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 +1,859 @@
|
|
|
1
|
+
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
|
|
2
|
+
#define USE_CUB
|
|
3
|
+
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
|
|
4
|
+
|
|
5
|
+
#ifdef USE_CUB
|
|
6
|
+
#include <cub/cub.cuh>
|
|
7
|
+
using namespace cub;
|
|
8
|
+
#endif // USE_CUB
|
|
9
|
+
|
|
10
|
+
#include "ssm-scan.cuh"
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
// Minimum number of tokens to use SSD (State Space Duality) matmul path instead of scan path.
|
|
14
|
+
// For n_tok <= this threshold, the scan kernel is used (lower overhead for short sequences).
|
|
15
|
+
#define SSM_SSD_MIN_TOKENS 128
|
|
16
|
+
|
|
17
|
+
// prepare_dt kernel dimensions: one block per (head, seq), each block handles DT_MAX_ITEMS items.
|
|
18
|
+
#define SSM_SSD_DT_BLOCK 256
|
|
19
|
+
#define SSM_SSD_DT_MAX_ITEMS 32
|
|
20
|
+
|
|
21
|
+
// Maximum tokens the SSD path supports, derived from the prepare_dt kernel block capacity.
|
|
22
|
+
#define SSM_SSD_MAX_TOKENS (SSM_SSD_DT_BLOCK * SSM_SSD_DT_MAX_ITEMS)
|
|
23
|
+
|
|
24
|
+
// Chunk size for chunked SSD. Caps matmul cost at O(chunk^2) per chunk.
|
|
25
|
+
#define SSM_SSD_CHUNK_SIZE 256
|
|
26
|
+
|
|
27
|
+
// We would like to keep pragma unroll for cases where L_template is not 0,
|
|
28
|
+
// so we suppress the clang transformation warning.
|
|
29
|
+
#ifdef __clang__
|
|
30
|
+
#pragma clang diagnostic push
|
|
31
|
+
#pragma clang diagnostic ignored "-Wpass-failed"
|
|
32
|
+
#endif // __clang__
|
|
33
|
+
template <size_t splitD, size_t N, size_t L_template>
|
|
34
|
+
__global__ void __launch_bounds__(splitD, 1)
|
|
35
|
+
ssm_scan_f32(const float * src0_ptr, const float * src1_ptr, const float * src2_ptr,
|
|
36
|
+
const float * src3_ptr, const float * src4_ptr, const float * src5_ptr,
|
|
37
|
+
const int32_t * src6_ptr, float * dst_ptr,
|
|
38
|
+
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
|
|
39
|
+
const int src2_nb1, const int src2_nb2, const int src3_nb1,
|
|
40
|
+
const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
|
|
41
|
+
const int64_t s_off, const int64_t d_inner, const int64_t L_param)
|
|
42
|
+
{
|
|
43
|
+
const float * GGML_CUDA_RESTRICT src0 = src0_ptr;
|
|
44
|
+
const float * GGML_CUDA_RESTRICT src1 = src1_ptr;
|
|
45
|
+
const float * GGML_CUDA_RESTRICT src2 = src2_ptr;
|
|
46
|
+
const float * GGML_CUDA_RESTRICT src3 = src3_ptr;
|
|
47
|
+
const float * GGML_CUDA_RESTRICT src4 = src4_ptr;
|
|
48
|
+
const float * GGML_CUDA_RESTRICT src5 = src5_ptr;
|
|
49
|
+
const int32_t * GGML_CUDA_RESTRICT src6 = src6_ptr;
|
|
50
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
51
|
+
const size_t L = L_template == 0 ? L_param : L_template;
|
|
52
|
+
ggml_cuda_pdl_sync();
|
|
53
|
+
const float *s0_block = (const float *)((const char *)src0 + src6[blockIdx.x] * src0_nb3 + blockIdx.y * splitD * src0_nb2);
|
|
54
|
+
const float *x_block = (const float *)((const char *)src1 + (blockIdx.x * src1_nb3) + blockIdx.y * splitD * sizeof(float));
|
|
55
|
+
const float *dt_block = (const float *)((const char *)src2 + (blockIdx.x * src2_nb2) + blockIdx.y * splitD * sizeof(float));
|
|
56
|
+
const float *A_block = (const float *)((const char *)src3 + blockIdx.y * splitD * src3_nb1);
|
|
57
|
+
const float *B_block = (const float *)((const char *)src4 + (blockIdx.x * src4_nb3));
|
|
58
|
+
const float *C_block = (const float *)((const char *)src5 + (blockIdx.x * src5_nb3));
|
|
59
|
+
float *y_block = (float *)((char *)dst + (blockIdx.x * d_inner * L * sizeof(float)) + blockIdx.y * splitD * sizeof(float));
|
|
60
|
+
float *s_block = (float *)((char *)dst + s_off + blockIdx.x * src0_nb3 + blockIdx.y * splitD * src0_nb2);
|
|
61
|
+
|
|
62
|
+
const int stride_x = src1_nb2 / sizeof(float);
|
|
63
|
+
const int stride_dt = src2_nb1 / sizeof(float);
|
|
64
|
+
const int stride_B = src4_nb2 / sizeof(float);
|
|
65
|
+
const int stride_C = src5_nb2 / sizeof(float);
|
|
66
|
+
const int stride_y = d_inner;
|
|
67
|
+
|
|
68
|
+
float regA[N];
|
|
69
|
+
float regs0[N];
|
|
70
|
+
|
|
71
|
+
__shared__ float smemB[N];
|
|
72
|
+
__shared__ float smemC[N];
|
|
73
|
+
|
|
74
|
+
#ifdef USE_CUB
|
|
75
|
+
using BlockLoad = cub::BlockLoad<float, splitD, N, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
|
|
76
|
+
using BlockStore = cub::BlockStore<float, splitD, N, cub::BLOCK_STORE_WARP_TRANSPOSE>;
|
|
77
|
+
|
|
78
|
+
union CubTempStorage {
|
|
79
|
+
typename BlockLoad::TempStorage load_temp;
|
|
80
|
+
typename BlockStore::TempStorage store_temp;
|
|
81
|
+
};
|
|
82
|
+
__shared__ CubTempStorage cub_temp_storage;
|
|
83
|
+
|
|
84
|
+
BlockLoad(cub_temp_storage.load_temp).Load(A_block, regA);
|
|
85
|
+
__syncthreads();
|
|
86
|
+
BlockLoad(cub_temp_storage.load_temp).Load(s0_block, regs0);
|
|
87
|
+
#else
|
|
88
|
+
const int stride_s0 = src0_nb2 / sizeof(float);
|
|
89
|
+
const int stride_A = src3_nb1 / sizeof(float);
|
|
90
|
+
#pragma unroll
|
|
91
|
+
for (size_t n = 0; n < N; ++n)
|
|
92
|
+
{
|
|
93
|
+
regA[n] = A_block[threadIdx.x * stride_A + n];
|
|
94
|
+
regs0[n] = s0_block[threadIdx.x * stride_s0 + n];
|
|
95
|
+
}
|
|
96
|
+
#endif
|
|
97
|
+
|
|
98
|
+
#pragma unroll
|
|
99
|
+
for (size_t i = 0; i < L; i++)
|
|
100
|
+
{
|
|
101
|
+
if (threadIdx.x < N)
|
|
102
|
+
{
|
|
103
|
+
smemB[threadIdx.x] = B_block[i * stride_B + threadIdx.x];
|
|
104
|
+
smemC[threadIdx.x] = C_block[i * stride_C + threadIdx.x];
|
|
105
|
+
}
|
|
106
|
+
__syncthreads();
|
|
107
|
+
|
|
108
|
+
float dt_soft_plus = dt_block[i * stride_dt + threadIdx.x];
|
|
109
|
+
if (dt_soft_plus <= 20.0f)
|
|
110
|
+
{
|
|
111
|
+
dt_soft_plus = log1pf(expf(dt_soft_plus));
|
|
112
|
+
}
|
|
113
|
+
float x_dt = x_block[i * stride_x + threadIdx.x] * dt_soft_plus;
|
|
114
|
+
|
|
115
|
+
float sumf = 0.0f;
|
|
116
|
+
#pragma unroll
|
|
117
|
+
for (size_t n = 0; n < N; n++)
|
|
118
|
+
{
|
|
119
|
+
float state = regs0[n] * expf(dt_soft_plus * regA[n]) + smemB[n] * x_dt;
|
|
120
|
+
sumf += state * smemC[n];
|
|
121
|
+
regs0[n] = state;
|
|
122
|
+
}
|
|
123
|
+
y_block[i * stride_y + threadIdx.x] = sumf;
|
|
124
|
+
__syncthreads();
|
|
125
|
+
}
|
|
126
|
+
|
|
127
|
+
#ifdef USE_CUB
|
|
128
|
+
BlockStore(cub_temp_storage.store_temp).Store(s_block, regs0);
|
|
129
|
+
#else
|
|
130
|
+
const int stride_s = stride_s0;
|
|
131
|
+
#pragma unroll
|
|
132
|
+
for (size_t n = 0; n < N; ++n)
|
|
133
|
+
{
|
|
134
|
+
s_block[threadIdx.x * stride_s + n] = regs0[n];
|
|
135
|
+
}
|
|
136
|
+
#endif
|
|
137
|
+
}
|
|
138
|
+
#ifdef __clang__
|
|
139
|
+
#pragma clang diagnostic pop
|
|
140
|
+
#endif // __clang__
|
|
141
|
+
|
|
142
|
+
// assumes as many threads as d_state
|
|
143
|
+
template <int c_factor, int d_state>
|
|
144
|
+
__global__ void __launch_bounds__(d_state, 1)
|
|
145
|
+
ssm_scan_f32_group(
|
|
146
|
+
const float * src0_ptr, const float * src1_ptr, const float * src2_ptr,
|
|
147
|
+
const float * src3_ptr, const float * src4_ptr, const float * src5_ptr,
|
|
148
|
+
const int32_t * src6_ptr, float * dst_ptr,
|
|
149
|
+
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
|
|
150
|
+
const int src2_nb1, const int src2_nb2, const int src3_nb1,
|
|
151
|
+
const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
|
|
152
|
+
const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok, const int64_t K) {
|
|
153
|
+
const float * GGML_CUDA_RESTRICT src0 = src0_ptr;
|
|
154
|
+
const float * GGML_CUDA_RESTRICT src1 = src1_ptr;
|
|
155
|
+
const float * GGML_CUDA_RESTRICT src2 = src2_ptr;
|
|
156
|
+
const float * GGML_CUDA_RESTRICT src3 = src3_ptr;
|
|
157
|
+
const float * GGML_CUDA_RESTRICT src4 = src4_ptr;
|
|
158
|
+
const float * GGML_CUDA_RESTRICT src5 = src5_ptr;
|
|
159
|
+
const int32_t * GGML_CUDA_RESTRICT src6 = src6_ptr;
|
|
160
|
+
float * GGML_CUDA_RESTRICT dst = dst_ptr;
|
|
161
|
+
|
|
162
|
+
const int warp = threadIdx.x / WARP_SIZE;
|
|
163
|
+
const int lane = threadIdx.x % WARP_SIZE;
|
|
164
|
+
const int warp_idx = blockIdx.x * c_factor + warp;
|
|
165
|
+
|
|
166
|
+
const int head_idx = warp_idx / d_head;
|
|
167
|
+
const int head_off = (warp_idx % d_head) * sizeof(float);
|
|
168
|
+
const int seq_idx = blockIdx.y;
|
|
169
|
+
|
|
170
|
+
const int group_off = (head_idx / (n_head / n_group)) * d_state * sizeof(float);
|
|
171
|
+
|
|
172
|
+
ggml_cuda_pdl_sync();
|
|
173
|
+
// TODO: refactor strides to be in elements/floats instead of bytes to be cleaner and consistent with the rest of the codebase
|
|
174
|
+
const float * s0_warp = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
|
|
175
|
+
const float * x_warp = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + (warp_idx * sizeof(float)));
|
|
176
|
+
const float * dt_warp = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
|
|
177
|
+
const float * A_warp = (const float *) ((const char *) src3 + head_idx * src3_nb1);
|
|
178
|
+
const float * B_warp = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
|
|
179
|
+
const float * C_warp = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
|
|
180
|
+
float * y_warp = dst + (seq_idx * n_tok * n_head * d_head) + warp_idx;
|
|
181
|
+
float * s_warp = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
|
|
182
|
+
|
|
183
|
+
// strides across n_seq_tokens
|
|
184
|
+
const int stride_x = src1_nb2 / sizeof(float);
|
|
185
|
+
const int stride_dt = src2_nb1 / sizeof(float);
|
|
186
|
+
const int stride_B = src4_nb2 / sizeof(float);
|
|
187
|
+
const int stride_C = src5_nb2 / sizeof(float);
|
|
188
|
+
const int stride_y = n_head * d_head;
|
|
189
|
+
|
|
190
|
+
float state[c_factor];
|
|
191
|
+
float state_sum = 0.0f;
|
|
192
|
+
|
|
193
|
+
#pragma unroll
|
|
194
|
+
for (int j = 0; j < c_factor; j++) {
|
|
195
|
+
state[j] = s0_warp[WARP_SIZE * j + lane];
|
|
196
|
+
}
|
|
197
|
+
|
|
198
|
+
for (int64_t i = 0; i < n_tok; i++) {
|
|
199
|
+
// NOTE: dt_soft_plus, dA and x_dt have the same value for a warp here.
|
|
200
|
+
// Recalculation is intentional; sharing via shuffles/smem proved slower due to sync overhead.
|
|
201
|
+
const float dt_soft_plus = (dt_warp[i * stride_dt] <= 20.0f ? log1pf(expf(dt_warp[i * stride_dt])) : dt_warp[i * stride_dt]);
|
|
202
|
+
|
|
203
|
+
state_sum = 0.0f;
|
|
204
|
+
const float dA = expf(dt_soft_plus * A_warp[0]);
|
|
205
|
+
const float x_dt = x_warp[i * stride_x] * dt_soft_plus;
|
|
206
|
+
#pragma unroll
|
|
207
|
+
for (int j = 0; j < c_factor; j++) {
|
|
208
|
+
const float B_val = B_warp[i * stride_B + WARP_SIZE * j + lane];
|
|
209
|
+
const float C_val = C_warp[i * stride_C + WARP_SIZE * j + lane];
|
|
210
|
+
state[j] = (state[j] * dA) + (B_val * x_dt);
|
|
211
|
+
state_sum += state[j] * C_val;
|
|
212
|
+
}
|
|
213
|
+
|
|
214
|
+
// parallel accumulation for output
|
|
215
|
+
state_sum = warp_reduce_sum(state_sum);
|
|
216
|
+
|
|
217
|
+
if (lane == 0) {
|
|
218
|
+
y_warp[i * stride_y] = state_sum;
|
|
219
|
+
}
|
|
220
|
+
|
|
221
|
+
// Slot 0 is the final state written below; slots 1..K-1 are rollback snapshots.
|
|
222
|
+
const int64_t slot = n_tok - 1 - i;
|
|
223
|
+
if (K > 1 && slot > 0 && slot < K) {
|
|
224
|
+
float * s_snapshot_warp = (float *) ((char *) dst + s_off + (slot * gridDim.y + seq_idx) * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
|
|
225
|
+
#pragma unroll
|
|
226
|
+
for (int j = 0; j < c_factor; j++) {
|
|
227
|
+
s_snapshot_warp[WARP_SIZE * j + lane] = state[j];
|
|
228
|
+
}
|
|
229
|
+
}
|
|
230
|
+
}
|
|
231
|
+
|
|
232
|
+
// write back the state
|
|
233
|
+
#pragma unroll
|
|
234
|
+
for (int j = 0; j < c_factor; j++) {
|
|
235
|
+
s_warp[WARP_SIZE * j + lane] = state[j];
|
|
236
|
+
}
|
|
237
|
+
}
|
|
238
|
+
|
|
239
|
+
static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
|
|
240
|
+
const float * src4, const float * src5, const int32_t * src6, float * dst,
|
|
241
|
+
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
|
|
242
|
+
const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
|
|
243
|
+
const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
|
|
244
|
+
const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
|
|
245
|
+
const int64_t K, cudaStream_t stream) {
|
|
246
|
+
// NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
|
|
247
|
+
if (src3_nb1 == sizeof(float)) {
|
|
248
|
+
// Mamba-2
|
|
249
|
+
if (d_state == 128) {
|
|
250
|
+
constexpr int threads = 128;
|
|
251
|
+
constexpr int num_warps = threads/WARP_SIZE;
|
|
252
|
+
|
|
253
|
+
const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
|
|
254
|
+
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
|
|
255
|
+
ggml_cuda_kernel_launch(ssm_scan_f32_group<128/WARP_SIZE, 128>, launch_params,
|
|
256
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
257
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
|
|
258
|
+
src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok, K);
|
|
259
|
+
} else if (d_state == 256) { // Falcon-H1
|
|
260
|
+
constexpr int threads = 256;
|
|
261
|
+
constexpr int num_warps = threads/WARP_SIZE;
|
|
262
|
+
|
|
263
|
+
const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
|
|
264
|
+
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
|
|
265
|
+
ggml_cuda_kernel_launch(ssm_scan_f32_group<256/WARP_SIZE, 256>, launch_params,
|
|
266
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
267
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
|
|
268
|
+
src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok, K);
|
|
269
|
+
} else {
|
|
270
|
+
GGML_ABORT("doesn't support d_state!=(128 or 256).");
|
|
271
|
+
}
|
|
272
|
+
} else {
|
|
273
|
+
// Mamba-1
|
|
274
|
+
GGML_ASSERT(K == 1);
|
|
275
|
+
constexpr int threads = 128;
|
|
276
|
+
GGML_ASSERT(n_head % threads == 0);
|
|
277
|
+
GGML_ASSERT(head_dim == 1);
|
|
278
|
+
GGML_ASSERT(n_group == 1);
|
|
279
|
+
const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
|
|
280
|
+
if (d_state == 16) {
|
|
281
|
+
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
|
|
282
|
+
switch (n_tok)
|
|
283
|
+
{
|
|
284
|
+
case 1:
|
|
285
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 1>, launch_params,
|
|
286
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
287
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
288
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
289
|
+
break;
|
|
290
|
+
case 2:
|
|
291
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 2>, launch_params,
|
|
292
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
293
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
294
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
295
|
+
break;
|
|
296
|
+
case 3:
|
|
297
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 3>, launch_params,
|
|
298
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
299
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
300
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
301
|
+
break;
|
|
302
|
+
case 4:
|
|
303
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 4>, launch_params,
|
|
304
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
305
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
306
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
307
|
+
break;
|
|
308
|
+
case 5:
|
|
309
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 5>, launch_params,
|
|
310
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
311
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
312
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
313
|
+
break;
|
|
314
|
+
case 6:
|
|
315
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 6>, launch_params,
|
|
316
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
317
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
318
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
319
|
+
break;
|
|
320
|
+
case 7:
|
|
321
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 7>, launch_params,
|
|
322
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
323
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
324
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
325
|
+
break;
|
|
326
|
+
case 8:
|
|
327
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 8>, launch_params,
|
|
328
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
329
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
330
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
331
|
+
break;
|
|
332
|
+
default:
|
|
333
|
+
ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 0>, launch_params,
|
|
334
|
+
src0, src1, src2, src3, src4, src5, src6, dst,
|
|
335
|
+
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
|
|
336
|
+
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
|
|
337
|
+
break;
|
|
338
|
+
}
|
|
339
|
+
} else {
|
|
340
|
+
GGML_ABORT("doesn't support d_state!=16.");
|
|
341
|
+
}
|
|
342
|
+
}
|
|
343
|
+
}
|
|
344
|
+
|
|
345
|
+
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
|
346
|
+
// ============================================================================
|
|
347
|
+
// SSD (State Space Duality) kernels for Mamba-2 prefill (n_tok > SSM_SSD_MIN_TOKENS)
|
|
348
|
+
//
|
|
349
|
+
// Instead of a sequential scan, SSD reformulates the output as:
|
|
350
|
+
// Y = (L (.) (C @ B^T)) @ (X * dt) + decay * C @ s_init
|
|
351
|
+
// where L is a causal decay mask derived from A and dt.
|
|
352
|
+
//
|
|
353
|
+
// This converts the O(T*N) sequential scan into parallel matmuls.
|
|
354
|
+
// ============================================================================
|
|
355
|
+
// Softplus(dt) and inclusive prefix sum per head using CUB BlockScan.
|
|
356
|
+
// Grid: (n_head, n_seqs)
|
|
357
|
+
template <int BLOCK_SIZE, int MAX_ITEMS>
|
|
358
|
+
__global__ void ssm_ssd_prepare_dt_kernel(
|
|
359
|
+
const float * __restrict__ dt_raw,
|
|
360
|
+
float * __restrict__ dt_sp_out,
|
|
361
|
+
float * __restrict__ cs_out,
|
|
362
|
+
const int n_head, const int n_tok,
|
|
363
|
+
const int dt_stride_tok, // elements between tokens in dt
|
|
364
|
+
const int dt_stride_seq) { // elements between sequences in dt
|
|
365
|
+
|
|
366
|
+
const int h = blockIdx.x;
|
|
367
|
+
const int s = blockIdx.y;
|
|
368
|
+
|
|
369
|
+
const float * dt_seq = dt_raw + s * dt_stride_seq;
|
|
370
|
+
|
|
371
|
+
float * dt_sp_seq = dt_sp_out + s * n_tok * n_head;
|
|
372
|
+
float * cs_seq = cs_out + s * n_tok * n_head;
|
|
373
|
+
|
|
374
|
+
const int items_per_thread = (n_tok + BLOCK_SIZE - 1) / BLOCK_SIZE;
|
|
375
|
+
|
|
376
|
+
// Phase 1: softplus with interleaved distribution (t = i*BLOCK_SIZE + threadIdx.x).
|
|
377
|
+
// Each warp reads BLOCK_SIZE consecutive tokens, giving coalesced dt_raw loads
|
|
378
|
+
// (stride n_head between threads vs. items_per_thread*n_head in blocked layout).
|
|
379
|
+
float local_vals[MAX_ITEMS];
|
|
380
|
+
for (int i = 0; i < items_per_thread; i++) {
|
|
381
|
+
const int t = i * BLOCK_SIZE + threadIdx.x;
|
|
382
|
+
if (t < n_tok) {
|
|
383
|
+
float val = dt_seq[h + t * dt_stride_tok];
|
|
384
|
+
float sp = (val <= 20.0f) ? log1pf(expf(val)) : val;
|
|
385
|
+
local_vals[i] = sp;
|
|
386
|
+
dt_sp_seq[t * n_head + h] = sp;
|
|
387
|
+
} else {
|
|
388
|
+
local_vals[i] = 0.0f;
|
|
389
|
+
}
|
|
390
|
+
}
|
|
391
|
+
|
|
392
|
+
// Phase 2+3: per-step inclusive scan to build cs[] in token order.
|
|
393
|
+
// With interleaved distribution the per-thread total scan would not give token-order
|
|
394
|
+
// prefix sums, so we scan one BLOCK_SIZE slab at a time and carry a running total.
|
|
395
|
+
#ifdef USE_CUB
|
|
396
|
+
using BlockScan = cub::BlockScan<float, BLOCK_SIZE>;
|
|
397
|
+
__shared__ typename BlockScan::TempStorage scan_temp;
|
|
398
|
+
__shared__ float step_total;
|
|
399
|
+
|
|
400
|
+
float running = 0.0f;
|
|
401
|
+
for (int i = 0; i < items_per_thread; i++) {
|
|
402
|
+
float inclusive;
|
|
403
|
+
BlockScan(scan_temp).InclusiveSum(local_vals[i], inclusive);
|
|
404
|
+
const int t = i * BLOCK_SIZE + threadIdx.x;
|
|
405
|
+
if (t < n_tok) {
|
|
406
|
+
cs_seq[t * n_head + h] = running + inclusive;
|
|
407
|
+
}
|
|
408
|
+
if (threadIdx.x == BLOCK_SIZE - 1) {
|
|
409
|
+
step_total = inclusive;
|
|
410
|
+
}
|
|
411
|
+
__syncthreads();
|
|
412
|
+
running += step_total;
|
|
413
|
+
}
|
|
414
|
+
#else
|
|
415
|
+
// Fallback: sequential prefix scan in shared memory, one slab at a time.
|
|
416
|
+
__shared__ float sdata[BLOCK_SIZE];
|
|
417
|
+
float running = 0.0f;
|
|
418
|
+
for (int i = 0; i < items_per_thread; i++) {
|
|
419
|
+
const int t = i * BLOCK_SIZE + threadIdx.x;
|
|
420
|
+
sdata[threadIdx.x] = local_vals[i];
|
|
421
|
+
__syncthreads();
|
|
422
|
+
if (threadIdx.x == 0) {
|
|
423
|
+
for (int j = 1; j < BLOCK_SIZE; j++) {
|
|
424
|
+
sdata[j] += sdata[j - 1];
|
|
425
|
+
}
|
|
426
|
+
}
|
|
427
|
+
__syncthreads();
|
|
428
|
+
if (t < n_tok) {
|
|
429
|
+
cs_seq[t * n_head + h] = running + sdata[threadIdx.x];
|
|
430
|
+
}
|
|
431
|
+
running += sdata[BLOCK_SIZE - 1];
|
|
432
|
+
__syncthreads();
|
|
433
|
+
}
|
|
434
|
+
#endif
|
|
435
|
+
}
|
|
436
|
+
|
|
437
|
+
// Prepare SSD matmul inputs for one chunk: X_dt, B_weighted, C_scaled.
|
|
438
|
+
// T_matmul controls precision for X_dt, B_weighted (float or half).
|
|
439
|
+
// C_scaled is always float (pairs with float s_cur in step 3c).
|
|
440
|
+
// Computation is always FP32; only the final store converts to T_matmul.
|
|
441
|
+
// Also materializes the causal M matrix = exp(A*(cs_out - cs_in)) * CB (fused with prep to save a launch).
|
|
442
|
+
// Grid: (ceil(max(C*head_dim, d_state*C, chunk_len^2) / BLOCK), n_head, n_seqs)
|
|
443
|
+
template <int BLOCK_SIZE, typename T_matmul>
|
|
444
|
+
__global__ void ssm_ssd_pre_matmul_kernel(
|
|
445
|
+
const float * __restrict__ cs, // {n_tok, n_head} cumulative dt sums
|
|
446
|
+
const float * __restrict__ dt_sp, // {n_tok, n_head} softplus(dt)
|
|
447
|
+
const float * __restrict__ A, // {1, n_head}
|
|
448
|
+
const float * __restrict__ x, // {head_dim, n_head, n_tok, n_seqs}
|
|
449
|
+
const float * __restrict__ B, // {d_state, n_group, n_tok, n_seqs}
|
|
450
|
+
const float * __restrict__ C_src, // {d_state, n_group, n_tok, n_seqs}
|
|
451
|
+
T_matmul * __restrict__ X_dt, // {head_dim, C, n_head} x * dt, d-fastest
|
|
452
|
+
T_matmul * __restrict__ B_weighted, // {d_state, C, n_head} B * decay_from_end
|
|
453
|
+
float * __restrict__ C_scaled, // {d_state, C, n_head} C * decay_to_pos (always float)
|
|
454
|
+
const float * __restrict__ CB, // {chunk_len, chunk_len, n_group, n_seqs}
|
|
455
|
+
half * __restrict__ M_out, // {chunk_len, chunk_len, n_head, n_seqs}
|
|
456
|
+
const int chunk_len, const int head_dim, const int n_head, const int n_group,
|
|
457
|
+
const int d_state, const int A_stride,
|
|
458
|
+
const int x_stride_tok, const int x_stride_seq,
|
|
459
|
+
const int B_stride_tok, const int B_stride_seq,
|
|
460
|
+
const int C_stride_tok, const int C_stride_seq,
|
|
461
|
+
const int chunk_offset,
|
|
462
|
+
const int n_tok_total) {
|
|
463
|
+
|
|
464
|
+
const int h = blockIdx.y;
|
|
465
|
+
const int s = blockIdx.z;
|
|
466
|
+
const int g = h / (n_head / n_group);
|
|
467
|
+
|
|
468
|
+
const float A_h = A[h * A_stride];
|
|
469
|
+
const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
|
|
470
|
+
|
|
471
|
+
const int cs_seq_off = s * n_tok_total * n_head;
|
|
472
|
+
const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
|
|
473
|
+
const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
|
|
474
|
+
|
|
475
|
+
// Prepare X_dt = x * dt, stored d-fastest for coalesced reads and writes.
|
|
476
|
+
const int n_xdt = chunk_len * head_dim;
|
|
477
|
+
if (idx < n_xdt) {
|
|
478
|
+
const int d = idx % head_dim;
|
|
479
|
+
const int t = idx / head_dim;
|
|
480
|
+
|
|
481
|
+
const float x_val = x[s * x_stride_seq + (chunk_offset + t) * x_stride_tok + d + h * head_dim];
|
|
482
|
+
const float dt_val = dt_sp[cs_seq_off + (chunk_offset + t) * n_head + h];
|
|
483
|
+
|
|
484
|
+
X_dt[d + t * head_dim + h * n_xdt + s * n_xdt * n_head] = (T_matmul)(x_val * dt_val);
|
|
485
|
+
}
|
|
486
|
+
|
|
487
|
+
// Prepare B_weighted and C_scaled together: both share the same index space (d_state * chunk_len)
|
|
488
|
+
// and the same cs_t load, so merging halves the cs[] global memory traffic.
|
|
489
|
+
const int n_bw = d_state * chunk_len;
|
|
490
|
+
if (idx < n_bw) {
|
|
491
|
+
const int n = idx % d_state;
|
|
492
|
+
const int t = idx / d_state;
|
|
493
|
+
|
|
494
|
+
const float cs_t = cs[cs_seq_off + (chunk_offset + t) * n_head + h] - cs_base;
|
|
495
|
+
|
|
496
|
+
const float B_val = B[s * B_stride_seq + (chunk_offset + t) * B_stride_tok + g * d_state + n];
|
|
497
|
+
B_weighted[n + t * d_state + h * n_bw + s * n_bw * n_head] = (T_matmul)(B_val * __expf(A_h * (cs_last - cs_t)));
|
|
498
|
+
|
|
499
|
+
const float C_val = C_src[s * C_stride_seq + (chunk_offset + t) * C_stride_tok + g * d_state + n];
|
|
500
|
+
C_scaled[n + t * d_state + h * n_bw + s * n_bw * n_head] = C_val * __expf(A_h * cs_t);
|
|
501
|
+
}
|
|
502
|
+
|
|
503
|
+
// Materialize M = exp(A*(cs_out - cs_in)) * CB with causal mask.
|
|
504
|
+
const int n_M = chunk_len * chunk_len;
|
|
505
|
+
if (idx < n_M) {
|
|
506
|
+
const int t_out = idx % chunk_len;
|
|
507
|
+
const int t_in = idx / chunk_len;
|
|
508
|
+
|
|
509
|
+
half val;
|
|
510
|
+
if (t_in <= t_out) {
|
|
511
|
+
const float cs_out = cs[cs_seq_off + (chunk_offset + t_out) * n_head + h] - cs_base;
|
|
512
|
+
const float cs_in = cs[cs_seq_off + (chunk_offset + t_in) * n_head + h] - cs_base;
|
|
513
|
+
const float decay = __expf(A_h * (cs_out - cs_in));
|
|
514
|
+
const float * CB_g = CB + (int64_t)s * chunk_len * chunk_len * n_group
|
|
515
|
+
+ (int64_t)g * chunk_len * chunk_len;
|
|
516
|
+
const float cb_val = CB_g[t_out + t_in * chunk_len];
|
|
517
|
+
val = __float2half(decay * cb_val);
|
|
518
|
+
} else {
|
|
519
|
+
val = __float2half(0.0f);
|
|
520
|
+
}
|
|
521
|
+
|
|
522
|
+
M_out[(int64_t)s * n_M * n_head + (int64_t)h * n_M + t_in * chunk_len + t_out] = val;
|
|
523
|
+
}
|
|
524
|
+
}
|
|
525
|
+
|
|
526
|
+
// Scale running state in-place: s_cur *= decay_total(chunk).
|
|
527
|
+
// Called BEFORE cuBLAS state update (beta=1) to fuse inter-chunk decay.
|
|
528
|
+
// Eliminates the s_old buffer and D2D memcpy vs the old approach of:
|
|
529
|
+
// memcpy(s_old, s_cur) -> cuBLAS(beta=0) -> s_cur += decay * s_old
|
|
530
|
+
// Grid: (ceil(d_state * head_dim / BLOCK), n_head, n_seqs)
|
|
531
|
+
template <int BLOCK_SIZE>
|
|
532
|
+
__global__ void ssm_ssd_scale_state_kernel(
|
|
533
|
+
float * __restrict__ s_cur, // {d_state, head_dim, n_head, n_seqs}
|
|
534
|
+
const float * __restrict__ cs, // {n_tok, n_head} cumulative dt sums
|
|
535
|
+
const float * __restrict__ A, // {1, n_head}
|
|
536
|
+
const int d_state, const int head_dim, const int n_head,
|
|
537
|
+
const int chunk_offset, const int chunk_len,
|
|
538
|
+
const int n_tok_total, const int A_stride) {
|
|
539
|
+
|
|
540
|
+
const int h = blockIdx.y;
|
|
541
|
+
const int s = blockIdx.z;
|
|
542
|
+
const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
|
|
543
|
+
const int state_per_head = d_state * head_dim;
|
|
544
|
+
if (idx >= state_per_head) return;
|
|
545
|
+
|
|
546
|
+
const float A_h = A[h * A_stride];
|
|
547
|
+
const int cs_seq_off = s * n_tok_total * n_head;
|
|
548
|
+
const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
|
|
549
|
+
const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
|
|
550
|
+
const float decay_total = __expf(A_h * cs_last);
|
|
551
|
+
|
|
552
|
+
const int off = s * state_per_head * n_head + h * state_per_head + idx;
|
|
553
|
+
s_cur[off] *= decay_total;
|
|
554
|
+
}
|
|
555
|
+
|
|
556
|
+
// Copy initial state from src0[ids[s]] into s_cur for each sequence.
|
|
557
|
+
// Grid: (ceil(d_state * head_dim * n_head / BLOCK), n_seqs)
|
|
558
|
+
template <int BLOCK_SIZE>
|
|
559
|
+
__global__ void ssm_ssd_init_state_kernel(
|
|
560
|
+
const float * __restrict__ src0, // {d_state, head_dim, n_head, n_rs}
|
|
561
|
+
const int32_t * __restrict__ ids, // {n_seqs}
|
|
562
|
+
float * __restrict__ s_cur, // {d_state, head_dim, n_head, n_seqs}
|
|
563
|
+
const int state_size, // d_state * head_dim * n_head
|
|
564
|
+
const int64_t s0_stride_seq) { // elements between state rows
|
|
565
|
+
const int s = blockIdx.y;
|
|
566
|
+
const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
|
|
567
|
+
if (idx >= state_size) return;
|
|
568
|
+
|
|
569
|
+
const float * s_src = src0 + (int64_t)ids[s] * s0_stride_seq;
|
|
570
|
+
s_cur[s * state_size + idx] = s_src[idx];
|
|
571
|
+
}
|
|
572
|
+
|
|
573
|
+
// SSD (State Space Duality) dispatch for Mamba-2 prefill.
|
|
574
|
+
// Chunked matmuls: CB, materialize M + cuBLAS Y, S@C, B@X_dt.
|
|
575
|
+
// All strides are in elements (floats), not bytes.
|
|
576
|
+
static void ssm_scan_ssd_f32_cuda(
|
|
577
|
+
ggml_backend_cuda_context & ctx,
|
|
578
|
+
const float * src0_d, const float * src1_d, const float * src2_d, const float * src3_d,
|
|
579
|
+
const float * src4_d, const float * src5_d, const int32_t * src6_d, float * dst_d,
|
|
580
|
+
const int64_t s0_stride_seq, // state (src0) stride between seqs
|
|
581
|
+
const int x_stride_tok, const int x_stride_seq, // x (src1) strides
|
|
582
|
+
const int dt_stride_tok, const int dt_stride_seq, // dt (src2) strides
|
|
583
|
+
const int A_stride, // A (src3) stride between heads
|
|
584
|
+
const int B_stride_tok, const int B_stride_seq, // B (src4) strides
|
|
585
|
+
const int C_stride_tok, const int C_stride_seq, // C (src5) strides
|
|
586
|
+
const int64_t s_off, const int64_t d_state, const int64_t head_dim,
|
|
587
|
+
const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq) {
|
|
588
|
+
|
|
589
|
+
cudaStream_t stream = ctx.stream();
|
|
590
|
+
const int64_t d_inner = head_dim * n_head;
|
|
591
|
+
|
|
592
|
+
const int64_t chunk_size = SSM_SSD_CHUNK_SIZE;
|
|
593
|
+
const int64_t n_chunks = (n_tok + chunk_size - 1) / chunk_size;
|
|
594
|
+
|
|
595
|
+
const int64_t state_per_head = d_state * head_dim;
|
|
596
|
+
|
|
597
|
+
using matmul_t = half;
|
|
598
|
+
static constexpr cudaDataType_t matmul_dtype = CUDA_R_16F;
|
|
599
|
+
|
|
600
|
+
ggml_cuda_pool_alloc<float> dt_sp_buf(ctx.pool(), n_tok * n_head * n_seq);
|
|
601
|
+
ggml_cuda_pool_alloc<float> cs_buf(ctx.pool(), n_tok * n_head * n_seq);
|
|
602
|
+
ggml_cuda_pool_alloc<float> CB_buf(ctx.pool(), chunk_size * chunk_size * n_group * n_seq);
|
|
603
|
+
ggml_cuda_pool_alloc<matmul_t> X_dt_buf(ctx.pool(), chunk_size * head_dim * n_head * n_seq);
|
|
604
|
+
ggml_cuda_pool_alloc<matmul_t> B_w_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
|
|
605
|
+
ggml_cuda_pool_alloc<float> C_s_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
|
|
606
|
+
float * dt_sp = dt_sp_buf.get();
|
|
607
|
+
float * cs = cs_buf.get();
|
|
608
|
+
float * CB = CB_buf.get();
|
|
609
|
+
matmul_t * X_dt = X_dt_buf.get();
|
|
610
|
+
matmul_t * B_weighted = B_w_buf.get();
|
|
611
|
+
float * C_scaled = C_s_buf.get();
|
|
612
|
+
float * s_cur = (float *)((char *)dst_d + s_off); // write state directly to dst
|
|
613
|
+
|
|
614
|
+
// Step 1: softplus(dt) and parallel prefix sum over full sequence
|
|
615
|
+
{
|
|
616
|
+
dim3 grid(n_head, n_seq);
|
|
617
|
+
ssm_ssd_prepare_dt_kernel<SSM_SSD_DT_BLOCK, SSM_SSD_DT_MAX_ITEMS><<<grid, SSM_SSD_DT_BLOCK, 0, stream>>>(
|
|
618
|
+
src2_d, dt_sp, cs, n_head, n_tok, dt_stride_tok, dt_stride_seq);
|
|
619
|
+
CUDA_CHECK(cudaGetLastError());
|
|
620
|
+
}
|
|
621
|
+
|
|
622
|
+
// Step 2: initialize running state from src0[ids[s]]
|
|
623
|
+
{
|
|
624
|
+
constexpr int BLOCK = 256;
|
|
625
|
+
const int64_t state_size = d_state * head_dim * n_head;
|
|
626
|
+
dim3 grid((state_size + BLOCK - 1) / BLOCK, n_seq);
|
|
627
|
+
ssm_ssd_init_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
|
|
628
|
+
src0_d, src6_d, s_cur, state_size, s0_stride_seq);
|
|
629
|
+
CUDA_CHECK(cudaGetLastError());
|
|
630
|
+
}
|
|
631
|
+
|
|
632
|
+
// Step 3: chunked SSD loop
|
|
633
|
+
// Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
|
|
634
|
+
cublasHandle_t handle = ctx.cublas_handle();
|
|
635
|
+
const float alpha_one = 1.0f;
|
|
636
|
+
const float beta_zero = 0.0f;
|
|
637
|
+
const float beta_one = 1.0f;
|
|
638
|
+
const int lda_C_src = C_stride_tok; // leading dim for C in CB = C^T @ B
|
|
639
|
+
const int ldb_B_src = B_stride_tok; // leading dim for B in CB = C^T @ B
|
|
640
|
+
|
|
641
|
+
// Scratch buffer for causal M matrix, reused across chunks (max size at chunk_size)
|
|
642
|
+
const int64_t n_M_max = chunk_size * chunk_size;
|
|
643
|
+
ggml_cuda_pool_alloc<half> M_buf(ctx.pool(), n_M_max * n_head * n_seq);
|
|
644
|
+
half * M_mat = M_buf.get();
|
|
645
|
+
|
|
646
|
+
for (int64_t k = 0; k < n_chunks; k++) {
|
|
647
|
+
const int64_t chunk_offset = k * chunk_size;
|
|
648
|
+
const int64_t chunk_len = (chunk_offset + chunk_size <= n_tok) ? chunk_size : (n_tok - chunk_offset);
|
|
649
|
+
|
|
650
|
+
// 3a: CB = C^T @ B per group
|
|
651
|
+
for (int64_t s = 0; s < n_seq; s++) {
|
|
652
|
+
const float * C_s = src5_d + s * C_stride_seq + chunk_offset * C_stride_tok;
|
|
653
|
+
const float * B_s = src4_d + s * B_stride_seq + chunk_offset * B_stride_tok;
|
|
654
|
+
float * CB_s = CB + s * chunk_len * chunk_len * n_group;
|
|
655
|
+
|
|
656
|
+
if (n_group == 1) {
|
|
657
|
+
CUBLAS_CHECK(cublasSgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N,
|
|
658
|
+
chunk_len, chunk_len, d_state,
|
|
659
|
+
&alpha_one, C_s, lda_C_src, B_s, ldb_B_src,
|
|
660
|
+
&beta_zero, CB_s, (int)chunk_len));
|
|
661
|
+
} else {
|
|
662
|
+
CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
|
|
663
|
+
chunk_len, chunk_len, d_state,
|
|
664
|
+
&alpha_one,
|
|
665
|
+
C_s, CUDA_R_32F, lda_C_src, d_state,
|
|
666
|
+
B_s, CUDA_R_32F, ldb_B_src, d_state,
|
|
667
|
+
&beta_zero,
|
|
668
|
+
CB_s, CUDA_R_32F, (int)chunk_len, (long long)(chunk_len * chunk_len),
|
|
669
|
+
n_group,
|
|
670
|
+
CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
|
|
671
|
+
}
|
|
672
|
+
}
|
|
673
|
+
|
|
674
|
+
// 3b: prepare X_dt, B_weighted, C_scaled + materialize causal M matrix
|
|
675
|
+
const int64_t n_M = chunk_len * chunk_len;
|
|
676
|
+
{
|
|
677
|
+
constexpr int BLOCK = 256;
|
|
678
|
+
const int64_t n_xdt = chunk_len * head_dim;
|
|
679
|
+
const int64_t n_bw = d_state * chunk_len;
|
|
680
|
+
int64_t max_work = n_xdt;
|
|
681
|
+
if (n_bw > max_work) max_work = n_bw;
|
|
682
|
+
if (n_M > max_work) max_work = n_M;
|
|
683
|
+
dim3 grid((max_work + BLOCK - 1) / BLOCK, n_head, n_seq);
|
|
684
|
+
ssm_ssd_pre_matmul_kernel<BLOCK, matmul_t><<<grid, BLOCK, 0, stream>>>(
|
|
685
|
+
cs, dt_sp, src3_d, src1_d, src4_d, src5_d,
|
|
686
|
+
X_dt, B_weighted, C_scaled,
|
|
687
|
+
CB, M_mat,
|
|
688
|
+
chunk_len, head_dim, n_head, n_group, d_state, A_stride,
|
|
689
|
+
x_stride_tok, x_stride_seq, B_stride_tok, B_stride_seq, C_stride_tok, C_stride_seq,
|
|
690
|
+
chunk_offset, n_tok);
|
|
691
|
+
CUDA_CHECK(cudaGetLastError());
|
|
692
|
+
}
|
|
693
|
+
|
|
694
|
+
// 3c: dst = S_cur^T @ C_scaled (state contribution)
|
|
695
|
+
{
|
|
696
|
+
const int64_t stride_S = state_per_head;
|
|
697
|
+
const int64_t stride_Cs = d_state * chunk_len;
|
|
698
|
+
|
|
699
|
+
for (int64_t s = 0; s < n_seq; s++) {
|
|
700
|
+
float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
|
|
701
|
+
|
|
702
|
+
CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
|
|
703
|
+
head_dim, chunk_len, d_state,
|
|
704
|
+
&alpha_one,
|
|
705
|
+
s_cur + s * stride_S * n_head, CUDA_R_32F, d_state, stride_S,
|
|
706
|
+
C_scaled + s * stride_Cs * n_head, CUDA_R_32F, d_state, stride_Cs,
|
|
707
|
+
&beta_zero,
|
|
708
|
+
dst_chunk, CUDA_R_32F, d_inner, head_dim,
|
|
709
|
+
n_head,
|
|
710
|
+
CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
|
|
711
|
+
}
|
|
712
|
+
}
|
|
713
|
+
|
|
714
|
+
// 3d: dst += X_dt @ M^T (intra-chunk contribution, adds to 3c result)
|
|
715
|
+
// M is stored as M[t_out, t_in] (lower-triangular), transpose needed for Y = X @ M^T.
|
|
716
|
+
{
|
|
717
|
+
const int64_t stride_M = n_M;
|
|
718
|
+
const int64_t stride_X_h = (int64_t)chunk_len * head_dim;
|
|
719
|
+
|
|
720
|
+
for (int64_t s = 0; s < n_seq; s++) {
|
|
721
|
+
float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
|
|
722
|
+
CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
|
|
723
|
+
head_dim, chunk_len, chunk_len,
|
|
724
|
+
&alpha_one,
|
|
725
|
+
X_dt + s * stride_X_h * n_head, matmul_dtype, head_dim, stride_X_h,
|
|
726
|
+
M_mat + s * stride_M * n_head, matmul_dtype, chunk_len, stride_M,
|
|
727
|
+
&beta_one,
|
|
728
|
+
dst_chunk, CUDA_R_32F, d_inner, head_dim,
|
|
729
|
+
n_head,
|
|
730
|
+
CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
|
|
731
|
+
}
|
|
732
|
+
}
|
|
733
|
+
|
|
734
|
+
// 3e: s_cur = B_weighted @ X_dt^T + decay_total * s_cur_old (state update)
|
|
735
|
+
{
|
|
736
|
+
// Scale s_cur in-place by per-head decay_total BEFORE cuBLAS overwrites it
|
|
737
|
+
constexpr int BLOCK = 256;
|
|
738
|
+
dim3 grid((state_per_head + BLOCK - 1) / BLOCK, n_head, n_seq);
|
|
739
|
+
ssm_ssd_scale_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
|
|
740
|
+
s_cur, cs, src3_d,
|
|
741
|
+
d_state, head_dim, n_head,
|
|
742
|
+
chunk_offset, chunk_len, n_tok, A_stride);
|
|
743
|
+
CUDA_CHECK(cudaGetLastError());
|
|
744
|
+
|
|
745
|
+
// cuBLAS with beta=1: s_cur = B_weighted @ X_dt^T + 1.0 * s_cur (already scaled)
|
|
746
|
+
const int64_t stride_Bw = d_state * chunk_len;
|
|
747
|
+
const int64_t stride_X = chunk_len * head_dim;
|
|
748
|
+
const int64_t stride_S = state_per_head;
|
|
749
|
+
|
|
750
|
+
for (int64_t s = 0; s < n_seq; s++) {
|
|
751
|
+
// X_dt is d-fastest {hd, C}, read as OP_T to get {C, hd}
|
|
752
|
+
CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
|
|
753
|
+
d_state, head_dim, chunk_len,
|
|
754
|
+
&alpha_one,
|
|
755
|
+
B_weighted + s * stride_Bw * n_head, matmul_dtype, d_state, stride_Bw,
|
|
756
|
+
X_dt + s * stride_X * n_head, matmul_dtype, head_dim, stride_X,
|
|
757
|
+
&beta_one,
|
|
758
|
+
s_cur + s * stride_S * n_head, CUDA_R_32F, d_state, stride_S,
|
|
759
|
+
n_head,
|
|
760
|
+
CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
|
|
761
|
+
}
|
|
762
|
+
}
|
|
763
|
+
}
|
|
764
|
+
}
|
|
765
|
+
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
|
766
|
+
|
|
767
|
+
void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
|
768
|
+
const struct ggml_tensor * src0 = dst->src[0]; // s
|
|
769
|
+
const struct ggml_tensor * src1 = dst->src[1]; // x
|
|
770
|
+
const struct ggml_tensor * src2 = dst->src[2]; // dt
|
|
771
|
+
const struct ggml_tensor * src3 = dst->src[3]; // A
|
|
772
|
+
const struct ggml_tensor * src4 = dst->src[4]; // B
|
|
773
|
+
const struct ggml_tensor * src5 = dst->src[5]; // C
|
|
774
|
+
const struct ggml_tensor * src6 = dst->src[6]; // ids
|
|
775
|
+
|
|
776
|
+
const int64_t nc = src0->ne[0]; // d_state
|
|
777
|
+
const int64_t nr = src0->ne[1]; // head_dim or 1
|
|
778
|
+
const int64_t nh = src1->ne[1]; // n_head
|
|
779
|
+
const int64_t ng = src4->ne[1]; // n_group
|
|
780
|
+
const int64_t n_t = src1->ne[2]; // number of tokens per sequence
|
|
781
|
+
const int64_t n_s = src1->ne[3]; // number of sequences in the batch
|
|
782
|
+
const int32_t K_param = ggml_get_op_params_i32(dst, 0);
|
|
783
|
+
const int64_t K = K_param > 0 ? K_param : 1;
|
|
784
|
+
|
|
785
|
+
const int64_t s_off = ggml_nelements(src1) * sizeof(float);
|
|
786
|
+
|
|
787
|
+
GGML_ASSERT(ggml_nelements(src1) + K*nc*nr*nh*n_s == ggml_nelements(dst));
|
|
788
|
+
GGML_ASSERT(src0->nb[0] == sizeof(float));
|
|
789
|
+
GGML_ASSERT(src1->nb[0] == sizeof(float));
|
|
790
|
+
GGML_ASSERT(src2->nb[0] == sizeof(float));
|
|
791
|
+
GGML_ASSERT(src3->nb[0] == sizeof(float));
|
|
792
|
+
GGML_ASSERT(src4->nb[0] == sizeof(float));
|
|
793
|
+
GGML_ASSERT(src5->nb[0] == sizeof(float));
|
|
794
|
+
GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
|
|
795
|
+
GGML_ASSERT(src3->ne[0] == 1 || K == 1);
|
|
796
|
+
|
|
797
|
+
const float * src0_d = (const float *) src0->data;
|
|
798
|
+
const float * src1_d = (const float *) src1->data;
|
|
799
|
+
const float * src2_d = (const float *) src2->data;
|
|
800
|
+
const float * src3_d = (const float *) src3->data;
|
|
801
|
+
const float * src4_d = (const float *) src4->data;
|
|
802
|
+
const float * src5_d = (const float *) src5->data;
|
|
803
|
+
const int32_t * src6_d = (const int32_t *) src6->data;
|
|
804
|
+
float * dst_d = (float *) dst->data;
|
|
805
|
+
cudaStream_t stream = ctx.stream();
|
|
806
|
+
|
|
807
|
+
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
808
|
+
GGML_ASSERT(src6->type == GGML_TYPE_I32);
|
|
809
|
+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
810
|
+
|
|
811
|
+
// Byte strides are narrowed to int for both scan and SSD paths.
|
|
812
|
+
GGML_ASSERT(src0->nb[2] <= (size_t)INT_MAX);
|
|
813
|
+
GGML_ASSERT(src0->nb[3] <= (size_t)INT_MAX);
|
|
814
|
+
GGML_ASSERT(src1->nb[2] <= (size_t)INT_MAX);
|
|
815
|
+
GGML_ASSERT(src1->nb[3] <= (size_t)INT_MAX);
|
|
816
|
+
GGML_ASSERT(src2->nb[1] <= (size_t)INT_MAX);
|
|
817
|
+
GGML_ASSERT(src2->nb[2] <= (size_t)INT_MAX);
|
|
818
|
+
GGML_ASSERT(src3->nb[1] <= (size_t)INT_MAX);
|
|
819
|
+
GGML_ASSERT(src4->nb[2] <= (size_t)INT_MAX);
|
|
820
|
+
GGML_ASSERT(src4->nb[3] <= (size_t)INT_MAX);
|
|
821
|
+
GGML_ASSERT(src5->nb[2] <= (size_t)INT_MAX);
|
|
822
|
+
GGML_ASSERT(src5->nb[3] <= (size_t)INT_MAX);
|
|
823
|
+
|
|
824
|
+
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
|
825
|
+
// Mamba-2 with scalar A per head: use SSD matmul path for long sequences.
|
|
826
|
+
// Requires NVIDIA Turing+ otherwise fallback to scan.
|
|
827
|
+
const bool is_mamba2 = (src3->nb[1] == sizeof(float));
|
|
828
|
+
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
|
|
829
|
+
const bool use_ssd = is_mamba2 && n_t > SSM_SSD_MIN_TOKENS
|
|
830
|
+
&& K == 1
|
|
831
|
+
&& n_t <= SSM_SSD_MAX_TOKENS
|
|
832
|
+
&& GGML_CUDA_CC_IS_NVIDIA(cc)
|
|
833
|
+
&& cc >= GGML_CUDA_CC_TURING
|
|
834
|
+
&& nr % 8 == 0; // cuBLAS requires 8-element (16-byte) alignment
|
|
835
|
+
|
|
836
|
+
if (use_ssd) {
|
|
837
|
+
// ssm_ssd_init_state_kernel uses flat linear indexing within each sequence,
|
|
838
|
+
// so src0 must be fully contiguous across all inner dimensions.
|
|
839
|
+
// The scan path handles non-contiguous nb[2] via src0_nb2 but does not handle nb[1].
|
|
840
|
+
GGML_ASSERT(src0->nb[1] == nc * sizeof(float));
|
|
841
|
+
GGML_ASSERT(src0->nb[2] == nc * nr * sizeof(float));
|
|
842
|
+
|
|
843
|
+
ssm_scan_ssd_f32_cuda(ctx,
|
|
844
|
+
src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
|
|
845
|
+
(int64_t)(src0->nb[3] / sizeof(float)),
|
|
846
|
+
(int)(src1->nb[2] / sizeof(float)), (int)(src1->nb[3] / sizeof(float)),
|
|
847
|
+
(int)(src2->nb[1] / sizeof(float)), (int)(src2->nb[2] / sizeof(float)),
|
|
848
|
+
(int)(src3->nb[1] / sizeof(float)),
|
|
849
|
+
(int)(src4->nb[2] / sizeof(float)), (int)(src4->nb[3] / sizeof(float)),
|
|
850
|
+
(int)(src5->nb[2] / sizeof(float)), (int)(src5->nb[3] / sizeof(float)),
|
|
851
|
+
s_off, nc, nr, nh, ng, n_t, n_s);
|
|
852
|
+
return;
|
|
853
|
+
}
|
|
854
|
+
#endif
|
|
855
|
+
ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
|
|
856
|
+
src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
|
|
857
|
+
src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
|
|
858
|
+
s_off, nc, nr, nh, ng, n_t, n_s, K, stream);
|
|
859
|
+
}
|