml/include/ggml.h +2963 -0
- data/vendor/ggml/include/gguf.h +211 -0
- data/vendor/ggml/requirements.txt +12 -0
- data/vendor/ggml/scripts/gen-authors.sh +9 -0
- data/vendor/ggml/scripts/make-release-checks.sh +96 -0
- data/vendor/ggml/scripts/make-release-desc.sh +69 -0
- data/vendor/ggml/scripts/make-release-summary.txt +33 -0
- data/vendor/ggml/scripts/release.sh +195 -0
- data/vendor/ggml/scripts/sync-llama-am.sh +167 -0
- data/vendor/ggml/scripts/sync-llama.last +1 -0
- data/vendor/ggml/scripts/sync-llama.sh +21 -0
- data/vendor/ggml/scripts/sync-whisper-am.sh +138 -0
- data/vendor/ggml/scripts/sync-whisper.last +1 -0
- data/vendor/ggml/scripts/sync-whisper.sh +17 -0
- data/vendor/ggml/src/CMakeLists.txt +634 -0
- data/vendor/ggml/src/ggml-alloc.c +1249 -0
- data/vendor/ggml/src/ggml-backend-dl.cpp +48 -0
- data/vendor/ggml/src/ggml-backend-dl.h +45 -0
- data/vendor/ggml/src/ggml-backend-impl.h +291 -0
- data/vendor/ggml/src/ggml-backend-meta.cpp +2511 -0
- data/vendor/ggml/src/ggml-backend-reg.cpp +605 -0
- data/vendor/ggml/src/ggml-backend.cpp +2517 -0
- data/vendor/ggml/src/ggml-blas/CMakeLists.txt +101 -0
- data/vendor/ggml/src/ggml-blas/ggml-blas.cpp +530 -0
- data/vendor/ggml/src/ggml-cann/CMakeLists.txt +89 -0
- data/vendor/ggml/src/ggml-cann/acl_tensor.cpp +195 -0
- data/vendor/ggml/src/ggml-cann/acl_tensor.h +349 -0
- data/vendor/ggml/src/ggml-cann/aclnn_ops.cpp +4479 -0
- data/vendor/ggml/src/ggml-cann/aclnn_ops.h +1191 -0
- data/vendor/ggml/src/ggml-cann/common.h +651 -0
- data/vendor/ggml/src/ggml-cann/ggml-cann.cpp +3070 -0
- data/vendor/ggml/src/ggml-common.h +1911 -0
- data/vendor/ggml/src/ggml-cpu/CMakeLists.txt +685 -0
- data/vendor/ggml/src/ggml-cpu/amx/amx.cpp +249 -0
- data/vendor/ggml/src/ggml-cpu/amx/amx.h +8 -0
- data/vendor/ggml/src/ggml-cpu/amx/common.h +115 -0
- data/vendor/ggml/src/ggml-cpu/amx/mmq.cpp +2511 -0
- data/vendor/ggml/src/ggml-cpu/amx/mmq.h +10 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +41 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/quants.c +4319 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/repack.cpp +5156 -0
- data/vendor/ggml/src/ggml-cpu/arch/loongarch/quants.c +2309 -0
- data/vendor/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
- data/vendor/ggml/src/ggml-cpu/arch/powerpc/quants.c +2304 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/quants.c +6596 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1703 -0
- data/vendor/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
- data/vendor/ggml/src/ggml-cpu/arch/s390/quants.c +1465 -0
- data/vendor/ggml/src/ggml-cpu/arch/wasm/quants.c +1292 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/quants.c +4108 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/repack.cpp +6407 -0
- data/vendor/ggml/src/ggml-cpu/arch-fallback.h +353 -0
- data/vendor/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
- data/vendor/ggml/src/ggml-cpu/binary-ops.h +16 -0
- data/vendor/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
- data/vendor/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/vendor/ggml/src/ggml-cpu/common.h +95 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu-impl.h +539 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu.c +3936 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu.cpp +712 -0
- data/vendor/ggml/src/ggml-cpu/hbm.cpp +55 -0
- data/vendor/ggml/src/ggml-cpu/hbm.h +8 -0
- data/vendor/ggml/src/ggml-cpu/iqp.cpp +1253 -0
- data/vendor/ggml/src/ggml-cpu/iqp.h +39 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/CMakeLists.txt +14 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.cpp +1101 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.h +101 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +1919 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
- data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4164 -0
- data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
- data/vendor/ggml/src/ggml-cpu/ops.cpp +12164 -0
- data/vendor/ggml/src/ggml-cpu/ops.h +124 -0
- data/vendor/ggml/src/ggml-cpu/quants.c +1339 -0
- data/vendor/ggml/src/ggml-cpu/quants.h +106 -0
- data/vendor/ggml/src/ggml-cpu/repack.cpp +4836 -0
- data/vendor/ggml/src/ggml-cpu/repack.h +245 -0
- data/vendor/ggml/src/ggml-cpu/simd-gemm.h +226 -0
- data/vendor/ggml/src/ggml-cpu/simd-mappings.h +1319 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime.cpp +1742 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime.h +21 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +1027 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_kernels.h +189 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
- data/vendor/ggml/src/ggml-cpu/traits.cpp +36 -0
- data/vendor/ggml/src/ggml-cpu/traits.h +38 -0
- data/vendor/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
- data/vendor/ggml/src/ggml-cpu/unary-ops.h +35 -0
- data/vendor/ggml/src/ggml-cpu/vec.cpp +613 -0
- data/vendor/ggml/src/ggml-cpu/vec.h +1570 -0
- data/vendor/ggml/src/ggml-cuda/CMakeLists.txt +267 -0
- data/vendor/ggml/src/ggml-cuda/acc.cu +61 -0
- data/vendor/ggml/src/ggml-cuda/acc.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/add-id.cu +58 -0
- data/vendor/ggml/src/ggml-cuda/add-id.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/allreduce.cu +971 -0
- data/vendor/ggml/src/ggml-cuda/allreduce.cuh +29 -0
- data/vendor/ggml/src/ggml-cuda/arange.cu +34 -0
- data/vendor/ggml/src/ggml-cuda/arange.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/argmax.cu +91 -0
- data/vendor/ggml/src/ggml-cuda/argmax.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/argsort.cu +292 -0
- data/vendor/ggml/src/ggml-cuda/argsort.cuh +20 -0
- data/vendor/ggml/src/ggml-cuda/binbcast.cu +574 -0
- data/vendor/ggml/src/ggml-cuda/binbcast.cuh +12 -0
- data/vendor/ggml/src/ggml-cuda/clamp.cu +45 -0
- data/vendor/ggml/src/ggml-cuda/clamp.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
- data/vendor/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/common.cuh +1678 -0
- data/vendor/ggml/src/ggml-cuda/concat.cu +242 -0
- data/vendor/ggml/src/ggml-cuda/concat.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cu +88 -0
- data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cu +115 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d.cu +167 -0
- data/vendor/ggml/src/ggml-cuda/conv2d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/convert.cu +705 -0
- data/vendor/ggml/src/ggml-cuda/convert.cuh +66 -0
- data/vendor/ggml/src/ggml-cuda/count-equal.cu +64 -0
- data/vendor/ggml/src/ggml-cuda/count-equal.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/cp-async.cuh +57 -0
- data/vendor/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
- data/vendor/ggml/src/ggml-cuda/cpy.cu +617 -0
- data/vendor/ggml/src/ggml-cuda/cpy.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cu +177 -0
- data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/cumsum.cu +307 -0
- data/vendor/ggml/src/ggml-cuda/cumsum.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/dequantize.cuh +452 -0
- data/vendor/ggml/src/ggml-cuda/diag.cu +77 -0
- data/vendor/ggml/src/ggml-cuda/diag.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/diagmask.cu +40 -0
- data/vendor/ggml/src/ggml-cuda/diagmask.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/dsv4-hc.cu +294 -0
- data/vendor/ggml/src/ggml-cuda/dsv4-hc.cuh +6 -0
- data/vendor/ggml/src/ggml-cuda/fattn-common.cuh +1289 -0
- data/vendor/ggml/src/ggml-cuda/fattn-mma-f16.cuh +2136 -0
- data/vendor/ggml/src/ggml-cuda/fattn-swizzle.cuh +126 -0
- data/vendor/ggml/src/ggml-cuda/fattn-tile.cu +60 -0
- data/vendor/ggml/src/ggml-cuda/fattn-tile.cuh +1355 -0
- data/vendor/ggml/src/ggml-cuda/fattn-vec.cuh +611 -0
- data/vendor/ggml/src/ggml-cuda/fattn.cu +722 -0
- data/vendor/ggml/src/ggml-cuda/fattn.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/fill.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/fill.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/fwht.cu +101 -0
- data/vendor/ggml/src/ggml-cuda/fwht.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/gated_delta_net.cu +327 -0
- data/vendor/ggml/src/ggml-cuda/gated_delta_net.cuh +14 -0
- data/vendor/ggml/src/ggml-cuda/getrows.cu +490 -0
- data/vendor/ggml/src/ggml-cuda/getrows.cuh +15 -0
- data/vendor/ggml/src/ggml-cuda/ggml-cuda.cu +5781 -0
- data/vendor/ggml/src/ggml-cuda/gla.cu +93 -0
- data/vendor/ggml/src/ggml-cuda/gla.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/im2col.cu +267 -0
- data/vendor/ggml/src/ggml-cuda/im2col.cuh +6 -0
- data/vendor/ggml/src/ggml-cuda/lightning-indexer.cu +588 -0
- data/vendor/ggml/src/ggml-cuda/lightning-indexer.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/mean.cu +77 -0
- data/vendor/ggml/src/ggml-cuda/mean.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/mma.cuh +1456 -0
- data/vendor/ggml/src/ggml-cuda/mmf.cu +191 -0
- data/vendor/ggml/src/ggml-cuda/mmf.cuh +908 -0
- data/vendor/ggml/src/ggml-cuda/mmid.cu +180 -0
- data/vendor/ggml/src/ggml-cuda/mmid.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-ampere.cuh +383 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-blackwell.cuh +37 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-cdna.cuh +185 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-dp4a.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-older.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna2.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh +290 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3.cuh +274 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna4.cuh +290 -0
- data/vendor/ggml/src/ggml-cuda/mmq-load-tiles.cuh +1768 -0
- data/vendor/ggml/src/ggml-cuda/mmq-vec-dot.cuh +1240 -0
- data/vendor/ggml/src/ggml-cuda/mmq.cu +386 -0
- data/vendor/ggml/src/ggml-cuda/mmq.cuh +1598 -0
- data/vendor/ggml/src/ggml-cuda/mmvf.cu +875 -0
- data/vendor/ggml/src/ggml-cuda/mmvf.cuh +14 -0
- data/vendor/ggml/src/ggml-cuda/mmvq.cu +1504 -0
- data/vendor/ggml/src/ggml-cuda/mmvq.cuh +18 -0
- data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cu +65 -0
- data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/norm.cu +698 -0
- data/vendor/ggml/src/ggml-cuda/norm.cuh +18 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/out-prod.cu +125 -0
- data/vendor/ggml/src/ggml-cuda/out-prod.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/pad.cu +106 -0
- data/vendor/ggml/src/ggml-cuda/pad.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
- data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pool1d.cu +85 -0
- data/vendor/ggml/src/ggml-cuda/pool1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pool2d.cu +94 -0
- data/vendor/ggml/src/ggml-cuda/pool2d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/quantize.cu +697 -0
- data/vendor/ggml/src/ggml-cuda/quantize.cuh +70 -0
- data/vendor/ggml/src/ggml-cuda/reduce_rows.cuh +43 -0
- data/vendor/ggml/src/ggml-cuda/roll.cu +67 -0
- data/vendor/ggml/src/ggml-cuda/roll.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/rope.cu +941 -0
- data/vendor/ggml/src/ggml-cuda/rope.cuh +11 -0
- data/vendor/ggml/src/ggml-cuda/scale.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/scale.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/set-rows.cu +398 -0
- data/vendor/ggml/src/ggml-cuda/set-rows.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/set.cu +39 -0
- data/vendor/ggml/src/ggml-cuda/set.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/snake.cu +72 -0
- data/vendor/ggml/src/ggml-cuda/snake.cuh +8 -0
- data/vendor/ggml/src/ggml-cuda/softcap.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/softcap.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/softmax.cu +480 -0
- data/vendor/ggml/src/ggml-cuda/softmax.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/solve_tri.cu +273 -0
- data/vendor/ggml/src/ggml-cuda/solve_tri.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/ssm-conv.cu +206 -0
- data/vendor/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/ssm-scan.cu +859 -0
- data/vendor/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/sum.cu +41 -0
- data/vendor/ggml/src/ggml-cuda/sum.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/sumrows.cu +47 -0
- data/vendor/ggml/src/ggml-cuda/sumrows.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +111 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/top-k.cu +275 -0
- data/vendor/ggml/src/ggml-cuda/top-k.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/topk-moe.cu +442 -0
- data/vendor/ggml/src/ggml-cuda/topk-moe.cuh +31 -0
- data/vendor/ggml/src/ggml-cuda/tri.cu +136 -0
- data/vendor/ggml/src/ggml-cuda/tri.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/tsembd.cu +47 -0
- data/vendor/ggml/src/ggml-cuda/tsembd.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/unary.cu +721 -0
- data/vendor/ggml/src/ggml-cuda/unary.cuh +123 -0
- data/vendor/ggml/src/ggml-cuda/upscale.cu +293 -0
- data/vendor/ggml/src/ggml-cuda/upscale.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/vecdotq.cuh +1371 -0
- data/vendor/ggml/src/ggml-cuda/vendors/cuda.h +28 -0
- data/vendor/ggml/src/ggml-cuda/vendors/hip.h +301 -0
- data/vendor/ggml/src/ggml-cuda/vendors/musa.h +151 -0
- data/vendor/ggml/src/ggml-cuda/wkv.cu +253 -0
- data/vendor/ggml/src/ggml-cuda/wkv.cuh +7 -0
- data/vendor/ggml/src/ggml-et/CMakeLists.txt +246 -0
- data/vendor/ggml/src/ggml-et/cmake/embed_one_kernel.cmake +15 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.cpp.in +6 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.hpp.in +12 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.cpp.in +18 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.h.in +13 -0
- data/vendor/ggml/src/ggml-et/et-kernels/CMakeLists.txt +137 -0
- data/vendor/ggml/src/ggml-et/et-kernels/scripts/check_unimplemented_instructions.sh +36 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/RunBackend.sh +23 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/block_ops.h +997 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/clamp_f32.c +120 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/concat_f32.c +175 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f16.c +107 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f32.c +248 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/conv_2d_f32_me.c +807 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cpy_f32_f16.c +110 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/crt.S +15 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cumsum_f32.c +96 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/diag_f32.c +90 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/el_map_f32.c +377 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/fill_f32.c +87 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f16_me.c +1000 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f32.c +217 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/gated_delta_net_f32.c +346 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/get_rows_f32.c +612 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ggml_tensor.h +44 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/glu_f32.c +606 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/group_norm_f32.c +171 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/im2col.c +130 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/l2_norm_f32.c +237 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/linker.ld +85 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/math_fp.h +299 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mean_f32.c +220 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/memops.c +181 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0.c +358 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0_matrix_engine.c +368 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q8_0.c +413 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16.c +142 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16_matrix_engine.c +329 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32.c +137 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32_matrix_engine.c +155 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q4_0.c +169 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q8_0.c +160 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_f32.c +288 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/norm_f32.c +328 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/pad_f32.c +165 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/platform.h +545 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/quants.h +72 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/repeat_f32.c +118 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_f32.c +270 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_mul_f32.c +290 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rope_f32.c +656 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv6_f32.c +184 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv7_f32.c +272 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/scale_f32.c +94 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/set_f32.c +101 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/set_rows_f32.c +394 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/softmax_f32.c +698 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/solve_tri_f32.c +109 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/sqr_f32.c +88 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_conv_f32.c +129 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_scan_f32.c +282 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/sum_rows_f32.c +103 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/tensor.h +897 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/tri_f32.c +244 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/uberkernel.c +497 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/unary_f32.c +705 -0
- data/vendor/ggml/src/ggml-et/ggml-et-common.h +86 -0
- data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.cpp +502 -0
- data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.h +54 -0
- data/vendor/ggml/src/ggml-et/ggml-et-kernels.cpp +508 -0
- data/vendor/ggml/src/ggml-et/ggml-et-kernels.h +48 -0
- data/vendor/ggml/src/ggml-et/ggml-et-memops.cpp +36 -0
- data/vendor/ggml/src/ggml-et/ggml-et-memops.h +18 -0
- data/vendor/ggml/src/ggml-et/ggml-et-ops.cpp +2584 -0
- data/vendor/ggml/src/ggml-et/ggml-et-ops.h +393 -0
- data/vendor/ggml/src/ggml-et/ggml-et-uberkernel-common.h +17 -0
- data/vendor/ggml/src/ggml-et/ggml-et.cpp +1880 -0
- data/vendor/ggml/src/ggml-feats.h +166 -0
- data/vendor/ggml/src/ggml-hexagon/CMakeLists.txt +113 -0
- data/vendor/ggml/src/ggml-hexagon/ggml-hexagon.cpp +6523 -0
- data/vendor/ggml/src/ggml-hexagon/htp/CMakeLists.txt +58 -0
- data/vendor/ggml/src/ggml-hexagon/htp/act-ops.c +585 -0
- data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.c +398 -0
- data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.h +40 -0
- data/vendor/ggml/src/ggml-hexagon/htp/argsort-ops.c +514 -0
- data/vendor/ggml/src/ggml-hexagon/htp/binary-ops.c +873 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +155 -0
- data/vendor/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cpy-ops.c +354 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cumsum-ops.c +265 -0
- data/vendor/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
- data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.c +104 -0
- data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.h +408 -0
- data/vendor/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
- data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +2479 -0
- data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +344 -0
- data/vendor/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1157 -0
- data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.c +260 -0
- data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.h +77 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-bitmap.h +24 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-dma.h +2 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-utils.h +65 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1378 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.c +170 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.h +160 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-ctx.h +146 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-ops.h +242 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.c +300 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.h +29 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp_iface.idl +23 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-arith.h +636 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-base.h +292 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-exp.h +255 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +303 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-log.h +118 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-norm.h +454 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-quant.h +165 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-reduce.h +336 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-scale.h +199 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +181 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +189 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-utils.h +24 -0
- data/vendor/ggml/src/ggml-hexagon/htp/im2col-ops.c +306 -0
- data/vendor/ggml/src/ggml-hexagon/htp/main.c +1245 -0
- data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.c +4078 -0
- data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.h +778 -0
- data/vendor/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
- data/vendor/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
- data/vendor/ggml/src/ggml-hexagon/htp/rope-ops.c +733 -0
- data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.c +230 -0
- data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.h +74 -0
- data/vendor/ggml/src/ggml-hexagon/htp/softmax-ops.c +407 -0
- data/vendor/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
- data/vendor/ggml/src/ggml-hexagon/htp/ssm-conv.c +433 -0
- data/vendor/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.c +1264 -0
- data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.h +176 -0
- data/vendor/ggml/src/ggml-hexagon/htp/work-queue.c +244 -0
- data/vendor/ggml/src/ggml-hexagon/htp/work-queue.h +38 -0
- data/vendor/ggml/src/ggml-hexagon/htp-drv.cpp +425 -0
- data/vendor/ggml/src/ggml-hexagon/htp-drv.h +123 -0
- data/vendor/ggml/src/ggml-hexagon/htp-opnode.h +372 -0
- data/vendor/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/vendor/ggml/src/ggml-hexagon/libggml-htp.inf +36 -0
- data/vendor/ggml/src/ggml-hip/CMakeLists.txt +153 -0
- data/vendor/ggml/src/ggml-impl.h +783 -0
- data/vendor/ggml/src/ggml-metal/CMakeLists.txt +240 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-common.cpp +474 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-common.h +56 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-context.m +759 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.cpp +2330 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.h +355 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.m +2473 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-impl.h +1345 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-ops.cpp +5610 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-ops.h +106 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.cpp +3732 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.h +77 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal.cpp +1006 -0
- data/vendor/ggml/src/ggml-metal/kernels/argsort.metal +337 -0
- data/vendor/ggml/src/ggml-metal/kernels/binbcast.metal +228 -0
- data/vendor/ggml/src/ggml-metal/kernels/common.h +126 -0
- data/vendor/ggml/src/ggml-metal/kernels/conv.metal +724 -0
- data/vendor/ggml/src/ggml-metal/kernels/dequantize.h +735 -0
- data/vendor/ggml/src/ggml-metal/kernels/fa.metal +2445 -0
- data/vendor/ggml/src/ggml-metal/kernels/gated_delta_net.metal +250 -0
- data/vendor/ggml/src/ggml-metal/kernels/misc.metal +595 -0
- data/vendor/ggml/src/ggml-metal/kernels/mul_mm.metal +853 -0
- data/vendor/ggml/src/ggml-metal/kernels/mul_mv.metal +3225 -0
- data/vendor/ggml/src/ggml-metal/kernels/norm.metal +308 -0
- data/vendor/ggml/src/ggml-metal/kernels/pool.metal +148 -0
- data/vendor/ggml/src/ggml-metal/kernels/quantize.h +262 -0
- data/vendor/ggml/src/ggml-metal/kernels/quantize.metal +480 -0
- data/vendor/ggml/src/ggml-metal/kernels/reduce.metal +228 -0
- data/vendor/ggml/src/ggml-metal/kernels/rope.metal +333 -0
- data/vendor/ggml/src/ggml-metal/kernels/softmax.metal +223 -0
- data/vendor/ggml/src/ggml-metal/kernels/solve_tri.metal +75 -0
- data/vendor/ggml/src/ggml-metal/kernels/ssm.metal +473 -0
- data/vendor/ggml/src/ggml-metal/kernels/tri.metal +69 -0
- data/vendor/ggml/src/ggml-metal/kernels/unary.metal +400 -0
- data/vendor/ggml/src/ggml-metal/kernels/upscale.metal +179 -0
- data/vendor/ggml/src/ggml-metal/kernels/wkv.metal +179 -0
- data/vendor/ggml/src/ggml-musa/CMakeLists.txt +123 -0
- data/vendor/ggml/src/ggml-musa/mudnn.cu +112 -0
- data/vendor/ggml/src/ggml-musa/mudnn.cuh +12 -0
- data/vendor/ggml/src/ggml-opencl/CMakeLists.txt +246 -0
- data/vendor/ggml/src/ggml-opencl/cl-program-cache.cpp +453 -0
- data/vendor/ggml/src/ggml-opencl/cl-program-cache.h +75 -0
- data/vendor/ggml/src/ggml-opencl/fa_tune.h +92 -0
- data/vendor/ggml/src/ggml-opencl/ggml-opencl.cpp +26982 -0
- data/vendor/ggml/src/ggml-opencl/kernels/abs.cl +113 -0
- data/vendor/ggml/src/ggml-opencl/kernels/add.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
- data/vendor/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
- data/vendor/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
- data/vendor/ggml/src/ggml-opencl/kernels/concat.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
- data/vendor/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cpy.cl +288 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cvt.cl +2663 -0
- data/vendor/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/vendor/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
- data/vendor/ggml/src/ggml-opencl/kernels/div.cl +138 -0
- data/vendor/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
- data/vendor/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
- data/vendor/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +421 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +422 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +2656 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +2011 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1840 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +376 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_q8_1_dp4a.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_q8_1_dp4a.cl +169 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_q8_1_dp4a.cl +209 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_q8_1_dp4a.cl +200 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_0_f32_ns.cl +221 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_1_dp4a.cl +226 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_q8_1_dp4a.cl +143 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_f32.cl +139 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_q8_1_dp4a.cl +127 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +490 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_q8_1_dp4a.cl +281 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_q8_1_dp4a.cl +235 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_q8_1_dp4a.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +245 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32_tiled.cl +136 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_q8_1_dp4a.cl +144 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_q8_1_dp4a.cl +212 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +257 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +266 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +307 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl +383 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +384 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +823 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_o4.cl +349 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_tiled.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +296 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +299 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +453 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +409 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_o4.cl +372 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_tiled.cl +196 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +280 -0
- data/vendor/ggml/src/ggml-opencl/kernels/get_rows.cl +193 -0
- data/vendor/ggml/src/ggml-opencl/kernels/glu.cl +443 -0
- data/vendor/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
- data/vendor/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
- data/vendor/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
- data/vendor/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_add_id_glu.cl +76 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_combine.cl +79 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_quant_a_q8_1.cl +64 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +155 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +196 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +189 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +101 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1237 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_mrow.cl +306 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +182 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +205 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +244 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/norm.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
- data/vendor/ggml/src/ggml-opencl/kernels/quant_a_q8_1.cl +42 -0
- data/vendor/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
- data/vendor/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
- data/vendor/ggml/src/ggml-opencl/kernels/rms_norm.cl +369 -0
- data/vendor/ggml/src/ggml-opencl/kernels/rope.cl +759 -0
- data/vendor/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
- data/vendor/ggml/src/ggml-opencl/kernels/set_rows.cl +708 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
- data/vendor/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
- data/vendor/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
- data/vendor/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
- data/vendor/ggml/src/ggml-opencl/kernels/ssm_scan.cl +216 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
- data/vendor/ggml/src/ggml-opencl/kernels/transpose.cl +143 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
- data/vendor/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
- data/vendor/ggml/src/ggml-opencl/libdl.h +79 -0
- data/vendor/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/vendor/ggml/src/ggml-openvino/ggml-decoder.cpp +2057 -0
- data/vendor/ggml/src/ggml-openvino/ggml-decoder.h +417 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +493 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.h +216 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino.cpp +1566 -0
- data/vendor/ggml/src/ggml-openvino/ggml-quants.cpp +1273 -0
- data/vendor/ggml/src/ggml-openvino/ggml-quants.h +186 -0
- data/vendor/ggml/src/ggml-openvino/model-cache.cpp +272 -0
- data/vendor/ggml/src/ggml-openvino/model-cache.h +56 -0
- data/vendor/ggml/src/ggml-openvino/openvino/decoder.h +133 -0
- data/vendor/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/vendor/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/node_context.h +171 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/add.cpp +45 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cont.cpp +40 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cpy.cpp +295 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cumsum.cpp +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/diag.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/fill.cpp +34 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +241 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +327 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gather_matmul.hpp +43 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +142 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +75 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu_quick.cpp +64 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +110 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +61 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +283 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +100 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/permute.cpp +147 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/pool_2d.cpp +53 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/repeat.cpp +48 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/reshape.cpp +113 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +81 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/roll.cpp +36 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/rope.cpp +303 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/scale.cpp +86 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/set.cpp +76 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +95 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/softmax.cpp +166 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/solve_tri.cpp +108 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/sqr.cpp +35 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +61 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/transpose.cpp +54 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/tri.cpp +82 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/view.cpp +245 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op_table.cpp +88 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op_table.h +67 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp +212 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_dequantization_subgraph.h +44 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
- data/vendor/ggml/src/ggml-openvino/openvino/translate_session.cpp +432 -0
- data/vendor/ggml/src/ggml-openvino/openvino/translate_session.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/utils.cpp +886 -0
- data/vendor/ggml/src/ggml-openvino/openvino/utils.h +88 -0
- data/vendor/ggml/src/ggml-openvino/utils.cpp +1496 -0
- data/vendor/ggml/src/ggml-openvino/utils.h +195 -0
- data/vendor/ggml/src/ggml-opt.cpp +1094 -0
- data/vendor/ggml/src/ggml-quants.c +5667 -0
- data/vendor/ggml/src/ggml-quants.h +115 -0
- data/vendor/ggml/src/ggml-rpc/CMakeLists.txt +49 -0
- data/vendor/ggml/src/ggml-rpc/ggml-rpc.cpp +2351 -0
- data/vendor/ggml/src/ggml-rpc/transport-apple.cpp +481 -0
- data/vendor/ggml/src/ggml-rpc/transport-apple.h +27 -0
- data/vendor/ggml/src/ggml-rpc/transport.cpp +739 -0
- data/vendor/ggml/src/ggml-rpc/transport.h +38 -0
- data/vendor/ggml/src/ggml-sycl/CMakeLists.txt +218 -0
- data/vendor/ggml/src/ggml-sycl/add-id.cpp +81 -0
- data/vendor/ggml/src/ggml-sycl/add-id.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/backend.hpp +52 -0
- data/vendor/ggml/src/ggml-sycl/base.hpp +36 -0
- data/vendor/ggml/src/ggml-sycl/binbcast.cpp +650 -0
- data/vendor/ggml/src/ggml-sycl/binbcast.hpp +69 -0
- data/vendor/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
- data/vendor/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/common.cpp +153 -0
- data/vendor/ggml/src/ggml-sycl/common.hpp +1029 -0
- data/vendor/ggml/src/ggml-sycl/concat.cpp +509 -0
- data/vendor/ggml/src/ggml-sycl/concat.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/conv.cpp +101 -0
- data/vendor/ggml/src/ggml-sycl/conv.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-dw.cpp +170 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv2d.cpp +150 -0
- data/vendor/ggml/src/ggml-sycl/conv2d.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv3d.cpp +224 -0
- data/vendor/ggml/src/ggml-sycl/conv3d.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/convert.cpp +877 -0
- data/vendor/ggml/src/ggml-sycl/convert.hpp +64 -0
- data/vendor/ggml/src/ggml-sycl/count-equal.cpp +79 -0
- data/vendor/ggml/src/ggml-sycl/count-equal.hpp +9 -0
- data/vendor/ggml/src/ggml-sycl/cpy.cpp +1433 -0
- data/vendor/ggml/src/ggml-sycl/cpy.hpp +537 -0
- data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
- data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
- data/vendor/ggml/src/ggml-sycl/cumsum.cpp +148 -0
- data/vendor/ggml/src/ggml-sycl/cumsum.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/dequantize.hpp +1681 -0
- data/vendor/ggml/src/ggml-sycl/diag.cpp +67 -0
- data/vendor/ggml/src/ggml-sycl/diag.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/dmmv.cpp +2227 -0
- data/vendor/ggml/src/ggml-sycl/dmmv.hpp +27 -0
- data/vendor/ggml/src/ggml-sycl/dpct/helper.hpp +3782 -0
- data/vendor/ggml/src/ggml-sycl/dsv4-hc.cpp +280 -0
- data/vendor/ggml/src/ggml-sycl/dsv4-hc.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/element_wise.cpp +1432 -0
- data/vendor/ggml/src/ggml-sycl/element_wise.hpp +135 -0
- data/vendor/ggml/src/ggml-sycl/esimd.hpp +589 -0
- data/vendor/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
- data/vendor/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
- data/vendor/ggml/src/ggml-sycl/fattn-common.hpp +1185 -0
- data/vendor/ggml/src/ggml-sycl/fattn-mkl.cpp +689 -0
- data/vendor/ggml/src/ggml-sycl/fattn-onednn.cpp +450 -0
- data/vendor/ggml/src/ggml-sycl/fattn-onednn.hpp +18 -0
- data/vendor/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
- data/vendor/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
- data/vendor/ggml/src/ggml-sycl/fattn-vec.hpp +684 -0
- data/vendor/ggml/src/ggml-sycl/fattn.cpp +452 -0
- data/vendor/ggml/src/ggml-sycl/fattn.hpp +42 -0
- data/vendor/ggml/src/ggml-sycl/fill.cpp +55 -0
- data/vendor/ggml/src/ggml-sycl/fill.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/fusion.cpp +212 -0
- data/vendor/ggml/src/ggml-sycl/fusion.hpp +17 -0
- data/vendor/ggml/src/ggml-sycl/fwht.cpp +119 -0
- data/vendor/ggml/src/ggml-sycl/fwht.hpp +12 -0
- data/vendor/ggml/src/ggml-sycl/gated_delta_net.cpp +366 -0
- data/vendor/ggml/src/ggml-sycl/gated_delta_net.hpp +19 -0
- data/vendor/ggml/src/ggml-sycl/gemm.hpp +93 -0
- data/vendor/ggml/src/ggml-sycl/getrows.cpp +368 -0
- data/vendor/ggml/src/ggml-sycl/getrows.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/ggml-sycl.cpp +7028 -0
- data/vendor/ggml/src/ggml-sycl/gla.cpp +106 -0
- data/vendor/ggml/src/ggml-sycl/gla.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/im2col.cpp +400 -0
- data/vendor/ggml/src/ggml-sycl/im2col.hpp +23 -0
- data/vendor/ggml/src/ggml-sycl/lightning-indexer.cpp +197 -0
- data/vendor/ggml/src/ggml-sycl/lightning-indexer.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/mem.cpp +162 -0
- data/vendor/ggml/src/ggml-sycl/mem.hpp +16 -0
- data/vendor/ggml/src/ggml-sycl/mmq.cpp +3030 -0
- data/vendor/ggml/src/ggml-sycl/mmq.hpp +33 -0
- data/vendor/ggml/src/ggml-sycl/mmvq.cpp +3033 -0
- data/vendor/ggml/src/ggml-sycl/mmvq.hpp +76 -0
- data/vendor/ggml/src/ggml-sycl/norm.cpp +963 -0
- data/vendor/ggml/src/ggml-sycl/norm.hpp +32 -0
- data/vendor/ggml/src/ggml-sycl/opt-step.cpp +131 -0
- data/vendor/ggml/src/ggml-sycl/opt-step.hpp +6 -0
- data/vendor/ggml/src/ggml-sycl/outprod.cpp +83 -0
- data/vendor/ggml/src/ggml-sycl/outprod.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/pad.cpp +97 -0
- data/vendor/ggml/src/ggml-sycl/pad.hpp +24 -0
- data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.cpp +100 -0
- data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/pool.cpp +185 -0
- data/vendor/ggml/src/ggml-sycl/pool.hpp +22 -0
- data/vendor/ggml/src/ggml-sycl/presets.hpp +79 -0
- data/vendor/ggml/src/ggml-sycl/quantize.hpp +133 -0
- data/vendor/ggml/src/ggml-sycl/quants.hpp +204 -0
- data/vendor/ggml/src/ggml-sycl/repeat_back.cpp +76 -0
- data/vendor/ggml/src/ggml-sycl/repeat_back.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/roll.cpp +122 -0
- data/vendor/ggml/src/ggml-sycl/roll.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/rope.cpp +651 -0
- data/vendor/ggml/src/ggml-sycl/rope.hpp +26 -0
- data/vendor/ggml/src/ggml-sycl/set.cpp +73 -0
- data/vendor/ggml/src/ggml-sycl/set.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/set_rows.cpp +577 -0
- data/vendor/ggml/src/ggml-sycl/set_rows.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/softmax.cpp +425 -0
- data/vendor/ggml/src/ggml-sycl/softmax.hpp +24 -0
- data/vendor/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
- data/vendor/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/ssm_conv.cpp +136 -0
- data/vendor/ggml/src/ggml-sycl/ssm_conv.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/ssm_scan.cpp +170 -0
- data/vendor/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/sycl_hw.cpp +67 -0
- data/vendor/ggml/src/ggml-sycl/sycl_hw.hpp +38 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/topk-moe.cpp +620 -0
- data/vendor/ggml/src/ggml-sycl/topk-moe.hpp +12 -0
- data/vendor/ggml/src/ggml-sycl/tsembd.cpp +73 -0
- data/vendor/ggml/src/ggml-sycl/tsembd.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/type.hpp +112 -0
- data/vendor/ggml/src/ggml-sycl/upscale.cpp +410 -0
- data/vendor/ggml/src/ggml-sycl/upscale.hpp +9 -0
- data/vendor/ggml/src/ggml-sycl/vecdotq.hpp +1726 -0
- data/vendor/ggml/src/ggml-sycl/wkv.cpp +293 -0
- data/vendor/ggml/src/ggml-sycl/wkv.hpp +10 -0
- data/vendor/ggml/src/ggml-threading.cpp +12 -0
- data/vendor/ggml/src/ggml-threading.h +14 -0
- data/vendor/ggml/src/ggml-version.h.in +4 -0
- data/vendor/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/vendor/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +149 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend.cpp +72 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/vendor/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/vendor/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/vendor/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +194 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +54 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
- data/vendor/ggml/src/ggml-vulkan/CMakeLists.txt +254 -0
- data/vendor/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
- data/vendor/ggml/src/ggml-vulkan/ggml-vulkan.cpp +20349 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +69 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp +28 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp +20 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +86 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp +114 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +53 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +473 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +424 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +25 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +350 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp +67 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose_02.comp +61 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +135 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp +78 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp +75 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +83 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +741 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +1466 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.glsl +13 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +49 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +40 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_0.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_tq2_0.comp +31 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +28 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fa_types.glsl +55 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/bfloat16.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e2m1.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e4m3.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/integer_dot.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +760 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +223 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +647 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +487 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +146 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +121 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.glsl +11 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +75 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gla.comp +82 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +138 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +124 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/lightning_indexer.comp +151 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +264 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +230 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +132 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +95 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +105 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +93 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +124 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +156 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +128 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +134 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +165 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_tq2_0.comp +102 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +141 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +564 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +475 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +690 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +670 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +89 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +317 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +488 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +99 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +194 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/out_prod.comp +59 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +64 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool1d.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +127 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +150 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +218 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +195 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +54 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp +62 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp +79 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large_common.glsl +53 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp +81 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp +134 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +47 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_clamp.comp +12 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp +118 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +221 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp +246 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_radix_select.comp +144 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +1938 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +178 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/utils.glsl +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +1367 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
- data/vendor/ggml/src/ggml-webgpu/CMakeLists.txt +83 -0
- data/vendor/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3457 -0
- data/vendor/ggml/src/ggml-webgpu/ggml-webgpu.cpp +4790 -0
- data/vendor/ggml/src/ggml-webgpu/pre_wgsl.hpp +808 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +102 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +967 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +92 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +127 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d_dw.wgsl +100 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +84 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +89 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +408 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_decls.tmpl +134 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_staging.tmpl +136 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +240 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +503 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +794 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +178 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +77 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +1052 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +159 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +210 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +167 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1560 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +70 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +151 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rope.wgsl +228 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +152 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +65 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +108 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.wgsl +210 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +120 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +247 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/vendor/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
- data/vendor/ggml/src/ggml-zdnn/common.hpp +59 -0
- data/vendor/ggml/src/ggml-zdnn/ggml-zdnn.cpp +638 -0
- data/vendor/ggml/src/ggml-zdnn/mmf.cpp +80 -0
- data/vendor/ggml/src/ggml-zdnn/mmf.hpp +12 -0
- data/vendor/ggml/src/ggml-zdnn/utils.cpp +79 -0
- data/vendor/ggml/src/ggml-zdnn/utils.hpp +19 -0
- data/vendor/ggml/src/ggml-zendnn/CMakeLists.txt +91 -0
- data/vendor/ggml/src/ggml-zendnn/ggml-zendnn.cpp +838 -0
- data/vendor/ggml/src/ggml.c +8088 -0
- data/vendor/ggml/src/ggml.cpp +26 -0
- data/vendor/ggml/src/gguf.cpp +1706 -0
- metadata +1459 -0
|
@@ -0,0 +1,1496 @@
|
|
|
1
|
+
#include "utils.h"
|
|
2
|
+
|
|
3
|
+
#include "ggml-impl.h"
|
|
4
|
+
#include "ggml-openvino-extra.h"
|
|
5
|
+
#include "ggml-openvino/ggml-decoder.h"
|
|
6
|
+
#include "ggml.h"
|
|
7
|
+
#include "model-cache.h"
|
|
8
|
+
#include "openvino/frontend.h"
|
|
9
|
+
#include "openvino/input_model.h"
|
|
10
|
+
|
|
11
|
+
#include <algorithm>
|
|
12
|
+
#include <cassert>
|
|
13
|
+
#include <cmath>
|
|
14
|
+
#include <cstddef>
|
|
15
|
+
#include <cstdint>
|
|
16
|
+
#include <cstdlib>
|
|
17
|
+
#include <cstring>
|
|
18
|
+
#include <fstream>
|
|
19
|
+
#include <functional>
|
|
20
|
+
#include <future>
|
|
21
|
+
#include <iomanip>
|
|
22
|
+
#include <iostream>
|
|
23
|
+
#include <memory>
|
|
24
|
+
#include <openvino/core/any.hpp>
|
|
25
|
+
#include <openvino/core/graph_util.hpp>
|
|
26
|
+
#include <openvino/core/shape.hpp>
|
|
27
|
+
#include <openvino/core/type/float16.hpp>
|
|
28
|
+
#include <openvino/frontend/manager.hpp>
|
|
29
|
+
#include <openvino/openvino.hpp>
|
|
30
|
+
#include <openvino/runtime/compiled_model.hpp>
|
|
31
|
+
#include <openvino/runtime/infer_request.hpp>
|
|
32
|
+
#include <openvino/runtime/intel_gpu/ocl/ocl.hpp>
|
|
33
|
+
#include <openvino/runtime/intel_npu/properties.hpp>
|
|
34
|
+
#include <openvino/runtime/properties.hpp>
|
|
35
|
+
#include <openvino/runtime/tensor.hpp>
|
|
36
|
+
#include <optional>
|
|
37
|
+
#include <string>
|
|
38
|
+
#include <unordered_map>
|
|
39
|
+
#include <vector>
|
|
40
|
+
|
|
41
|
+
// Suppress deprecation warning for ov::Tensor::data()
|
|
42
|
+
#pragma GCC diagnostic push
|
|
43
|
+
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
|
|
44
|
+
|
|
45
|
+
enum ggml_status ov_graph_compute(ggml_cgraph * cgraph, ggml_backend_t backend) {
|
|
46
|
+
ggml_backend_openvino_context * ctx = (ggml_backend_openvino_context *) backend->context;
|
|
47
|
+
try {
|
|
48
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DUMP_CGRAPH")) {
|
|
49
|
+
std::string filename = "cgraph_ov.txt";
|
|
50
|
+
GgmlOvDecoder::dump_cgraph(cgraph, filename);
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
const auto is_static = ggml_openvino_is_npu() || ggml_openvino_getenv_int("GGML_OPENVINO_FORCE_STATIC");
|
|
54
|
+
|
|
55
|
+
GGML_ASSERT(ctx->runtime_context != nullptr);
|
|
56
|
+
std::shared_ptr<ov_runtime_context> r_ctx = std::static_pointer_cast<ov_runtime_context>(ctx->runtime_context);
|
|
57
|
+
|
|
58
|
+
return is_static ? ov_graph_compute_static(cgraph, r_ctx) : ov_graph_compute_dynamic(cgraph, r_ctx);
|
|
59
|
+
} catch (const ov::Exception & e) {
|
|
60
|
+
GGML_LOG_ERROR("GGML OpenVINO backend ov::Exception: %s\n", e.what());
|
|
61
|
+
return GGML_STATUS_FAILED;
|
|
62
|
+
} catch (const std::exception & e) {
|
|
63
|
+
GGML_LOG_ERROR("GGML OpenVINO backend std::exception: %s\n", e.what());
|
|
64
|
+
return GGML_STATUS_FAILED;
|
|
65
|
+
} catch (...) {
|
|
66
|
+
GGML_LOG_ERROR("GGML OpenVINO backend unknown exception\n");
|
|
67
|
+
return GGML_STATUS_FAILED;
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
|
|
71
|
+
// For a KV cache input, return an ov::Tensor sized to n_kv (== attention_size
|
|
72
|
+
// for that layer) instead of the fully-allocated ctx_per_seq. Pre-conditions:
|
|
73
|
+
// * non-static (CPU/GPU) backend, single sequence, seq_active_start == 0
|
|
74
|
+
// * ggml KV layout is a contiguous [1, 1, ctx_per_seq, n_heads_kv*head_size]
|
|
75
|
+
// so the first n_kv rows are the live prefix and shrinking the ctx axis
|
|
76
|
+
// gives a valid tensor over the same host storage
|
|
77
|
+
// * not an SWA layer (ring cache): once the window has wrapped the first
|
|
78
|
+
// n_kv rows no longer contain the live prefix
|
|
79
|
+
// On any unmet pre-condition returns std::nullopt; the caller falls back to
|
|
80
|
+
// the full-size tensor.
|
|
81
|
+
static std::optional<ov::Tensor> try_make_kv_sliced_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
|
|
82
|
+
const std::string & name,
|
|
83
|
+
const ggml_tensor * ggml_tensor) {
|
|
84
|
+
static const bool kv_slice_disabled = ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_KV_SLICE");
|
|
85
|
+
if (kv_slice_disabled) {
|
|
86
|
+
return std::nullopt;
|
|
87
|
+
}
|
|
88
|
+
if (ggml_decoder->is_static() || ggml_decoder->is_stateful()) {
|
|
89
|
+
return std::nullopt;
|
|
90
|
+
}
|
|
91
|
+
if (ggml_tensor->op != GGML_OP_NONE || ggml_tensor->view_src != nullptr) {
|
|
92
|
+
return std::nullopt;
|
|
93
|
+
}
|
|
94
|
+
const auto * op = ggml_decoder->get_tensor_used_op(ggml_tensor);
|
|
95
|
+
if (!GgmlOvDecoder::is_kvcache(ggml_tensor, op)) {
|
|
96
|
+
return std::nullopt;
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
const auto & compute_params = ggml_decoder->get_compute_params();
|
|
100
|
+
if (compute_params.n_seq_active != 1 || compute_params.seq_active_start != 0) {
|
|
101
|
+
return std::nullopt;
|
|
102
|
+
}
|
|
103
|
+
|
|
104
|
+
int layer;
|
|
105
|
+
if (auto layer_opt = extract_layer_from_name(name); layer_opt.has_value()) {
|
|
106
|
+
layer = layer_opt.value();
|
|
107
|
+
} else {
|
|
108
|
+
return std::nullopt;
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
const bool is_swa = ggml_decoder->is_swa_layer(layer);
|
|
112
|
+
if (is_swa) {
|
|
113
|
+
return std::nullopt;
|
|
114
|
+
}
|
|
115
|
+
const int ctx_per_seq = ggml_decoder->get_ctx_per_seq();
|
|
116
|
+
const int n_kv = compute_params.attention_size;
|
|
117
|
+
if (ctx_per_seq <= 0 || n_kv <= 0 || n_kv >= ctx_per_seq) {
|
|
118
|
+
return std::nullopt;
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
ov::Shape full_shape = ggml_decoder->get_shape(ggml_tensor);
|
|
122
|
+
if (full_shape.size() != 4 || full_shape[0] != 1 || full_shape[1] != 1 ||
|
|
123
|
+
static_cast<int>(full_shape[2]) != ctx_per_seq) {
|
|
124
|
+
return std::nullopt;
|
|
125
|
+
}
|
|
126
|
+
|
|
127
|
+
ov::Shape sliced_shape = full_shape;
|
|
128
|
+
sliced_shape[2] = static_cast<size_t>(n_kv);
|
|
129
|
+
|
|
130
|
+
// Disabling for now as gpu has bug with in-place ScatterUpdate with remote tensors, can re-enable once CVS-186519 is fixed
|
|
131
|
+
// if (ggml_openvino_buffer_is_remote(ggml_tensor)) {
|
|
132
|
+
// auto remote_context = ggml_openvino_get_remote_context();
|
|
133
|
+
// auto gpu_context = remote_context->as<ov::intel_gpu::ocl::ClContext>();
|
|
134
|
+
// return gpu_context.create_tensor(ggml_decoder->get_ov_type(ggml_tensor), sliced_shape, ggml_tensor->data);
|
|
135
|
+
// }
|
|
136
|
+
|
|
137
|
+
return ov::Tensor(ggml_decoder->get_ov_type(ggml_tensor), sliced_shape, ggml_tensor->data);
|
|
138
|
+
}
|
|
139
|
+
|
|
140
|
+
static uint64_t ggml_openvino_model_cache_extra_cfg(const std::string & device, bool stateful) {
|
|
141
|
+
const char * manual_gqa_env = ggml_openvino_getenv_str("GGML_OPENVINO_MANUAL_GQA_ATTN");
|
|
142
|
+
const bool manual_gqa_enabled = manual_gqa_env != nullptr ?
|
|
143
|
+
ggml_openvino_getenv_int("GGML_OPENVINO_MANUAL_GQA_ATTN") > 0 :
|
|
144
|
+
device == "GPU";
|
|
145
|
+
|
|
146
|
+
uint64_t extra_cfg = 0;
|
|
147
|
+
extra_cfg = extra_cfg * 131 + (stateful ? 1u : 0u);
|
|
148
|
+
extra_cfg = extra_cfg * 131 + (ggml_openvino_reduce_compile_mem_enabled() ? 1u : 0u);
|
|
149
|
+
extra_cfg = extra_cfg * 131 + (ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_KV_SLICE") ? 1u : 0u);
|
|
150
|
+
extra_cfg = extra_cfg * 131 + (manual_gqa_enabled ? 1u : 0u);
|
|
151
|
+
return extra_cfg;
|
|
152
|
+
}
|
|
153
|
+
|
|
154
|
+
ov::Tensor create_ov_output_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
|
|
155
|
+
std::shared_ptr<ov::InferRequest> infer_request,
|
|
156
|
+
int output_index,
|
|
157
|
+
const ggml_tensor * ggml_tensor) {
|
|
158
|
+
if (auto sliced = try_make_kv_sliced_tensor(ggml_decoder, std::string(ggml_tensor->name), ggml_tensor)) {
|
|
159
|
+
return *sliced;
|
|
160
|
+
}
|
|
161
|
+
|
|
162
|
+
// Disabling for now as gpu has bug with in-place ScatterUpdate with remote tensors, can re-enable once CVS-186519 is fixed
|
|
163
|
+
// if (ggml_tensor->extra != nullptr && !ggml_decoder->is_splited_model()) {
|
|
164
|
+
// auto * extra_base = static_cast<ggml_openvino_extra_base *>(ggml_tensor->extra);
|
|
165
|
+
// if (extra_base->type == ggml_openvino_extra_base::Type::TENSOR) {
|
|
166
|
+
// auto * tensor_extra = static_cast<ggml_openvino_tensor_extra *>(extra_base);
|
|
167
|
+
// return *tensor_extra->tensor;
|
|
168
|
+
// }
|
|
169
|
+
// }
|
|
170
|
+
|
|
171
|
+
auto output_type = ggml_decoder->get_ov_type(ggml_tensor);
|
|
172
|
+
ov::Shape output_shape;
|
|
173
|
+
void * output_data = ggml_tensor->data;
|
|
174
|
+
if (ggml_decoder->is_static()) {
|
|
175
|
+
output_shape = infer_request->get_output_tensor(output_index).get_shape();
|
|
176
|
+
} else {
|
|
177
|
+
// For a CPY into a padded view_src (e.g. a padded KV cache buffer), the
|
|
178
|
+
// OV ScatterUpdate node outputs the full view_src shape, not the CPY node's
|
|
179
|
+
// own (smaller) shape. Using the CPY shape here causes set_output_tensor to
|
|
180
|
+
// fail with a shape-incompatibility error. Use view_src's shape and data
|
|
181
|
+
// pointer instead so the OV tensor matches the model output exactly.
|
|
182
|
+
if (ggml_tensor->op == GGML_OP_CPY && ggml_tensor->view_src != nullptr &&
|
|
183
|
+
ggml_nbytes(ggml_tensor) != ggml_nbytes(ggml_tensor->view_src)) {
|
|
184
|
+
output_shape = ggml_decoder->get_shape(ggml_tensor->view_src);
|
|
185
|
+
output_data = ggml_tensor->view_src->data;
|
|
186
|
+
} else {
|
|
187
|
+
output_shape = ggml_decoder->get_shape(ggml_tensor);
|
|
188
|
+
}
|
|
189
|
+
}
|
|
190
|
+
ov::Tensor output_tensor(output_type, output_shape, output_data);
|
|
191
|
+
return output_tensor;
|
|
192
|
+
}
|
|
193
|
+
|
|
194
|
+
enum ggml_status ov_graph_compute_dynamic(ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx) {
|
|
195
|
+
auto & core = ov_singleton_core();
|
|
196
|
+
const auto & config = ggml_openvino_get_compile_config();
|
|
197
|
+
const auto & device = r_ctx->device;
|
|
198
|
+
const auto & stateful = r_ctx->stateful;
|
|
199
|
+
static auto is_static = false;
|
|
200
|
+
|
|
201
|
+
static const bool cache_disabled = ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_CACHE");
|
|
202
|
+
|
|
203
|
+
// is_model_splitted is O(n_nodes^2) plus a create_weight_nodes scan and takes ~20 ms
|
|
204
|
+
// on a Llama-1B decode graph. It is called once per graph_compute invocation but the
|
|
205
|
+
// graph shape is identical across all decode steps, so memoize by graph_key: compute
|
|
206
|
+
// graph_key first (a few hundred us), and if the same key is already in decoder_cache
|
|
207
|
+
// we know the graph is not splitted (only not-splitted graphs get inserted there).
|
|
208
|
+
graph_key key(cgraph);
|
|
209
|
+
bool key_seen = false;
|
|
210
|
+
if (!cache_disabled) {
|
|
211
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
212
|
+
key_seen = r_ctx->decoder_cache.find(key) != r_ctx->decoder_cache.end();
|
|
213
|
+
}
|
|
214
|
+
|
|
215
|
+
bool model_is_splitted = key_seen ? false : is_model_splitted(cgraph);
|
|
216
|
+
|
|
217
|
+
if (is_naive(cgraph)) {
|
|
218
|
+
if (!model_is_splitted) {
|
|
219
|
+
return naive_compute(cgraph, core, device, config);
|
|
220
|
+
}
|
|
221
|
+
}
|
|
222
|
+
|
|
223
|
+
auto start_time = ggml_time_us();
|
|
224
|
+
|
|
225
|
+
std::shared_ptr<GgmlOvDecoder> ggml_decoder;
|
|
226
|
+
std::shared_ptr<ov::InferRequest> infer_request;
|
|
227
|
+
ModelParams m_params;
|
|
228
|
+
ComputeParams c_params;
|
|
229
|
+
std::tie(m_params, c_params) = GgmlOvDecoder::compute_llm_params(cgraph, is_static);
|
|
230
|
+
|
|
231
|
+
const bool cache_enabled = !model_is_splitted && !cache_disabled;
|
|
232
|
+
bool cache_hit = false;
|
|
233
|
+
|
|
234
|
+
int64_t decoder_end_time;
|
|
235
|
+
int64_t conversion_end_time;
|
|
236
|
+
int64_t compile_end_time;
|
|
237
|
+
int64_t infer_end_time;
|
|
238
|
+
int64_t ov_raw_infer_start;
|
|
239
|
+
|
|
240
|
+
{
|
|
241
|
+
std::shared_ptr<decoder_runtime_ctx> entry;
|
|
242
|
+
ModelParams old_m_params;
|
|
243
|
+
|
|
244
|
+
if (cache_enabled) {
|
|
245
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
246
|
+
auto it = r_ctx->decoder_cache.find(key);
|
|
247
|
+
cache_hit = it != r_ctx->decoder_cache.end();
|
|
248
|
+
if (cache_hit) {
|
|
249
|
+
entry = it->second;
|
|
250
|
+
} else {
|
|
251
|
+
r_ctx->clear_caches_locked();
|
|
252
|
+
auto mutex = std::make_shared<std::mutex>();
|
|
253
|
+
entry = std::make_shared<decoder_runtime_ctx>(mutex);
|
|
254
|
+
r_ctx->decoder_cache[key] = entry;
|
|
255
|
+
}
|
|
256
|
+
} else {
|
|
257
|
+
auto mutex = std::make_shared<std::mutex>();
|
|
258
|
+
entry = std::make_shared<decoder_runtime_ctx>(mutex);
|
|
259
|
+
cache_hit = false;
|
|
260
|
+
}
|
|
261
|
+
|
|
262
|
+
std::lock_guard<std::mutex> lock(*(entry->mutex));
|
|
263
|
+
|
|
264
|
+
if (cache_hit) {
|
|
265
|
+
ggml_decoder = entry->ptr;
|
|
266
|
+
old_m_params = ggml_decoder->get_model_params();
|
|
267
|
+
if (!ggml_decoder->is_splited_model()) {
|
|
268
|
+
cache_hit = old_m_params.can_reuse_dynamically(m_params);
|
|
269
|
+
}
|
|
270
|
+
}
|
|
271
|
+
|
|
272
|
+
std::vector<std::string> ov_input_names;
|
|
273
|
+
std::vector<std::string> ov_output_names;
|
|
274
|
+
|
|
275
|
+
if (cache_hit) {
|
|
276
|
+
std::map<std::string, std::shared_ptr<ov::Node>> model_weights;
|
|
277
|
+
ggml_decoder->set_compute_params(c_params);
|
|
278
|
+
ggml_decoder->set_model_params(m_params);
|
|
279
|
+
if (old_m_params.kv_buffer_changed(m_params)) {
|
|
280
|
+
ggml_decoder->update_io(cgraph);
|
|
281
|
+
}
|
|
282
|
+
ggml_decoder->add_extra_inputs();
|
|
283
|
+
{
|
|
284
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
285
|
+
infer_request = r_ctx->infer_request_cache.at(key);
|
|
286
|
+
ov_input_names = r_ctx->ov_input_names_cache.at(key);
|
|
287
|
+
ov_output_names = r_ctx->ov_output_names_cache.at(key);
|
|
288
|
+
}
|
|
289
|
+
|
|
290
|
+
if (stateful) {
|
|
291
|
+
const auto * inp_pos = get_inp_pos_tensor(cgraph);
|
|
292
|
+
int32_t * pos_data = (int32_t *) inp_pos->data;
|
|
293
|
+
auto pos_shape = ggml_decoder->get_shape(inp_pos);
|
|
294
|
+
if (pos_data[0] == 0) {
|
|
295
|
+
infer_request->reset_state();
|
|
296
|
+
r_ctx->stateful_kv_size = pos_shape[3];
|
|
297
|
+
} else if (r_ctx->stateful_kv_size == static_cast<size_t>(pos_data[0])) {
|
|
298
|
+
r_ctx->stateful_kv_size += pos_shape[3];
|
|
299
|
+
} else {
|
|
300
|
+
auto states = infer_request->query_state();
|
|
301
|
+
for (auto state : states) {
|
|
302
|
+
auto state_tensor = state.get_state();
|
|
303
|
+
auto state_tensor_shape = state_tensor.get_shape();
|
|
304
|
+
if (static_cast<uint32_t>(pos_data[0]) > r_ctx->stateful_kv_size) {
|
|
305
|
+
std::string state_name;
|
|
306
|
+
try {
|
|
307
|
+
state_name = r_ctx->kv_state_input_name_map.at(state.get_name());
|
|
308
|
+
} catch (...) {
|
|
309
|
+
GGML_LOG_ERROR(
|
|
310
|
+
"GGML OpenVINO backend stateful inference failed: no input found for the state\n");
|
|
311
|
+
return GGML_STATUS_FAILED;
|
|
312
|
+
}
|
|
313
|
+
auto kv_tensor = get_ov_input_tensor(ggml_decoder, state_name);
|
|
314
|
+
kv_tensor.set_shape({state_tensor_shape[0], kv_tensor.get_shape()[2], state_tensor_shape[2],
|
|
315
|
+
state_tensor_shape[3]});
|
|
316
|
+
state_tensor = kv_tensor;
|
|
317
|
+
state_tensor_shape = state_tensor.get_shape();
|
|
318
|
+
}
|
|
319
|
+
ov::Coordinate begin = {0, 0, 0, 0};
|
|
320
|
+
ov::Coordinate end = {state_tensor_shape[0], static_cast<uint32_t>(pos_data[0]),
|
|
321
|
+
state_tensor_shape[2], state_tensor_shape[3]};
|
|
322
|
+
ov::Tensor new_state_tensor(state_tensor, begin, end);
|
|
323
|
+
state.set_state(new_state_tensor);
|
|
324
|
+
}
|
|
325
|
+
r_ctx->stateful_kv_size = pos_data[0] + pos_shape[3];
|
|
326
|
+
}
|
|
327
|
+
}
|
|
328
|
+
|
|
329
|
+
decoder_end_time = ggml_time_us();
|
|
330
|
+
conversion_end_time = decoder_end_time;
|
|
331
|
+
compile_end_time = decoder_end_time;
|
|
332
|
+
} else {
|
|
333
|
+
// Fail fast: a cache-miss recompile feeds weight data to compile_model, but
|
|
334
|
+
// GGML_OPENVINO_RELEASE_WEIGHTS (or GGML_OPENVINO_MEMORY_OPTIMIZE on GPU)
|
|
335
|
+
// may have already dropped the host weight pages
|
|
336
|
+
// (they would read as zeros). That mode requires stable graph shapes.
|
|
337
|
+
if (ggml_openvino_weight_buffers_released()) {
|
|
338
|
+
GGML_ABORT(
|
|
339
|
+
"ggml-openvino: a new graph needs to be compiled but host weight buffers were already "
|
|
340
|
+
"released via GGML_OPENVINO_RELEASE_WEIGHTS/GGML_OPENVINO_MEMORY_OPTIMIZE. This mode requires "
|
|
341
|
+
"stable graph shapes; disable host weight release for dynamic workloads.");
|
|
342
|
+
}
|
|
343
|
+
if (cache_enabled) {
|
|
344
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
345
|
+
r_ctx->infer_request_cache.erase(key);
|
|
346
|
+
}
|
|
347
|
+
|
|
348
|
+
// Frontend-level compiled-model cache (GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR): if this model
|
|
349
|
+
// was compiled before, import the saved blob and skip requant + convert +
|
|
350
|
+
// compile. Only the dynamic single-model path is cached (split models compile
|
|
351
|
+
// two graphs and are left to the plugin-level ov::cache_dir). The decoder is
|
|
352
|
+
// still needed for I/O mapping, but can be built without weight nodes since
|
|
353
|
+
// the weights are baked into the imported CompiledModel.
|
|
354
|
+
const std::string model_cache_dir = ggml_openvino_model_cache_dir();
|
|
355
|
+
uint64_t model_fp = 0;
|
|
356
|
+
std::string blob_path, manifest_path;
|
|
357
|
+
bool imported = false;
|
|
358
|
+
// When the frontend model cache is active it supersedes the plugin-level
|
|
359
|
+
// ov::cache_dir: a blob exported from a model compiled WITH cache_dir cannot
|
|
360
|
+
// be re-imported (import returns an uninitialized model). Strip cache_dir /
|
|
361
|
+
// cache_mode from the config used for the cached compile and the import.
|
|
362
|
+
ov::AnyMap mc_config = config;
|
|
363
|
+
if (!model_cache_dir.empty()) {
|
|
364
|
+
mc_config.erase("CACHE_DIR");
|
|
365
|
+
mc_config.erase("CACHE_MODE");
|
|
366
|
+
}
|
|
367
|
+
if (!model_cache_dir.empty() && !model_is_splitted) {
|
|
368
|
+
const uint64_t extra_cfg = ggml_openvino_model_cache_extra_cfg(device, stateful);
|
|
369
|
+
model_fp = ggml_openvino_model_fingerprint(cgraph, device, /*fa=*/true, m_params.rope_params,
|
|
370
|
+
15, extra_cfg);
|
|
371
|
+
blob_path = ggml_openvino_model_cache_blob_path(model_cache_dir, model_fp);
|
|
372
|
+
manifest_path = ggml_openvino_model_cache_manifest_path(model_cache_dir, model_fp);
|
|
373
|
+
|
|
374
|
+
std::ifstream blob_in(blob_path, std::ios::binary);
|
|
375
|
+
bool blob_ok = blob_in.is_open();
|
|
376
|
+
bool manifest_ok = blob_ok && ggml_openvino_model_cache_verify_manifest(manifest_path, cgraph, model_fp);
|
|
377
|
+
if (blob_ok && manifest_ok) {
|
|
378
|
+
int64_t import_start = ggml_time_us();
|
|
379
|
+
try {
|
|
380
|
+
ov::CompiledModel cm;
|
|
381
|
+
auto remote_context = ggml_openvino_get_remote_context();
|
|
382
|
+
if (remote_context.has_value()) {
|
|
383
|
+
cm = core.import_model(blob_in, remote_context.value(), mc_config);
|
|
384
|
+
} else {
|
|
385
|
+
cm = core.import_model(blob_in, device, mc_config);
|
|
386
|
+
}
|
|
387
|
+
// Lightweight decoder: names-only weight map (membership is all the
|
|
388
|
+
// decoder needs; weights live in the imported model).
|
|
389
|
+
std::map<std::string, std::shared_ptr<ov::Node>> weight_names;
|
|
390
|
+
for (const auto & n : GgmlOvDecoder::collect_weight_names(cgraph)) {
|
|
391
|
+
weight_names[n] = nullptr;
|
|
392
|
+
}
|
|
393
|
+
ggml_decoder = std::make_shared<GgmlOvDecoder>(cgraph, m_params, c_params, weight_names,
|
|
394
|
+
is_static, stateful, model_is_splitted);
|
|
395
|
+
infer_request = std::make_shared<ov::InferRequest>(cm.create_infer_request());
|
|
396
|
+
entry->ptr = ggml_decoder;
|
|
397
|
+
// Names must match the decoder's ggml-tensor keys. The non-cached
|
|
398
|
+
// path keys off Parameter/Result *friendly names* (set by the
|
|
399
|
+
// frontend); export_model preserves these, and each compiled-model
|
|
400
|
+
// port's node is exactly that Parameter/Result. Use the port nodes
|
|
401
|
+
// directly (NOT get_runtime_model(), whose graph differs and is
|
|
402
|
+
// unsafe to deref this way).
|
|
403
|
+
for (const auto & p : cm.inputs()) {
|
|
404
|
+
ov_input_names.push_back(p.get_node()->get_friendly_name());
|
|
405
|
+
}
|
|
406
|
+
for (const auto & o : cm.outputs()) {
|
|
407
|
+
ov_output_names.push_back(o.get_node()->get_friendly_name());
|
|
408
|
+
}
|
|
409
|
+
imported = true;
|
|
410
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_PROFILING")) {
|
|
411
|
+
GGML_LOG_INFO(" - Model cache import time: %.3f ms \n",
|
|
412
|
+
(ggml_time_us() - import_start) / 1000.0);
|
|
413
|
+
}
|
|
414
|
+
GGML_LOG_INFO("ggml-openvino: model cache HIT %s\n", blob_path.c_str());
|
|
415
|
+
} catch (const std::exception & e) {
|
|
416
|
+
GGML_LOG_WARN("ggml-openvino: model cache import failed (%s), recompiling\n", e.what());
|
|
417
|
+
imported = false;
|
|
418
|
+
}
|
|
419
|
+
}
|
|
420
|
+
}
|
|
421
|
+
|
|
422
|
+
std::shared_ptr<ov::Model> model;
|
|
423
|
+
if (imported) {
|
|
424
|
+
decoder_end_time = conversion_end_time = compile_end_time = ggml_time_us();
|
|
425
|
+
} else {
|
|
426
|
+
auto model_weights = GgmlOvDecoder::create_weight_nodes(cgraph);
|
|
427
|
+
|
|
428
|
+
ggml_decoder = std::make_shared<GgmlOvDecoder>(cgraph, m_params, c_params, model_weights, is_static,
|
|
429
|
+
stateful, model_is_splitted);
|
|
430
|
+
decoder_end_time = ggml_time_us();
|
|
431
|
+
|
|
432
|
+
auto input_model = std::make_shared<ov::frontend::ggml::InputModel>(ggml_decoder);
|
|
433
|
+
model = ov::frontend::ggml::FrontEnd::convert(input_model);
|
|
434
|
+
ggml_decoder->clear_model_weights();
|
|
435
|
+
conversion_end_time = ggml_time_us();
|
|
436
|
+
|
|
437
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DUMP_IR")) {
|
|
438
|
+
char timestamped_filename[64];
|
|
439
|
+
auto timestamp = (long long) ggml_time_us();
|
|
440
|
+
snprintf(timestamped_filename, sizeof(timestamped_filename), "model_%lld.xml", timestamp);
|
|
441
|
+
ov::serialize(model, timestamped_filename);
|
|
442
|
+
}
|
|
443
|
+
|
|
444
|
+
// Use the cache-stripped config when the frontend model cache is active, so
|
|
445
|
+
// the resulting CompiledModel can be exported and later re-imported.
|
|
446
|
+
const ov::AnyMap & compile_config = model_cache_dir.empty() ? config : mc_config;
|
|
447
|
+
ov::CompiledModel compiled_model;
|
|
448
|
+
auto remote_context = ggml_openvino_get_remote_context();
|
|
449
|
+
if (remote_context.has_value()) {
|
|
450
|
+
compiled_model = core.compile_model(model, remote_context.value(), compile_config);
|
|
451
|
+
} else {
|
|
452
|
+
compiled_model = core.compile_model(model, device, compile_config);
|
|
453
|
+
}
|
|
454
|
+
compile_end_time = ggml_time_us();
|
|
455
|
+
|
|
456
|
+
// Export to the frontend model cache for next time. Publish the blob first,
|
|
457
|
+
// then the manifest, so a cache hit only sees fully written artifacts.
|
|
458
|
+
if (!model_cache_dir.empty() && !model_is_splitted && model_fp != 0) {
|
|
459
|
+
try {
|
|
460
|
+
const std::string blob_tmp = blob_path + ".tmp";
|
|
461
|
+
const std::string manifest_tmp = manifest_path + ".tmp";
|
|
462
|
+
if (ggml_openvino_model_cache_write_manifest(manifest_tmp, cgraph, model_fp)) {
|
|
463
|
+
std::ofstream blob_out(blob_tmp, std::ios::binary | std::ios::trunc);
|
|
464
|
+
if (blob_out.is_open()) {
|
|
465
|
+
compiled_model.export_model(blob_out);
|
|
466
|
+
blob_out.close();
|
|
467
|
+
if (blob_out.good()) {
|
|
468
|
+
if (std::rename(blob_tmp.c_str(), blob_path.c_str()) == 0 &&
|
|
469
|
+
std::rename(manifest_tmp.c_str(), manifest_path.c_str()) == 0) {
|
|
470
|
+
GGML_LOG_INFO("ggml-openvino: model cache WROTE %s\n", blob_path.c_str());
|
|
471
|
+
} else {
|
|
472
|
+
std::remove(blob_tmp.c_str());
|
|
473
|
+
std::remove(manifest_tmp.c_str());
|
|
474
|
+
}
|
|
475
|
+
} else {
|
|
476
|
+
std::remove(blob_tmp.c_str());
|
|
477
|
+
std::remove(manifest_tmp.c_str());
|
|
478
|
+
}
|
|
479
|
+
} else {
|
|
480
|
+
std::remove(manifest_tmp.c_str());
|
|
481
|
+
}
|
|
482
|
+
}
|
|
483
|
+
} catch (const std::exception & e) {
|
|
484
|
+
GGML_LOG_WARN("ggml-openvino: model cache export failed: %s\n", e.what());
|
|
485
|
+
}
|
|
486
|
+
}
|
|
487
|
+
|
|
488
|
+
infer_request = std::make_shared<ov::InferRequest>(compiled_model.create_infer_request());
|
|
489
|
+
entry->ptr = ggml_decoder;
|
|
490
|
+
|
|
491
|
+
for (const auto & ov_param : model->get_parameters()) {
|
|
492
|
+
ov_input_names.push_back(ov_param->get_friendly_name());
|
|
493
|
+
}
|
|
494
|
+
for (const auto & ov_output : model->get_results()) {
|
|
495
|
+
ov_output_names.push_back(ov_output->get_friendly_name());
|
|
496
|
+
}
|
|
497
|
+
} // end non-imported (compile) path
|
|
498
|
+
|
|
499
|
+
if (cache_enabled) {
|
|
500
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
501
|
+
r_ctx->infer_request_cache[key] = infer_request;
|
|
502
|
+
r_ctx->ov_input_names_cache[key] = ov_input_names;
|
|
503
|
+
r_ctx->ov_output_names_cache[key] = ov_output_names;
|
|
504
|
+
}
|
|
505
|
+
|
|
506
|
+
if (stateful && cache_enabled) {
|
|
507
|
+
const auto * inp_pos = get_inp_pos_tensor(cgraph);
|
|
508
|
+
auto pos_shape = ggml_decoder->get_shape(inp_pos);
|
|
509
|
+
r_ctx->stateful_kv_size = pos_shape[3];
|
|
510
|
+
const auto kv_param_res_names = ggml_decoder->get_kv_param_res_names();
|
|
511
|
+
for (const auto & pair : kv_param_res_names) {
|
|
512
|
+
r_ctx->kv_state_input_name_map[pair.first + pair.second] = pair.first;
|
|
513
|
+
}
|
|
514
|
+
}
|
|
515
|
+
}
|
|
516
|
+
|
|
517
|
+
for (size_t i = 0; i < ov_input_names.size(); i++) {
|
|
518
|
+
auto param_name = ov_input_names[i];
|
|
519
|
+
auto input_tensor = get_ov_input_tensor(ggml_decoder, param_name);
|
|
520
|
+
infer_request->set_input_tensor(i, input_tensor);
|
|
521
|
+
|
|
522
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_INPUT")) {
|
|
523
|
+
print_input_tensor_info(param_name, input_tensor);
|
|
524
|
+
}
|
|
525
|
+
}
|
|
526
|
+
|
|
527
|
+
for (size_t i = 0; i < ov_output_names.size(); i++) {
|
|
528
|
+
// Debug-only outputs added via GGML_OPENVINO_DEBUG_NODE (see
|
|
529
|
+
// translate_session.cpp) have no corresponding ggml tensor; leave
|
|
530
|
+
// them unbound so OpenVINO allocates its own tensor for them,
|
|
531
|
+
// rather than aliasing a ggml buffer that may be overwritten by a
|
|
532
|
+
// later in-place op before we get to read it.
|
|
533
|
+
const auto & model_outputs = ggml_decoder->get_model_outputs();
|
|
534
|
+
auto model_output_it = model_outputs.find(ov_output_names[i]);
|
|
535
|
+
if (model_output_it == model_outputs.end()) {
|
|
536
|
+
continue;
|
|
537
|
+
}
|
|
538
|
+
auto * ggml_tensor = model_output_it->second;
|
|
539
|
+
if (ggml_nbytes(ggml_tensor) == 0) {
|
|
540
|
+
continue;
|
|
541
|
+
}
|
|
542
|
+
auto output_tensor = create_ov_output_tensor(ggml_decoder, infer_request, i, ggml_tensor);
|
|
543
|
+
infer_request->set_output_tensor(i, output_tensor);
|
|
544
|
+
}
|
|
545
|
+
|
|
546
|
+
ov_raw_infer_start = ggml_time_us();
|
|
547
|
+
infer_request->infer();
|
|
548
|
+
infer_end_time = ggml_time_us();
|
|
549
|
+
|
|
550
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_OUTPUT") ||
|
|
551
|
+
ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
|
|
552
|
+
for (size_t i = 0; i < ov_output_names.size(); i++) {
|
|
553
|
+
const auto output_tensor = infer_request->get_output_tensor(i);
|
|
554
|
+
print_output_tensor_info(ov_output_names[i], output_tensor, output_tensor.data());
|
|
555
|
+
}
|
|
556
|
+
}
|
|
557
|
+
|
|
558
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_PROFILING")) {
|
|
559
|
+
GGML_LOG_INFO("\nGGML OpenVINO Backend: \n");
|
|
560
|
+
GGML_LOG_INFO(" - Graph decoder time: %.3f ms \n", (decoder_end_time - start_time) / 1000.0);
|
|
561
|
+
if (!cache_hit) {
|
|
562
|
+
GGML_LOG_INFO(" - Graph conversion time: %.3f ms \n",
|
|
563
|
+
(conversion_end_time - decoder_end_time) / 1000.0);
|
|
564
|
+
GGML_LOG_INFO(" - Graph compile time: %.3f ms \n", (compile_end_time - conversion_end_time) / 1000.0);
|
|
565
|
+
}
|
|
566
|
+
GGML_LOG_INFO(" - Graph inference time: %.3f ms \n", (infer_end_time - compile_end_time) / 1000.0);
|
|
567
|
+
GGML_LOG_INFO(" - OV raw infer time: %.3f ms \n", (infer_end_time - ov_raw_infer_start) / 1000.0);
|
|
568
|
+
}
|
|
569
|
+
}
|
|
570
|
+
|
|
571
|
+
// GGML_OPENVINO_RELEASE_WEIGHTS (or GGML_OPENVINO_MEMORY_OPTIMIZE on GPU): the plugin holds its own device copy of
|
|
572
|
+
// every weight after compile, so the host weight buffers can be dropped to reclaim
|
|
573
|
+
// RSS. The GPU backend uses a single dynamic-shape model for both prefill and decode,
|
|
574
|
+
// so once a graph is compiled it is reused for the whole session — the only thing
|
|
575
|
+
// that forces a recompile is clear_caches() on backend teardown. We therefore release
|
|
576
|
+
// on the first cache-hit (model compiled, plugin has its copy) and, crucially, pin the
|
|
577
|
+
// compiled-model cache so it survives backend teardown (see ggml_backend_openvino_free).
|
|
578
|
+
// Without the pin, a later test/context would recompile against the now-dropped pages.
|
|
579
|
+
// A genuinely new graph still fails fast at the cache-miss compile branch.
|
|
580
|
+
if (cache_hit && ggml_openvino_release_weights_enabled(device) &&
|
|
581
|
+
!ggml_openvino_weight_buffers_released()) {
|
|
582
|
+
ggml_openvino_release_weight_buffers();
|
|
583
|
+
}
|
|
584
|
+
|
|
585
|
+
return GGML_STATUS_SUCCESS;
|
|
586
|
+
}
|
|
587
|
+
|
|
588
|
+
enum ggml_status ov_graph_compute_static(ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx) {
|
|
589
|
+
auto & core = ov_singleton_core();
|
|
590
|
+
|
|
591
|
+
auto get_prefill_chunk_size = [] {
|
|
592
|
+
static const int chunk_size = []() {
|
|
593
|
+
int env_prefill_chunk_size = ggml_openvino_getenv_int("GGML_OPENVINO_PREFILL_CHUNK_SIZE");
|
|
594
|
+
return env_prefill_chunk_size > 0 ? env_prefill_chunk_size : 256;
|
|
595
|
+
}();
|
|
596
|
+
return chunk_size;
|
|
597
|
+
};
|
|
598
|
+
|
|
599
|
+
// Normally NPU, but honors GGML_OPENVINO_DEVICE so GGML_OPENVINO_FORCE_STATIC can run the
|
|
600
|
+
// static-shape path on CPU/GPU to isolate translation bugs from NPUW/NPU-driver issues.
|
|
601
|
+
static std::string device = ggml_openvino_get_device_name();
|
|
602
|
+
static auto is_static = true;
|
|
603
|
+
static auto stateful = false;
|
|
604
|
+
|
|
605
|
+
auto prefill_chunk_size = get_prefill_chunk_size();
|
|
606
|
+
const auto & config = ggml_openvino_get_compile_config();
|
|
607
|
+
|
|
608
|
+
if (is_naive(cgraph)) {
|
|
609
|
+
return naive_compute(cgraph, core, device, config);
|
|
610
|
+
}
|
|
611
|
+
|
|
612
|
+
auto start_time = ggml_time_us();
|
|
613
|
+
|
|
614
|
+
std::shared_ptr<GgmlOvDecoder> ggml_decoder;
|
|
615
|
+
std::shared_ptr<ov::InferRequest> infer_request;
|
|
616
|
+
ModelParams m_params;
|
|
617
|
+
ComputeParams c_params;
|
|
618
|
+
std::tie(m_params, c_params) = GgmlOvDecoder::compute_llm_params(cgraph, is_static);
|
|
619
|
+
|
|
620
|
+
const auto * inp_pos = get_inp_pos_tensor(cgraph);
|
|
621
|
+
const auto is_prefill = get_is_prefill(cgraph, inp_pos);
|
|
622
|
+
graph_key key(cgraph);
|
|
623
|
+
static const bool cache_enabled = !ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_CACHE");
|
|
624
|
+
bool cache_hit = false;
|
|
625
|
+
|
|
626
|
+
int64_t decoder_end_time;
|
|
627
|
+
int64_t conversion_end_time;
|
|
628
|
+
int64_t compile_end_time;
|
|
629
|
+
int64_t infer_end_time;
|
|
630
|
+
int64_t ov_raw_infer_start;
|
|
631
|
+
int64_t ov_raw_infer_total = 0;
|
|
632
|
+
|
|
633
|
+
std::shared_ptr<decoder_runtime_ctx> entry;
|
|
634
|
+
ModelParams old_m_params;
|
|
635
|
+
|
|
636
|
+
if (cache_enabled) {
|
|
637
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
638
|
+
auto it = r_ctx->decoder_cache.find(key);
|
|
639
|
+
cache_hit = it != r_ctx->decoder_cache.end();
|
|
640
|
+
if (cache_hit) {
|
|
641
|
+
entry = it->second;
|
|
642
|
+
} else {
|
|
643
|
+
r_ctx->clear_caches_locked();
|
|
644
|
+
auto mutex = std::make_shared<std::mutex>();
|
|
645
|
+
entry = std::make_shared<decoder_runtime_ctx>(mutex);
|
|
646
|
+
r_ctx->decoder_cache[key] = entry;
|
|
647
|
+
}
|
|
648
|
+
} else {
|
|
649
|
+
auto mutex = std::make_shared<std::mutex>();
|
|
650
|
+
entry = std::make_shared<decoder_runtime_ctx>(mutex);
|
|
651
|
+
cache_hit = false;
|
|
652
|
+
}
|
|
653
|
+
|
|
654
|
+
std::lock_guard<std::mutex> lock(*(entry->mutex));
|
|
655
|
+
|
|
656
|
+
if (cache_hit) {
|
|
657
|
+
ggml_decoder = entry->ptr;
|
|
658
|
+
old_m_params = ggml_decoder->get_model_params();
|
|
659
|
+
cache_hit = old_m_params.can_reuse_statically(m_params);
|
|
660
|
+
}
|
|
661
|
+
|
|
662
|
+
std::vector<std::string> ov_input_names_local;
|
|
663
|
+
std::vector<std::string> ov_output_names_local;
|
|
664
|
+
|
|
665
|
+
if (cache_hit) {
|
|
666
|
+
std::map<std::string, std::shared_ptr<ov::Node>> model_weights;
|
|
667
|
+
ggml_decoder->m_is_prefill = is_prefill;
|
|
668
|
+
ggml_decoder->set_model_params(m_params);
|
|
669
|
+
ggml_decoder->set_compute_params(c_params);
|
|
670
|
+
if (old_m_params.kv_buffer_changed(m_params)) {
|
|
671
|
+
ggml_decoder->update_io(cgraph);
|
|
672
|
+
}
|
|
673
|
+
ggml_decoder->add_extra_inputs();
|
|
674
|
+
{
|
|
675
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
676
|
+
infer_request =
|
|
677
|
+
is_prefill ? r_ctx->infer_request_cache_prefill.at(key) : r_ctx->infer_request_cache.at(key);
|
|
678
|
+
ov_input_names_local = r_ctx->ov_input_names_cache.at(key);
|
|
679
|
+
ov_output_names_local = r_ctx->ov_output_names_cache.at(key);
|
|
680
|
+
}
|
|
681
|
+
|
|
682
|
+
decoder_end_time = ggml_time_us();
|
|
683
|
+
conversion_end_time = decoder_end_time;
|
|
684
|
+
compile_end_time = decoder_end_time;
|
|
685
|
+
} else {
|
|
686
|
+
if (cache_enabled) {
|
|
687
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
688
|
+
r_ctx->infer_request_cache.erase(key);
|
|
689
|
+
r_ctx->infer_request_cache_prefill.erase(key);
|
|
690
|
+
}
|
|
691
|
+
|
|
692
|
+
std::shared_ptr<ov::Model> model;
|
|
693
|
+
auto model_weights = GgmlOvDecoder::create_weight_nodes(cgraph);
|
|
694
|
+
|
|
695
|
+
if (m_params.n_heads_kv == -1) {
|
|
696
|
+
// graph is not a LLM, e.g. context-shift graph
|
|
697
|
+
prefill_chunk_size = inp_pos->ne[0];
|
|
698
|
+
}
|
|
699
|
+
auto ggml_decoder_prefill = std::make_shared<GgmlOvDecoder>(
|
|
700
|
+
cgraph, m_params, c_params, model_weights, is_static, stateful, false, true, prefill_chunk_size);
|
|
701
|
+
auto ggml_decoder_decode = std::make_shared<GgmlOvDecoder>(cgraph, m_params, c_params, model_weights, is_static,
|
|
702
|
+
stateful, false, false, prefill_chunk_size);
|
|
703
|
+
decoder_end_time = ggml_time_us();
|
|
704
|
+
|
|
705
|
+
const bool dump_ir = ggml_openvino_getenv_int("GGML_OPENVINO_DUMP_IR");
|
|
706
|
+
const auto dump_ir_timestamp = static_cast<long long>(ggml_time_us());
|
|
707
|
+
|
|
708
|
+
auto build_static_model = [&core, &config, dump_ir, dump_ir_timestamp](
|
|
709
|
+
std::shared_ptr<GgmlOvDecoder> decoder,
|
|
710
|
+
const char * tag,
|
|
711
|
+
std::shared_ptr<ov::Model> & model,
|
|
712
|
+
ov::CompiledModel & compiled_model,
|
|
713
|
+
std::shared_ptr<ov::InferRequest> & infer_request,
|
|
714
|
+
int64_t & local_conversion_end_time,
|
|
715
|
+
int64_t & local_compile_end_time) {
|
|
716
|
+
auto input_model = std::make_shared<ov::frontend::ggml::InputModel>(decoder);
|
|
717
|
+
model = ov::frontend::ggml::FrontEnd::convert(input_model);
|
|
718
|
+
decoder->clear_model_weights();
|
|
719
|
+
local_conversion_end_time = ggml_time_us();
|
|
720
|
+
|
|
721
|
+
if (dump_ir) {
|
|
722
|
+
char timestamped_filename[64];
|
|
723
|
+
snprintf(timestamped_filename, sizeof(timestamped_filename), "model_%s_%lld.xml", tag,
|
|
724
|
+
dump_ir_timestamp);
|
|
725
|
+
ov::serialize(model, timestamped_filename);
|
|
726
|
+
}
|
|
727
|
+
|
|
728
|
+
compiled_model = core.compile_model(model, device, config);
|
|
729
|
+
infer_request = std::make_shared<ov::InferRequest>(compiled_model.create_infer_request());
|
|
730
|
+
local_compile_end_time = ggml_time_us();
|
|
731
|
+
};
|
|
732
|
+
std::shared_ptr<ov::Model> model_prefill;
|
|
733
|
+
std::shared_ptr<ov::Model> model_decode;
|
|
734
|
+
ov::CompiledModel compiled_model_prefill;
|
|
735
|
+
ov::CompiledModel compiled_model_decode;
|
|
736
|
+
std::shared_ptr<ov::InferRequest> infer_request_prefill;
|
|
737
|
+
std::shared_ptr<ov::InferRequest> infer_request_decode;
|
|
738
|
+
int64_t prefill_conversion_end_time;
|
|
739
|
+
int64_t decode_conversion_end_time;
|
|
740
|
+
int64_t prefill_compile_end_time;
|
|
741
|
+
int64_t decode_compile_end_time;
|
|
742
|
+
auto prefill_future = std::async(std::launch::async, build_static_model, ggml_decoder_prefill, "prefill",
|
|
743
|
+
std::ref(model_prefill), std::ref(compiled_model_prefill),
|
|
744
|
+
std::ref(infer_request_prefill), std::ref(prefill_conversion_end_time),
|
|
745
|
+
std::ref(prefill_compile_end_time));
|
|
746
|
+
auto decode_future = std::async(std::launch::async, build_static_model, ggml_decoder_decode, "decode",
|
|
747
|
+
std::ref(model_decode), std::ref(compiled_model_decode),
|
|
748
|
+
std::ref(infer_request_decode), std::ref(decode_conversion_end_time),
|
|
749
|
+
std::ref(decode_compile_end_time));
|
|
750
|
+
prefill_future.get();
|
|
751
|
+
decode_future.get();
|
|
752
|
+
conversion_end_time = std::max(prefill_conversion_end_time, decode_conversion_end_time);
|
|
753
|
+
compile_end_time = std::max(prefill_compile_end_time, decode_compile_end_time);
|
|
754
|
+
|
|
755
|
+
model = is_prefill ? model_prefill : model_decode;
|
|
756
|
+
ggml_decoder = is_prefill ? ggml_decoder_prefill : ggml_decoder_decode;
|
|
757
|
+
infer_request = is_prefill ? infer_request_prefill : infer_request_decode;
|
|
758
|
+
entry->ptr = ggml_decoder;
|
|
759
|
+
|
|
760
|
+
for (const auto & ov_param : model->get_parameters()) {
|
|
761
|
+
ov_input_names_local.push_back(ov_param->get_friendly_name());
|
|
762
|
+
}
|
|
763
|
+
for (const auto & ov_output : model->get_results()) {
|
|
764
|
+
ov_output_names_local.push_back(ov_output->get_friendly_name());
|
|
765
|
+
}
|
|
766
|
+
|
|
767
|
+
if (cache_enabled) {
|
|
768
|
+
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
|
|
769
|
+
r_ctx->infer_request_cache_prefill[key] = infer_request_prefill;
|
|
770
|
+
r_ctx->infer_request_cache[key] = infer_request_decode;
|
|
771
|
+
r_ctx->ov_input_names_cache[key] = ov_input_names_local;
|
|
772
|
+
r_ctx->ov_output_names_cache[key] = ov_output_names_local;
|
|
773
|
+
}
|
|
774
|
+
}
|
|
775
|
+
|
|
776
|
+
if (is_prefill) {
|
|
777
|
+
auto inp_len = get_inp_pos_n_tokens(cgraph, inp_pos);
|
|
778
|
+
for (int chunk_index = 0; chunk_index * prefill_chunk_size < inp_len; chunk_index++) {
|
|
779
|
+
for (size_t i = 0; i < ov_input_names_local.size(); i++) {
|
|
780
|
+
auto param_name = ov_input_names_local[i];
|
|
781
|
+
auto input_tensor = get_ov_input_tensor_static_prefill(ggml_decoder, param_name, chunk_index);
|
|
782
|
+
infer_request->set_input_tensor(i, input_tensor);
|
|
783
|
+
|
|
784
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_INPUT")) {
|
|
785
|
+
const auto input_tensor = infer_request->get_input_tensor(i);
|
|
786
|
+
print_input_tensor_info(param_name, input_tensor);
|
|
787
|
+
}
|
|
788
|
+
}
|
|
789
|
+
|
|
790
|
+
for (size_t i = 0; i < ov_output_names_local.size(); i++) {
|
|
791
|
+
const auto & model_outputs = ggml_decoder->get_model_outputs();
|
|
792
|
+
auto model_output_it = model_outputs.find(ov_output_names_local[i]);
|
|
793
|
+
if (model_output_it == model_outputs.end()) {
|
|
794
|
+
continue;
|
|
795
|
+
}
|
|
796
|
+
auto * ggml_tensor = model_output_it->second;
|
|
797
|
+
if (ggml_nbytes(ggml_tensor) == 0) {
|
|
798
|
+
// Zero-row in-place writeback (e.g. the empty s_copy defrag remainder). The OV
|
|
799
|
+
// Result is the full cache, so binding it over this 0-byte buffer overflows it.
|
|
800
|
+
continue;
|
|
801
|
+
}
|
|
802
|
+
auto output_tensor = create_ov_output_tensor(ggml_decoder, infer_request, i, ggml_tensor);
|
|
803
|
+
infer_request->set_output_tensor(i, output_tensor);
|
|
804
|
+
}
|
|
805
|
+
|
|
806
|
+
ov_raw_infer_start = ggml_time_us();
|
|
807
|
+
infer_request->infer();
|
|
808
|
+
ov_raw_infer_total += ggml_time_us() - ov_raw_infer_start;
|
|
809
|
+
|
|
810
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_OUTPUT") ||
|
|
811
|
+
ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
|
|
812
|
+
for (size_t i = 0; i < ov_output_names_local.size(); i++) {
|
|
813
|
+
const auto output_tensor = infer_request->get_output_tensor(i);
|
|
814
|
+
print_output_tensor_info(ov_output_names_local[i], output_tensor, output_tensor.data());
|
|
815
|
+
}
|
|
816
|
+
}
|
|
817
|
+
}
|
|
818
|
+
infer_end_time = ggml_time_us();
|
|
819
|
+
} else {
|
|
820
|
+
for (size_t i = 0; i < ov_input_names_local.size(); i++) {
|
|
821
|
+
auto param_name = ov_input_names_local[i];
|
|
822
|
+
auto input_tensor = get_ov_input_tensor_static_decode(ggml_decoder, param_name);
|
|
823
|
+
infer_request->set_input_tensor(i, input_tensor);
|
|
824
|
+
|
|
825
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_INPUT")) {
|
|
826
|
+
const auto input_tensor = infer_request->get_input_tensor(i);
|
|
827
|
+
print_input_tensor_info(param_name, input_tensor);
|
|
828
|
+
}
|
|
829
|
+
}
|
|
830
|
+
|
|
831
|
+
for (size_t i = 0; i < ov_output_names_local.size(); i++) {
|
|
832
|
+
const auto & model_outputs = ggml_decoder->get_model_outputs();
|
|
833
|
+
auto model_output_it = model_outputs.find(ov_output_names_local[i]);
|
|
834
|
+
if (model_output_it == model_outputs.end()) {
|
|
835
|
+
continue;
|
|
836
|
+
}
|
|
837
|
+
auto * ggml_tensor = model_output_it->second;
|
|
838
|
+
if (ggml_nbytes(ggml_tensor) == 0) {
|
|
839
|
+
continue;
|
|
840
|
+
}
|
|
841
|
+
auto output_tensor = create_ov_output_tensor(ggml_decoder, infer_request, i, ggml_tensor);
|
|
842
|
+
infer_request->set_output_tensor(i, output_tensor);
|
|
843
|
+
}
|
|
844
|
+
|
|
845
|
+
ov_raw_infer_start = ggml_time_us();
|
|
846
|
+
infer_request->infer();
|
|
847
|
+
infer_end_time = ggml_time_us();
|
|
848
|
+
ov_raw_infer_total = infer_end_time - ov_raw_infer_start;
|
|
849
|
+
|
|
850
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_OUTPUT") ||
|
|
851
|
+
ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
|
|
852
|
+
for (size_t i = 0; i < ov_output_names_local.size(); i++) {
|
|
853
|
+
const auto output_tensor = infer_request->get_output_tensor(i);
|
|
854
|
+
print_output_tensor_info(ov_output_names_local[i], output_tensor, output_tensor.data());
|
|
855
|
+
}
|
|
856
|
+
}
|
|
857
|
+
}
|
|
858
|
+
|
|
859
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_PROFILING")) {
|
|
860
|
+
GGML_LOG_INFO("\nGGML OpenVINO Backend: \n");
|
|
861
|
+
GGML_LOG_INFO(" - Graph decoder time: %.3f ms \n", (decoder_end_time - start_time) / 1000.0);
|
|
862
|
+
if (!cache_hit) {
|
|
863
|
+
GGML_LOG_INFO(" - Graph conversion time: %.3f ms \n", (conversion_end_time - decoder_end_time) / 1000.0);
|
|
864
|
+
GGML_LOG_INFO(" - Graph compile time: %.3f ms \n", (compile_end_time - conversion_end_time) / 1000.0);
|
|
865
|
+
}
|
|
866
|
+
GGML_LOG_INFO(" - Graph inference time: %.3f ms \n", (infer_end_time - compile_end_time) / 1000.0);
|
|
867
|
+
GGML_LOG_INFO(" - OV raw infer time: %.3f ms \n", ov_raw_infer_total / 1000.0);
|
|
868
|
+
}
|
|
869
|
+
|
|
870
|
+
return GGML_STATUS_SUCCESS;
|
|
871
|
+
}
|
|
872
|
+
|
|
873
|
+
// Detect whether a cgraph is a split subgraph or not.
|
|
874
|
+
// Step 1 compares each node's recorded use_count with actual fan-out references in node->src.
|
|
875
|
+
// Step 2 verifies that node inputs come from model nodes/weights/leafs; external sources imply split.
|
|
876
|
+
bool is_model_splitted(ggml_cgraph * cgraph) {
|
|
877
|
+
static const bool fallback_enabled = ggml_openvino_getenv_int("GGML_OPENVINO_ENABLE_FALLBACK") != 0;
|
|
878
|
+
if (!fallback_enabled) {
|
|
879
|
+
return false;
|
|
880
|
+
}
|
|
881
|
+
|
|
882
|
+
// Backend op tests execute each node through ggml_graph_view(), which preserves the original
|
|
883
|
+
// graph use_counts while exposing only one node. Treat those single-node views as regular
|
|
884
|
+
// naive graphs so intermediate ops do not look like split-model fragments.
|
|
885
|
+
if (cgraph->n_nodes <= 1 && cgraph->n_leafs == 0) {
|
|
886
|
+
return false;
|
|
887
|
+
}
|
|
888
|
+
|
|
889
|
+
// check the nodes of the model are used by the following nodes, through compare the node's use count and the count of nodes that use it as input. If does not match, return true, else return false.
|
|
890
|
+
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
891
|
+
ggml_tensor * node = cgraph->nodes[i];
|
|
892
|
+
int use_count = cgraph->use_counts[ggml_hash_find(&cgraph->visited_hash_set, node)];
|
|
893
|
+
// TODO: this is a workround for the tests case from llama.cpp, fix should from the root cause in the future.
|
|
894
|
+
if ((cgraph->n_nodes <= 1 && use_count == 0) ||
|
|
895
|
+
(cgraph->n_nodes <= 1 && node->op == GGML_OP_VIEW && use_count == 1 && node->src[0] != nullptr &&
|
|
896
|
+
node->src[0]->op == GGML_OP_NONE)) {
|
|
897
|
+
return false;
|
|
898
|
+
}
|
|
899
|
+
if (cgraph->n_nodes == 1 &&
|
|
900
|
+
(cgraph->nodes[0]->op == GGML_OP_TRANSPOSE || cgraph->nodes[0]->op == GGML_OP_PERMUTE)) {
|
|
901
|
+
return false;
|
|
902
|
+
}
|
|
903
|
+
int input_use_count = 0;
|
|
904
|
+
for (int j = 0; j < cgraph->n_nodes; j++) {
|
|
905
|
+
ggml_tensor * other_node = cgraph->nodes[j];
|
|
906
|
+
for (int k = 0; k < GGML_MAX_SRC; k++) {
|
|
907
|
+
if (other_node->src[k] == node) {
|
|
908
|
+
input_use_count++;
|
|
909
|
+
}
|
|
910
|
+
}
|
|
911
|
+
}
|
|
912
|
+
if (use_count != input_use_count && node->op != GGML_OP_NONE) {
|
|
913
|
+
return true;
|
|
914
|
+
}
|
|
915
|
+
}
|
|
916
|
+
// if all nodes's src node's src is not come from the nodes in the model, we think the model is splitted. This is a complementary check for the above check, because for some special case like the output node is not used by any node, the use count and input use count are both 0, we can not determine whether the model is splitted or not just based on the first check.
|
|
917
|
+
// Only weight-name membership is needed below. With GGML_OPENVINO_REDUCE_COMPILE_MEM
|
|
918
|
+
// use the name-only collector (no weight extraction); otherwise keep the original
|
|
919
|
+
// behavior of building (naive) weight nodes and take their names.
|
|
920
|
+
std::set<std::string> model_weights;
|
|
921
|
+
if (ggml_openvino_reduce_compile_mem_enabled()) {
|
|
922
|
+
model_weights = GgmlOvDecoder::collect_weight_names(cgraph);
|
|
923
|
+
} else {
|
|
924
|
+
for (const auto & kv : GgmlOvDecoder::create_weight_nodes(cgraph, true)) {
|
|
925
|
+
model_weights.insert(kv.first);
|
|
926
|
+
}
|
|
927
|
+
}
|
|
928
|
+
std::set<ggml_tensor *> model_nodes(cgraph->nodes, cgraph->nodes + cgraph->n_nodes);
|
|
929
|
+
// leaf nodes
|
|
930
|
+
std::set<ggml_tensor *> model_leafs(cgraph->leafs, cgraph->leafs + cgraph->n_leafs);
|
|
931
|
+
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
932
|
+
ggml_tensor * node = cgraph->nodes[i];
|
|
933
|
+
for (int j = 0; j < GGML_MAX_SRC; j++) {
|
|
934
|
+
ggml_tensor * src = node->src[j];
|
|
935
|
+
// the src is also not the model weights, we think the model is splitted.
|
|
936
|
+
// the src is also not in model leafs, we think the model is splitted.
|
|
937
|
+
if (src != nullptr && model_nodes.find(src) == model_nodes.end() &&
|
|
938
|
+
model_weights.find(std::string(src->name)) == model_weights.end() && !model_leafs.empty() == false &&
|
|
939
|
+
model_leafs.find(src) == model_leafs.end()) {
|
|
940
|
+
if (GgmlOvDecoder::is_inp_tok(src, node)) {
|
|
941
|
+
return false;
|
|
942
|
+
}
|
|
943
|
+
return true;
|
|
944
|
+
}
|
|
945
|
+
}
|
|
946
|
+
}
|
|
947
|
+
return false;
|
|
948
|
+
}
|
|
949
|
+
|
|
950
|
+
bool is_naive(ggml_cgraph * cgraph) {
|
|
951
|
+
constexpr int naive_graph_size_threshold = 20;
|
|
952
|
+
int count = 0;
|
|
953
|
+
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
954
|
+
if (cgraph->nodes[i]->op != GGML_OP_NONE) {
|
|
955
|
+
count++;
|
|
956
|
+
}
|
|
957
|
+
}
|
|
958
|
+
return count < naive_graph_size_threshold;
|
|
959
|
+
}
|
|
960
|
+
|
|
961
|
+
enum ggml_status naive_compute(ggml_cgraph * cgraph,
|
|
962
|
+
ov::Core & core,
|
|
963
|
+
const std::string & device,
|
|
964
|
+
const ov::AnyMap & config) {
|
|
965
|
+
if (cgraph->n_nodes == 1 && (cgraph->nodes[0]->op == GGML_OP_NONE || cgraph->nodes[0]->op == GGML_OP_VIEW)) {
|
|
966
|
+
return GGML_STATUS_SUCCESS;
|
|
967
|
+
}
|
|
968
|
+
|
|
969
|
+
bool naive = true;
|
|
970
|
+
auto model_weights = GgmlOvDecoder::create_weight_nodes(cgraph, naive);
|
|
971
|
+
auto decoder = std::make_shared<GgmlOvDecoder>(cgraph, model_weights);
|
|
972
|
+
auto input_model = std::make_shared<ov::frontend::ggml::InputModel>(decoder);
|
|
973
|
+
auto model = ov::frontend::ggml::FrontEnd::convert(input_model, naive);
|
|
974
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DUMP_IR")) {
|
|
975
|
+
ov::serialize(model, "IR_naive.xml");
|
|
976
|
+
}
|
|
977
|
+
|
|
978
|
+
std::shared_ptr<ov::InferRequest> infer_request;
|
|
979
|
+
auto remote_context = ggml_openvino_get_remote_context();
|
|
980
|
+
if (cgraph->nodes[0]->op == GGML_OP_MUL_MAT) {
|
|
981
|
+
// TODO ACCURACY hint triggers a bug in GPU plugin/driver on Lunar Lake. Remove once CVS-182166 is resolved
|
|
982
|
+
core.set_property(device, ov::hint::execution_mode(ov::hint::ExecutionMode::PERFORMANCE));
|
|
983
|
+
} else {
|
|
984
|
+
core.set_property(device, ov::hint::execution_mode(ov::hint::ExecutionMode::ACCURACY));
|
|
985
|
+
}
|
|
986
|
+
if (remote_context.has_value()) {
|
|
987
|
+
infer_request = std::make_shared<ov::InferRequest>(
|
|
988
|
+
core.compile_model(model, remote_context.value(), config).create_infer_request());
|
|
989
|
+
} else {
|
|
990
|
+
infer_request =
|
|
991
|
+
std::make_shared<ov::InferRequest>(core.compile_model(model, device, config).create_infer_request());
|
|
992
|
+
}
|
|
993
|
+
|
|
994
|
+
auto ov_params = model->get_parameters();
|
|
995
|
+
for (size_t i = 0; i < ov_params.size(); i++) {
|
|
996
|
+
auto param_name = ov_params[i]->get_friendly_name();
|
|
997
|
+
auto input_tensor = get_ov_input_tensor(decoder, param_name);
|
|
998
|
+
infer_request->set_input_tensor(i, input_tensor);
|
|
999
|
+
}
|
|
1000
|
+
|
|
1001
|
+
// Use get_output_tensor + memcpy instead of set_output_tensor to avoid memory overwritten
|
|
1002
|
+
// when i/o buffer overlaps, e.g. the cgraph is a single PERMUTE
|
|
1003
|
+
|
|
1004
|
+
infer_request->infer();
|
|
1005
|
+
|
|
1006
|
+
auto ov_results = model->get_results();
|
|
1007
|
+
for (size_t i = 0; i < ov_results.size(); i++) {
|
|
1008
|
+
auto output_tensor = infer_request->get_output_tensor(i);
|
|
1009
|
+
const auto & model_outputs = decoder->get_model_outputs();
|
|
1010
|
+
auto model_output_it = model_outputs.find(ov_results[i]->get_friendly_name());
|
|
1011
|
+
if (model_output_it == model_outputs.end()) {
|
|
1012
|
+
// Debug-only output added via GGML_OPENVINO_DEBUG_NODE; nothing to copy into.
|
|
1013
|
+
if (ggml_openvino_getenv_int("GGML_OPENVINO_DEBUG_OUTPUT") ||
|
|
1014
|
+
ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
|
|
1015
|
+
print_output_tensor_info(ov_results[i]->get_friendly_name(), output_tensor, output_tensor.data());
|
|
1016
|
+
}
|
|
1017
|
+
continue;
|
|
1018
|
+
}
|
|
1019
|
+
auto * ggml_tensor = model_output_it->second;
|
|
1020
|
+
std::memcpy(ggml_tensor->data, output_tensor.data(), output_tensor.get_byte_size());
|
|
1021
|
+
}
|
|
1022
|
+
return GGML_STATUS_SUCCESS;
|
|
1023
|
+
}
|
|
1024
|
+
|
|
1025
|
+
namespace {
|
|
1026
|
+
template <typename T> void set_zero_diagonal(std::vector<T> & matrix, size_t rows, size_t cols, T zero_value = T{}) {
|
|
1027
|
+
for (size_t i = 0; i < rows; ++i) {
|
|
1028
|
+
size_t diag_col = std::min(i, cols - 1);
|
|
1029
|
+
matrix[i * cols + diag_col] = zero_value;
|
|
1030
|
+
}
|
|
1031
|
+
}
|
|
1032
|
+
|
|
1033
|
+
ov::Tensor make_contiguous_split_input_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
|
|
1034
|
+
const struct ggml_tensor * ggml_tensor,
|
|
1035
|
+
const ov::Shape & input_shape) {
|
|
1036
|
+
const size_t element_size = ggml_type_size(ggml_tensor->type);
|
|
1037
|
+
const size_t block_size = ggml_blck_size(ggml_tensor->type);
|
|
1038
|
+
|
|
1039
|
+
GGML_ASSERT(block_size == 1 && "non-contiguous split inputs must be plain element types");
|
|
1040
|
+
|
|
1041
|
+
const struct ggml_tensor * source_tensor = ggml_tensor->view_src != nullptr ? ggml_tensor->view_src : ggml_tensor;
|
|
1042
|
+
const size_t source_offset = ggml_tensor->view_src != nullptr ? ggml_tensor->view_offs : 0;
|
|
1043
|
+
|
|
1044
|
+
std::vector<uint8_t> source_data(ggml_nbytes(source_tensor));
|
|
1045
|
+
ggml_backend_tensor_get(source_tensor, source_data.data(), 0, source_data.size());
|
|
1046
|
+
|
|
1047
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1048
|
+
auto * dst = static_cast<uint8_t *>(input_tensor.data());
|
|
1049
|
+
size_t dst_offset = 0;
|
|
1050
|
+
|
|
1051
|
+
for (size_t i3 = 0; i3 < static_cast<size_t>(ggml_tensor->ne[3]); ++i3) {
|
|
1052
|
+
for (size_t i2 = 0; i2 < static_cast<size_t>(ggml_tensor->ne[2]); ++i2) {
|
|
1053
|
+
for (size_t i1 = 0; i1 < static_cast<size_t>(ggml_tensor->ne[1]); ++i1) {
|
|
1054
|
+
for (size_t i0 = 0; i0 < static_cast<size_t>(ggml_tensor->ne[0]); ++i0) {
|
|
1055
|
+
const size_t src_offset = source_offset + i3 * ggml_tensor->nb[3] + i2 * ggml_tensor->nb[2] +
|
|
1056
|
+
i1 * ggml_tensor->nb[1] + i0 * ggml_tensor->nb[0];
|
|
1057
|
+
std::memcpy(dst + dst_offset, source_data.data() + src_offset, element_size);
|
|
1058
|
+
dst_offset += element_size;
|
|
1059
|
+
}
|
|
1060
|
+
}
|
|
1061
|
+
}
|
|
1062
|
+
}
|
|
1063
|
+
|
|
1064
|
+
return input_tensor;
|
|
1065
|
+
}
|
|
1066
|
+
|
|
1067
|
+
ov::Tensor convert_ggml_input_to_ov(std::shared_ptr<GgmlOvDecoder> ggml_decoder, const std::string & name) {
|
|
1068
|
+
const auto * ggml_tensor = ggml_decoder->get_input_ggml_tensor(name);
|
|
1069
|
+
|
|
1070
|
+
if (auto sliced = try_make_kv_sliced_tensor(ggml_decoder, name, ggml_tensor)) {
|
|
1071
|
+
return *sliced;
|
|
1072
|
+
}
|
|
1073
|
+
|
|
1074
|
+
if (ggml_tensor->extra != nullptr && !ggml_decoder->is_splited_model()) {
|
|
1075
|
+
auto * extra_base = static_cast<ggml_openvino_extra_base *>(ggml_tensor->extra);
|
|
1076
|
+
if (extra_base->type == ggml_openvino_extra_base::Type::TENSOR) {
|
|
1077
|
+
// GGML_LOG_DEBUG("Using ggml_tensor->extra as ov::Tensor for input: %s\n", name.c_str());
|
|
1078
|
+
auto * tensor_extra = static_cast<ggml_openvino_tensor_extra *>(extra_base);
|
|
1079
|
+
return *tensor_extra->tensor;
|
|
1080
|
+
}
|
|
1081
|
+
}
|
|
1082
|
+
|
|
1083
|
+
// GGML_LOG_DEBUG("Converting ggml tensor to ov::Tensor for input: %s\n", name.c_str());
|
|
1084
|
+
auto * input_data = ggml_tensor->data;
|
|
1085
|
+
ov::Shape input_shape;
|
|
1086
|
+
if (ggml_tensor->op == GGML_OP_VIEW && !ggml_decoder->is_splited_model()) {
|
|
1087
|
+
// This case is added to make test-backend-ops work
|
|
1088
|
+
input_shape = ggml_decoder->get_shape(ggml_tensor->view_src);
|
|
1089
|
+
} else {
|
|
1090
|
+
input_shape = ggml_decoder->get_shape(ggml_tensor);
|
|
1091
|
+
}
|
|
1092
|
+
|
|
1093
|
+
if (ggml_decoder->is_splited_model() && !ggml_is_contiguous(ggml_tensor)) {
|
|
1094
|
+
return make_contiguous_split_input_tensor(ggml_decoder, ggml_tensor, input_shape);
|
|
1095
|
+
}
|
|
1096
|
+
|
|
1097
|
+
auto input_tensor = ov::Tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape, input_data);
|
|
1098
|
+
return input_tensor;
|
|
1099
|
+
}
|
|
1100
|
+
} // namespace
|
|
1101
|
+
|
|
1102
|
+
ov::Tensor get_ov_input_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder, const std::string & param_name) {
|
|
1103
|
+
ov::Tensor input_tensor;
|
|
1104
|
+
auto extra_input = ggml_decoder->get_model_extra_inputs().find(param_name);
|
|
1105
|
+
if (extra_input != ggml_decoder->get_model_extra_inputs().end()) {
|
|
1106
|
+
input_tensor = ov::Tensor(extra_input->second.type, extra_input->second.shape);
|
|
1107
|
+
*input_tensor.data<int64_t>() = extra_input->second.value;
|
|
1108
|
+
} else {
|
|
1109
|
+
input_tensor = convert_ggml_input_to_ov(ggml_decoder, param_name);
|
|
1110
|
+
}
|
|
1111
|
+
return input_tensor;
|
|
1112
|
+
}
|
|
1113
|
+
|
|
1114
|
+
ov::Tensor get_ov_input_tensor_static_decode(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
|
|
1115
|
+
const std::string & param_name) {
|
|
1116
|
+
// NPU decoding stage
|
|
1117
|
+
if (ggml_decoder->get_model_extra_inputs().count(param_name)) {
|
|
1118
|
+
return get_ov_input_tensor(ggml_decoder, param_name);
|
|
1119
|
+
}
|
|
1120
|
+
const auto * ggml_tensor = ggml_decoder->get_input_ggml_tensor(param_name);
|
|
1121
|
+
const auto * op = ggml_decoder->get_tensor_used_op(ggml_tensor);
|
|
1122
|
+
|
|
1123
|
+
if (GgmlOvDecoder::is_inp_tok(ggml_tensor, op) || GgmlOvDecoder::is_inp_pos(ggml_tensor, op) ||
|
|
1124
|
+
GgmlOvDecoder::is_kv_idx(ggml_tensor, op)) {
|
|
1125
|
+
// IMROPE's inp_pos holds one value per t/h/w/e plane instead of a single position;
|
|
1126
|
+
// with a single decode token the planes are still contiguous, so a flat copy works.
|
|
1127
|
+
const int n_planes = GgmlOvDecoder::is_inp_pos(ggml_tensor, op) ? GgmlOvDecoder::get_inp_pos_n_planes(op) : 1;
|
|
1128
|
+
assert(ggml_tensor->ne[0] == n_planes);
|
|
1129
|
+
ov::Shape input_shape = {1, 1, 1, (size_t) n_planes};
|
|
1130
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1131
|
+
std::memcpy(input_tensor.data(), ggml_tensor->data, n_planes * ggml_type_size(ggml_tensor->type));
|
|
1132
|
+
return input_tensor;
|
|
1133
|
+
}
|
|
1134
|
+
|
|
1135
|
+
if (GgmlOvDecoder::is_output_idx(ggml_tensor, op)) {
|
|
1136
|
+
ov::Shape input_shape = {1, 1, 1, 1};
|
|
1137
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1138
|
+
int32_t inp_out_id = *((int32_t *) ggml_tensor->data);
|
|
1139
|
+
assert(ggml_tensor->ne[0] == 1);
|
|
1140
|
+
assert(inp_out_id == 0);
|
|
1141
|
+
*input_tensor.data<int32_t>() = inp_out_id;
|
|
1142
|
+
return input_tensor;
|
|
1143
|
+
}
|
|
1144
|
+
|
|
1145
|
+
if (GgmlOvDecoder::is_inp_mask(ggml_tensor, op)) {
|
|
1146
|
+
size_t context_size = ggml_decoder->get_ctx_size();
|
|
1147
|
+
if (ggml_tensor->type == GGML_TYPE_F16) {
|
|
1148
|
+
std::vector<ggml_fp16_t> padded_data =
|
|
1149
|
+
pad_input<ggml_fp16_t>(ggml_tensor, 1, context_size, GGML_FP32_TO_FP16(-INFINITY));
|
|
1150
|
+
ov::Tensor input_tensor(ov::element::f16, ov::Shape{1, 1, 1, context_size});
|
|
1151
|
+
std::memcpy(input_tensor.data(), padded_data.data(), padded_data.size() * sizeof(ggml_fp16_t));
|
|
1152
|
+
return input_tensor;
|
|
1153
|
+
}
|
|
1154
|
+
|
|
1155
|
+
std::vector<float> padded_data = pad_input<float>(ggml_tensor, 1, context_size, -INFINITY);
|
|
1156
|
+
ov::Tensor input_tensor(ov::element::f32, ov::Shape{1, 1, 1, context_size});
|
|
1157
|
+
auto * data_ptr = input_tensor.data<float>();
|
|
1158
|
+
std::copy(padded_data.begin(), padded_data.begin() + context_size, data_ptr);
|
|
1159
|
+
return input_tensor;
|
|
1160
|
+
}
|
|
1161
|
+
|
|
1162
|
+
return get_ov_input_tensor(ggml_decoder, param_name);
|
|
1163
|
+
}
|
|
1164
|
+
|
|
1165
|
+
ov::Tensor get_ov_input_tensor_static_prefill(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
|
|
1166
|
+
const std::string & param_name,
|
|
1167
|
+
int chunk_index) {
|
|
1168
|
+
// NPU prompt processing stage
|
|
1169
|
+
const size_t input_len = ggml_decoder->get_input_len();
|
|
1170
|
+
const size_t chunk_size = ggml_decoder->m_prefill_chunk_size;
|
|
1171
|
+
const size_t chunk_valid_size = std::min(chunk_size, input_len - chunk_index * chunk_size);
|
|
1172
|
+
const size_t chunk_pad_size = chunk_size - chunk_valid_size;
|
|
1173
|
+
|
|
1174
|
+
if (param_name == "chunk_valid_len") {
|
|
1175
|
+
ov::Tensor input_tensor(ov::element::i64, ov::Shape{1});
|
|
1176
|
+
*input_tensor.data<int64_t>() = (int64_t) chunk_valid_size;
|
|
1177
|
+
return input_tensor;
|
|
1178
|
+
}
|
|
1179
|
+
if (chunk_index > 0 && param_name == "cache_rs_reset_len") {
|
|
1180
|
+
// The recurrent-state clear belongs to the start of the sequence. Re-applying it on every
|
|
1181
|
+
// chunk would wipe the state accumulated by the preceding chunks, so disable it (a zero
|
|
1182
|
+
// length makes scale.cpp's keep-mask select every slot) after the first chunk.
|
|
1183
|
+
ov::Tensor input_tensor(ov::element::i64, ov::Shape{1});
|
|
1184
|
+
*input_tensor.data<int64_t>() = 0;
|
|
1185
|
+
return input_tensor;
|
|
1186
|
+
}
|
|
1187
|
+
if (ggml_decoder->get_model_extra_inputs().count(param_name)) {
|
|
1188
|
+
return get_ov_input_tensor(ggml_decoder, param_name);
|
|
1189
|
+
}
|
|
1190
|
+
const auto * ggml_tensor = ggml_decoder->get_input_ggml_tensor(param_name);
|
|
1191
|
+
const auto * op = ggml_decoder->get_tensor_used_op(ggml_tensor);
|
|
1192
|
+
|
|
1193
|
+
if (GgmlOvDecoder::is_inp_pos(ggml_tensor, op) && GgmlOvDecoder::get_inp_pos_n_planes(op) > 1) {
|
|
1194
|
+
// IMROPE: inp_pos stacks n_planes (t/h/w/e) position planes, each of length
|
|
1195
|
+
// input_len; pad every plane independently so they stay aligned to chunk_size.
|
|
1196
|
+
const int n_planes = GgmlOvDecoder::get_inp_pos_n_planes(op);
|
|
1197
|
+
const size_t element_size = ggml_type_size(ggml_tensor->type);
|
|
1198
|
+
ov::Shape input_shape = {1, 1, 1, (size_t) n_planes * chunk_size};
|
|
1199
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1200
|
+
for (int p = 0; p < n_planes; p++) {
|
|
1201
|
+
const char * src =
|
|
1202
|
+
(const char *) ggml_tensor->data + (p * input_len + chunk_index * chunk_size) * element_size;
|
|
1203
|
+
char * dst = (char *) input_tensor.data() + p * chunk_size * element_size;
|
|
1204
|
+
std::memcpy(dst, src, chunk_valid_size * element_size);
|
|
1205
|
+
if (chunk_pad_size > 0) {
|
|
1206
|
+
if (ggml_tensor->type == GGML_TYPE_I32) {
|
|
1207
|
+
int32_t last_value = *((const int32_t *) src + chunk_valid_size - 1);
|
|
1208
|
+
int32_t * out = (int32_t *) dst;
|
|
1209
|
+
std::fill(out + chunk_valid_size, out + chunk_size, last_value + 1);
|
|
1210
|
+
} else if (ggml_tensor->type == GGML_TYPE_I64) {
|
|
1211
|
+
int64_t last_value = *((const int64_t *) src + chunk_valid_size - 1);
|
|
1212
|
+
int64_t * out = (int64_t *) dst;
|
|
1213
|
+
std::fill(out + chunk_valid_size, out + chunk_size, last_value + 1);
|
|
1214
|
+
} else {
|
|
1215
|
+
throw std::runtime_error("Unexpected tensor type for " + param_name);
|
|
1216
|
+
}
|
|
1217
|
+
}
|
|
1218
|
+
}
|
|
1219
|
+
return input_tensor;
|
|
1220
|
+
}
|
|
1221
|
+
|
|
1222
|
+
if (GgmlOvDecoder::is_inp_tok(ggml_tensor, op) || GgmlOvDecoder::is_inp_pos(ggml_tensor, op) ||
|
|
1223
|
+
GgmlOvDecoder::is_kv_idx(ggml_tensor, op)) {
|
|
1224
|
+
ov::Shape input_shape = {1, 1, 1, chunk_size};
|
|
1225
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1226
|
+
// copy the chunk_index-th chunk from ggml_tensor
|
|
1227
|
+
size_t element_size = ggml_type_size(ggml_tensor->type);
|
|
1228
|
+
void * input_data = (char *) ggml_tensor->data + chunk_index * chunk_size * element_size;
|
|
1229
|
+
std::memcpy(input_tensor.data(), input_data, chunk_valid_size * element_size);
|
|
1230
|
+
// pad the rest with last_value + 1, so that kv's of padded positions are inserted
|
|
1231
|
+
// to the next row after the valids row in the kvcache
|
|
1232
|
+
if (chunk_pad_size > 0) {
|
|
1233
|
+
if (ggml_tensor->type == GGML_TYPE_I32) {
|
|
1234
|
+
int32_t last_value =
|
|
1235
|
+
*((int32_t *) ggml_tensor->data + (chunk_index * chunk_size + chunk_valid_size - 1));
|
|
1236
|
+
int32_t * output_data = input_tensor.data<int32_t>();
|
|
1237
|
+
std::fill(output_data + chunk_valid_size, output_data + chunk_size, last_value + 1);
|
|
1238
|
+
} else if (ggml_tensor->type == GGML_TYPE_I64) {
|
|
1239
|
+
int64_t last_value =
|
|
1240
|
+
*((int64_t *) ggml_tensor->data + (chunk_index * chunk_size + chunk_valid_size - 1));
|
|
1241
|
+
int64_t * output_data = input_tensor.data<int64_t>();
|
|
1242
|
+
std::fill(output_data + chunk_valid_size, output_data + chunk_size, last_value + 1);
|
|
1243
|
+
} else {
|
|
1244
|
+
throw std::runtime_error("Unexpected tensor type for " + param_name);
|
|
1245
|
+
}
|
|
1246
|
+
}
|
|
1247
|
+
return input_tensor;
|
|
1248
|
+
}
|
|
1249
|
+
|
|
1250
|
+
if (GgmlOvDecoder::is_output_idx(ggml_tensor, op)) {
|
|
1251
|
+
size_t output_len = ggml_decoder->get_compute_params().output_len;
|
|
1252
|
+
ov::Shape input_shape = {1, 1, 1, output_len};
|
|
1253
|
+
ov::Tensor input_tensor(ggml_decoder->get_ov_type(ggml_tensor), input_shape);
|
|
1254
|
+
if (ggml_tensor->ne[0] == 0) {
|
|
1255
|
+
*input_tensor.data<int32_t>() = 0;
|
|
1256
|
+
} else {
|
|
1257
|
+
auto * data_addr = input_tensor.data<int32_t>();
|
|
1258
|
+
for (size_t i = 0; i < output_len; i++) {
|
|
1259
|
+
data_addr[i] = ((int32_t *) ggml_tensor->data)[i] % chunk_size;
|
|
1260
|
+
}
|
|
1261
|
+
}
|
|
1262
|
+
return input_tensor;
|
|
1263
|
+
}
|
|
1264
|
+
|
|
1265
|
+
if (GgmlOvDecoder::is_inp_mask(ggml_tensor, op)) {
|
|
1266
|
+
size_t cols = ggml_tensor->ne[0];
|
|
1267
|
+
size_t rows = ggml_tensor->ne[1];
|
|
1268
|
+
size_t chunk_valid_rows = std::min(chunk_size, rows - chunk_index * chunk_size);
|
|
1269
|
+
size_t context_size = ggml_decoder->get_ctx_size();
|
|
1270
|
+
if (ggml_tensor->type == GGML_TYPE_F16) {
|
|
1271
|
+
const auto * ggml_data =
|
|
1272
|
+
static_cast<const ggml_fp16_t *>(ggml_tensor->data) + chunk_index * chunk_size * cols;
|
|
1273
|
+
std::vector<ggml_fp16_t> padded_data = pad_input<ggml_fp16_t>(ggml_data, chunk_valid_rows, cols, chunk_size,
|
|
1274
|
+
context_size, GGML_FP32_TO_FP16(-INFINITY));
|
|
1275
|
+
set_zero_diagonal(padded_data, chunk_size, context_size, GGML_FP32_TO_FP16(0.0f));
|
|
1276
|
+
ov::Tensor input_tensor(ov::element::f16, ov::Shape{1, 1, chunk_size, context_size});
|
|
1277
|
+
std::memcpy(input_tensor.data(), padded_data.data(), padded_data.size() * sizeof(ggml_fp16_t));
|
|
1278
|
+
return input_tensor;
|
|
1279
|
+
}
|
|
1280
|
+
|
|
1281
|
+
const auto * ggml_data = static_cast<const float *>(ggml_tensor->data) + chunk_index * chunk_size * cols;
|
|
1282
|
+
std::vector<float> padded_data =
|
|
1283
|
+
pad_input<float>(ggml_data, chunk_valid_rows, cols, chunk_size, context_size, -INFINITY);
|
|
1284
|
+
set_zero_diagonal(padded_data, chunk_size, context_size);
|
|
1285
|
+
ov::Tensor input_tensor(ov::element::f32, ov::Shape{1, 1, chunk_size, context_size});
|
|
1286
|
+
auto * data_ptr = input_tensor.data<float>();
|
|
1287
|
+
std::copy(padded_data.begin(), padded_data.begin() + chunk_size * context_size, data_ptr);
|
|
1288
|
+
return input_tensor;
|
|
1289
|
+
}
|
|
1290
|
+
|
|
1291
|
+
return get_ov_input_tensor(ggml_decoder, param_name);
|
|
1292
|
+
}
|
|
1293
|
+
|
|
1294
|
+
size_t checksum(const void * data, size_t size) {
|
|
1295
|
+
const uint8_t * bytes = static_cast<const uint8_t *>(data);
|
|
1296
|
+
size_t sum = 0;
|
|
1297
|
+
for (size_t i = 0; i < size; ++i) {
|
|
1298
|
+
sum += (uint8_t) i;
|
|
1299
|
+
sum += bytes[i];
|
|
1300
|
+
}
|
|
1301
|
+
return sum;
|
|
1302
|
+
}
|
|
1303
|
+
|
|
1304
|
+
bool save_ggml_tensor_data_to_txt(const ggml_tensor * tensor, const std::string & file_path) {
|
|
1305
|
+
if (tensor == nullptr || tensor->data == nullptr) {
|
|
1306
|
+
return false;
|
|
1307
|
+
}
|
|
1308
|
+
|
|
1309
|
+
std::ofstream out(file_path);
|
|
1310
|
+
if (!out.is_open()) {
|
|
1311
|
+
return false;
|
|
1312
|
+
}
|
|
1313
|
+
|
|
1314
|
+
const size_t n = ggml_nelements(tensor);
|
|
1315
|
+
out << "name: " << tensor->name << ", type: " << ggml_type_name(tensor->type) << ", shape: [" << tensor->ne[0]
|
|
1316
|
+
<< ", " << tensor->ne[1] << ", " << tensor->ne[2] << ", " << tensor->ne[3] << "]" << ", elements: " << n
|
|
1317
|
+
<< ", data:" << '\n';
|
|
1318
|
+
|
|
1319
|
+
switch (tensor->type) {
|
|
1320
|
+
case GGML_TYPE_F32: {
|
|
1321
|
+
const auto * data = static_cast<const float *>(tensor->data);
|
|
1322
|
+
for (size_t i = 0; i < n; ++i) {
|
|
1323
|
+
out << data[i] << '\n';
|
|
1324
|
+
}
|
|
1325
|
+
break;
|
|
1326
|
+
}
|
|
1327
|
+
case GGML_TYPE_F16: {
|
|
1328
|
+
const auto * data = static_cast<const ggml_fp16_t *>(tensor->data);
|
|
1329
|
+
for (size_t i = 0; i < n; ++i) {
|
|
1330
|
+
out << ggml_fp16_to_fp32(data[i]) << '\n';
|
|
1331
|
+
}
|
|
1332
|
+
break;
|
|
1333
|
+
}
|
|
1334
|
+
case GGML_TYPE_BF16: {
|
|
1335
|
+
const auto * data = static_cast<const ggml_bf16_t *>(tensor->data);
|
|
1336
|
+
for (size_t i = 0; i < n; ++i) {
|
|
1337
|
+
out << ggml_bf16_to_fp32(data[i]) << '\n';
|
|
1338
|
+
}
|
|
1339
|
+
break;
|
|
1340
|
+
}
|
|
1341
|
+
case GGML_TYPE_I32: {
|
|
1342
|
+
const auto * data = static_cast<const int32_t *>(tensor->data);
|
|
1343
|
+
for (size_t i = 0; i < n; ++i) {
|
|
1344
|
+
out << data[i] << '\n';
|
|
1345
|
+
}
|
|
1346
|
+
break;
|
|
1347
|
+
}
|
|
1348
|
+
case GGML_TYPE_I64: {
|
|
1349
|
+
const auto * data = static_cast<const int64_t *>(tensor->data);
|
|
1350
|
+
for (size_t i = 0; i < n; ++i) {
|
|
1351
|
+
out << data[i] << '\n';
|
|
1352
|
+
}
|
|
1353
|
+
break;
|
|
1354
|
+
}
|
|
1355
|
+
default:
|
|
1356
|
+
out << "unsupported tensor type for text dump" << '\n';
|
|
1357
|
+
return false;
|
|
1358
|
+
}
|
|
1359
|
+
|
|
1360
|
+
return true;
|
|
1361
|
+
}
|
|
1362
|
+
|
|
1363
|
+
void print_input_tensor_info(const std::string & name, const ov::Tensor & tensor) {
|
|
1364
|
+
std::cout << "Input name: " << name << ", Input shape: " << tensor.get_shape() << ", Address: " << tensor.data()
|
|
1365
|
+
<< std::endl;
|
|
1366
|
+
switch (tensor.get_element_type()) {
|
|
1367
|
+
case ov::element::f32: {
|
|
1368
|
+
if (name.find("self_kq_mask") == std::string::npos && name.find("KQ_mask") == std::string::npos) {
|
|
1369
|
+
std::cout << *(tensor.data<float>()) << std::endl;
|
|
1370
|
+
} else {
|
|
1371
|
+
size_t rows = tensor.get_shape()[2];
|
|
1372
|
+
size_t cols = tensor.get_shape()[3];
|
|
1373
|
+
auto * data = tensor.data<float>();
|
|
1374
|
+
for (size_t i = 0; i < rows; ++i) {
|
|
1375
|
+
for (size_t j = 0; j < cols; ++j) {
|
|
1376
|
+
float val = data[i * cols + j];
|
|
1377
|
+
if (std::isinf(val) && val < 0) {
|
|
1378
|
+
std::cout << std::setw(5) << "-inf";
|
|
1379
|
+
} else {
|
|
1380
|
+
std::cout << std::setw(5) << val;
|
|
1381
|
+
}
|
|
1382
|
+
}
|
|
1383
|
+
std::cout << std::endl;
|
|
1384
|
+
}
|
|
1385
|
+
}
|
|
1386
|
+
|
|
1387
|
+
break;
|
|
1388
|
+
}
|
|
1389
|
+
case ov::element::f16:
|
|
1390
|
+
std::cout << *(tensor.data<ov::float16>()) << std::endl;
|
|
1391
|
+
break;
|
|
1392
|
+
case ov::element::i32:
|
|
1393
|
+
for (size_t i = 0; i < tensor.get_size(); ++i) {
|
|
1394
|
+
std::cout << tensor.data<int32_t>()[i] << " ";
|
|
1395
|
+
}
|
|
1396
|
+
std::cout << std::endl;
|
|
1397
|
+
break;
|
|
1398
|
+
case ov::element::i64:
|
|
1399
|
+
for (size_t i = 0; i < tensor.get_size(); ++i) {
|
|
1400
|
+
std::cout << tensor.data<int64_t>()[i] << " ";
|
|
1401
|
+
}
|
|
1402
|
+
std::cout << std::endl;
|
|
1403
|
+
break;
|
|
1404
|
+
default:
|
|
1405
|
+
break;
|
|
1406
|
+
}
|
|
1407
|
+
}
|
|
1408
|
+
|
|
1409
|
+
void print_output_tensor_info(const std::string & name, const ov::Tensor & tensor, const void * output_dst) {
|
|
1410
|
+
std::cout << "Output name: " << name << ", Output shape: " << tensor.get_shape() << ", Address: " << output_dst
|
|
1411
|
+
<< std::endl;
|
|
1412
|
+
|
|
1413
|
+
auto print_float_stats = [](const std::string & type_name, size_t size, auto get_value) {
|
|
1414
|
+
if (size == 0) {
|
|
1415
|
+
return;
|
|
1416
|
+
}
|
|
1417
|
+
|
|
1418
|
+
float first = get_value(0);
|
|
1419
|
+
float min = first;
|
|
1420
|
+
float max = first;
|
|
1421
|
+
double sum = first;
|
|
1422
|
+
|
|
1423
|
+
for (size_t i = 1; i < size; ++i) {
|
|
1424
|
+
float v = get_value(i);
|
|
1425
|
+
if (v < min) {
|
|
1426
|
+
min = v;
|
|
1427
|
+
}
|
|
1428
|
+
if (v > max) {
|
|
1429
|
+
max = v;
|
|
1430
|
+
}
|
|
1431
|
+
sum += v;
|
|
1432
|
+
}
|
|
1433
|
+
double mean = sum / size;
|
|
1434
|
+
|
|
1435
|
+
std::cout << std::right << std::setw(6) << type_name << std::right << std::setw(12) << "First" << std::setw(12)
|
|
1436
|
+
<< "Min" << std::setw(12) << "Max" << std::setw(12) << "Mean" << std::endl;
|
|
1437
|
+
std::cout << std::right << std::setw(6) << "" << std::right << std::setw(12) << first << std::setw(12) << min
|
|
1438
|
+
<< std::setw(12) << max << std::setw(12) << mean << std::endl;
|
|
1439
|
+
};
|
|
1440
|
+
|
|
1441
|
+
switch (tensor.get_element_type()) {
|
|
1442
|
+
case ov::element::f32: {
|
|
1443
|
+
const float * data = tensor.data<float>();
|
|
1444
|
+
size_t size = tensor.get_size();
|
|
1445
|
+
print_float_stats("[f32]", size, [data](size_t i) { return data[i]; });
|
|
1446
|
+
break;
|
|
1447
|
+
}
|
|
1448
|
+
case ov::element::f16: {
|
|
1449
|
+
const ov::float16 * data = tensor.data<ov::float16>();
|
|
1450
|
+
size_t size = tensor.get_size();
|
|
1451
|
+
print_float_stats("[f16]", size, [data](size_t i) { return static_cast<float>(data[i]); });
|
|
1452
|
+
break;
|
|
1453
|
+
}
|
|
1454
|
+
default:
|
|
1455
|
+
break;
|
|
1456
|
+
}
|
|
1457
|
+
}
|
|
1458
|
+
|
|
1459
|
+
const ggml_tensor * get_inp_pos_tensor(ggml_cgraph * cgraph) {
|
|
1460
|
+
for (int i = 0; i < cgraph->n_nodes; ++i) {
|
|
1461
|
+
auto * op = cgraph->nodes[i];
|
|
1462
|
+
for (int j = 0; j < GGML_MAX_SRC; ++j) {
|
|
1463
|
+
auto * src = op->src[j];
|
|
1464
|
+
if (src == nullptr) {
|
|
1465
|
+
break;
|
|
1466
|
+
}
|
|
1467
|
+
if (GgmlOvDecoder::is_inp_pos(src, op)) {
|
|
1468
|
+
return src;
|
|
1469
|
+
}
|
|
1470
|
+
}
|
|
1471
|
+
}
|
|
1472
|
+
GGML_LOG_ERROR("get_inp_pos_tensor: inp_pos not found in cgraph");
|
|
1473
|
+
throw std::runtime_error("get_inp_pos_tensor: inp_pos not found in cgraph");
|
|
1474
|
+
}
|
|
1475
|
+
|
|
1476
|
+
int64_t get_inp_pos_n_tokens(ggml_cgraph * cgraph, const ggml_tensor * inp_pos) {
|
|
1477
|
+
// IMROPE stacks n_planes (t/h/w/e) position planes into inp_pos, so ne[0] is
|
|
1478
|
+
// n_planes * n_tokens. Callers that need a token count must divide the planes out.
|
|
1479
|
+
int n_planes = 1;
|
|
1480
|
+
for (int i = 0; i < cgraph->n_nodes; ++i) {
|
|
1481
|
+
auto * op = cgraph->nodes[i];
|
|
1482
|
+
for (int j = 0; j < GGML_MAX_SRC; ++j) {
|
|
1483
|
+
if (op->src[j] == inp_pos) {
|
|
1484
|
+
n_planes = GgmlOvDecoder::get_inp_pos_n_planes(op);
|
|
1485
|
+
break;
|
|
1486
|
+
}
|
|
1487
|
+
}
|
|
1488
|
+
}
|
|
1489
|
+
return inp_pos->ne[0] / n_planes;
|
|
1490
|
+
}
|
|
1491
|
+
|
|
1492
|
+
bool get_is_prefill(ggml_cgraph * cgraph, const ggml_tensor * inp_pos) {
|
|
1493
|
+
return get_inp_pos_n_tokens(cgraph, inp_pos) > 1;
|
|
1494
|
+
}
|
|
1495
|
+
|
|
1496
|
+
#pragma GCC diagnostic pop
|