whispercpp 1.3.5 → 1.3.6
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/LICENSE +1 -1
- data/README.md +99 -2
- data/ext/extconf.rb +1 -0
- data/ext/ruby_whisper.c +20 -4
- data/ext/ruby_whisper.h +30 -2
- data/ext/ruby_whisper_context.c +216 -124
- data/ext/ruby_whisper_context_params.c +163 -0
- data/ext/ruby_whisper_model.c +0 -1
- data/ext/ruby_whisper_params.c +0 -1
- data/ext/ruby_whisper_segment.c +0 -1
- data/ext/ruby_whisper_token.c +29 -9
- data/ext/ruby_whisper_transcribe.cpp +4 -1
- data/ext/ruby_whisper_vad_context.c +48 -1
- data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
- data/ext/ruby_whisper_vad_params.c +0 -1
- data/ext/ruby_whisper_vad_segment.c +0 -1
- data/ext/ruby_whisper_vad_segments.c +0 -1
- data/ext/sources/CMakeLists.txt +1 -1
- data/ext/sources/bindings/javascript/package.json +1 -1
- data/ext/sources/cmake/whisper-config.cmake.in +5 -40
- data/ext/sources/examples/bench/bench.cpp +23 -18
- data/ext/sources/examples/cli/cli.cpp +8 -0
- data/ext/sources/examples/common-ggml.cpp +2 -0
- data/ext/sources/examples/miniaudio.h +4507 -2131
- data/ext/sources/examples/server/server.cpp +18 -4
- data/ext/sources/examples/talk-llama/CMakeLists.txt +3 -2
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +7 -13
- data/ext/sources/examples/talk-llama/llama-adapter.h +4 -3
- data/ext/sources/examples/talk-llama/llama-arch.cpp +335 -17
- data/ext/sources/examples/talk-llama/llama-arch.h +42 -0
- data/ext/sources/examples/talk-llama/llama-batch.cpp +3 -1
- data/ext/sources/examples/talk-llama/llama-chat.cpp +21 -1
- data/ext/sources/examples/talk-llama/llama-chat.h +1 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +508 -520
- data/ext/sources/examples/talk-llama/llama-context.h +27 -28
- data/ext/sources/examples/talk-llama/llama-cparams.h +5 -0
- data/ext/sources/examples/talk-llama/llama-ext.h +12 -0
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +8 -8
- data/ext/sources/examples/talk-llama/llama-graph.cpp +583 -130
- data/ext/sources/examples/talk-llama/llama-graph.h +131 -10
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +57 -40
- data/ext/sources/examples/talk-llama/llama-hparams.h +79 -10
- data/ext/sources/examples/talk-llama/llama-impl.cpp +4 -4
- data/ext/sources/examples/talk-llama/llama-impl.h +13 -1
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +3 -1
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +274 -89
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +2 -3
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +275 -0
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +140 -0
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +11 -13
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +28 -11
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +527 -119
- data/ext/sources/examples/talk-llama/llama-model-loader.h +35 -5
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +60 -46
- data/ext/sources/examples/talk-llama/llama-model-saver.h +5 -2
- data/ext/sources/examples/talk-llama/llama-model.cpp +1365 -647
- data/ext/sources/examples/talk-llama/llama-model.h +72 -19
- data/ext/sources/examples/talk-llama/llama-quant.cpp +578 -346
- data/ext/sources/examples/talk-llama/{llama-sampling.cpp → llama-sampler.cpp} +190 -76
- data/ext/sources/examples/talk-llama/{llama-sampling.h → llama-sampler.h} +0 -2
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +118 -48
- data/ext/sources/examples/talk-llama/llama-vocab.h +5 -0
- data/ext/sources/examples/talk-llama/llama.cpp +76 -22
- data/ext/sources/examples/talk-llama/llama.h +63 -30
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +2 -3
- data/ext/sources/examples/talk-llama/models/apertus.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/arcee.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/arctic.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +4 -3
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +1 -2
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +3 -5
- data/ext/sources/examples/talk-llama/models/bert.cpp +13 -7
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +9 -24
- data/ext/sources/examples/talk-llama/models/bloom.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/chameleon.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/chatglm.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/codeshell.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/cogvlm.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/command-r.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/dbrx.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/deci.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +24 -21
- data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +445 -0
- data/ext/sources/examples/talk-llama/models/dots1.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/dream.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/eurobert.cpp +97 -0
- data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +145 -0
- data/ext/sources/examples/talk-llama/models/exaone.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/exaone4.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +2 -4
- data/ext/sources/examples/talk-llama/models/falcon.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma3.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +7 -7
- data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/glm4.cpp +14 -7
- data/ext/sources/examples/talk-llama/models/gpt2.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/gptneox.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/granite.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/grok.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +5 -7
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/internlm2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/jais.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/jais2.cpp +123 -0
- data/ext/sources/examples/talk-llama/models/jamba.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +381 -0
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +145 -124
- data/ext/sources/examples/talk-llama/models/llada-moe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/llada.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/llama.cpp +18 -11
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/{graph-context-mamba.cpp → mamba-base.cpp} +9 -3
- data/ext/sources/examples/talk-llama/models/mamba.cpp +1 -2
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +11 -5
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +14 -13
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/models.h +181 -46
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +2 -9
- data/ext/sources/examples/talk-llama/models/mpt.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +26 -14
- data/ext/sources/examples/talk-llama/models/nemotron.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/neo-bert.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/olmo.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/olmo2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/olmoe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/openelm.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/orion.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/paddleocr.cpp +122 -0
- data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/phi2.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/phi3.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/plamo.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/plamo2.cpp +9 -5
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/plm.cpp +15 -14
- data/ext/sources/examples/talk-llama/models/qwen.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/qwen2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/qwen3.cpp +12 -9
- data/ext/sources/examples/talk-llama/models/qwen35.cpp +381 -0
- data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +422 -0
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +15 -8
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +84 -432
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +9 -18
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +8 -17
- data/ext/sources/examples/talk-llama/models/refact.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/rnd1.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +2 -0
- data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +2 -0
- data/ext/sources/examples/talk-llama/models/seed-oss.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/smallthinker.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/smollm3.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/stablelm.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/starcoder.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/starcoder2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +165 -0
- data/ext/sources/examples/talk-llama/models/t5-dec.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/t5-enc.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/xverse.cpp +3 -3
- data/ext/sources/examples/talk-llama/unicode.cpp +21 -65
- data/ext/sources/ggml/CMakeLists.txt +9 -3
- data/ext/sources/ggml/include/ggml-backend.h +1 -1
- data/ext/sources/ggml/include/ggml-cann.h +1 -1
- data/ext/sources/ggml/include/ggml-cpu.h +5 -0
- data/ext/sources/ggml/include/ggml-openvino.h +37 -0
- data/ext/sources/ggml/include/ggml-opt.h +1 -1
- data/ext/sources/ggml/include/ggml-rpc.h +6 -1
- data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
- data/ext/sources/ggml/include/ggml.h +56 -9
- data/ext/sources/ggml/src/CMakeLists.txt +3 -0
- data/ext/sources/ggml/src/ggml-alloc.c +4 -9
- data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
- data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +28 -86
- data/ext/sources/ggml/src/ggml-backend.cpp +5 -2
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +6 -2
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +348 -189
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +40 -85
- data/ext/sources/ggml/src/ggml-cann/common.h +3 -4
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +44 -62
- data/ext/sources/ggml/src/ggml-common.h +11 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +16 -11
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +42 -19
- data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +85 -85
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +85 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2744 -548
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +1653 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1391 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -10
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +9 -9
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +118 -18
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +107 -26
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
- data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +59 -12
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +15 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +965 -252
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +584 -197
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +903 -188
- data/ext/sources/ggml/src/ggml-cpu/ops.h +1 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +40 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2890 -679
- data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +136 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +111 -3
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +17 -0
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +19 -10
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +32 -30
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +134 -18
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +41 -27
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +6 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +78 -64
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +384 -143
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +36 -22
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +3 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +26 -5
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +127 -12
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +263 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +595 -200
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +9 -8
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +173 -6
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +158 -85
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +34 -22
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +127 -67
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +157 -65
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +18 -76
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +2 -16
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +233 -133
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
- data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +56 -32
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +3 -3
- data/ext/sources/ggml/src/ggml-cuda/top-k.cu +0 -1
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +199 -135
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +55 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +31 -17
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +10 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +82 -45
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +334 -160
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +7 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +328 -197
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +765 -234
- data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +252 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +412 -265
- data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +23 -23
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +1 -1
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.h → hex-dma.h} +28 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +1 -1
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +27 -37
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +6 -35
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +240 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +251 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +20 -1347
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +211 -13
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +1119 -952
- data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +254 -244
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +36 -36
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +155 -138
- data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +339 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +209 -114
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
- data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +6 -0
- data/ext/sources/ggml/src/ggml-impl.h +62 -0
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +274 -73
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +22 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +102 -36
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +174 -23
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +580 -280
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +5 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +320 -107
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1068 -825
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +19 -1
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +3108 -636
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +41 -99
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +45 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +204 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
- data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
- data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
- data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +975 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +373 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1110 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +884 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +123 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +39 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +123 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +293 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +226 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +85 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +823 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.h +123 -0
- data/ext/sources/ggml/src/ggml-quants.c +96 -5
- data/ext/sources/ggml/src/ggml-quants.h +3 -0
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +15 -88
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +315 -10
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +69 -1
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +22 -1
- data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +78 -68
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1179 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +55 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1338 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +667 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +225 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +309 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +316 -51
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
- data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +13 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
- data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1250 -465
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +374 -170
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +66 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +389 -201
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +106 -58
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +128 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +8 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +3 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +36 -63
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +10 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +16 -10
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +55 -35
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1314 -109
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1660 -1371
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +141 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +65 -72
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +75 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +6 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +40 -5
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +105 -60
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +53 -259
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat.tmpl.wgsl → mul_mat.wgsl} +68 -257
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +692 -23
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_reg_tile.tmpl.wgsl → mul_mat_reg_tile.wgsl} +28 -128
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +31 -137
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +480 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +9 -36
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +40 -12
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +6 -1
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +9 -6
- data/ext/sources/ggml/src/ggml.c +167 -33
- data/ext/sources/ggml/src/gguf.cpp +229 -44
- data/ext/sources/src/whisper.cpp +6 -28
- data/sig/whisper.rbs +43 -2
- data/test/test_context_params.rb +82 -0
- data/test/test_token.rb +11 -0
- data/test/test_vad_context.rb +58 -8
- data/test/test_whisper.rb +20 -0
- data/whispercpp.gemspec +1 -1
- metadata +240 -28
- data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
- data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
data/ext/ruby_whisper_context.c
CHANGED
|
@@ -1,5 +1,3 @@
|
|
|
1
|
-
#include <ruby.h>
|
|
2
|
-
#include <ruby/memory_view.h>
|
|
3
1
|
#include "ruby_whisper.h"
|
|
4
2
|
|
|
5
3
|
extern ID id_to_s;
|
|
@@ -20,6 +18,7 @@ extern VALUE eError;
|
|
|
20
18
|
extern VALUE cModel;
|
|
21
19
|
|
|
22
20
|
extern const rb_data_type_t ruby_whisper_params_type;
|
|
21
|
+
extern const rb_data_type_t ruby_whisper_context_params_type;
|
|
23
22
|
extern VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self);
|
|
24
23
|
extern VALUE rb_whisper_model_s_new(VALUE context);
|
|
25
24
|
extern VALUE rb_whisper_segment_s_new(VALUE context, int index);
|
|
@@ -27,6 +26,27 @@ extern void prepare_transcription(ruby_whisper_params *rwp, VALUE *context);
|
|
|
27
26
|
|
|
28
27
|
ID transcribe_option_names[1];
|
|
29
28
|
|
|
29
|
+
typedef struct fill_samples_args {
|
|
30
|
+
float *dest;
|
|
31
|
+
VALUE *src;
|
|
32
|
+
int n_samples;
|
|
33
|
+
} fill_samples_args;
|
|
34
|
+
|
|
35
|
+
typedef struct full_args {
|
|
36
|
+
VALUE *context;
|
|
37
|
+
VALUE *params;
|
|
38
|
+
float *samples;
|
|
39
|
+
int n_samples;
|
|
40
|
+
} full_args;
|
|
41
|
+
|
|
42
|
+
typedef struct full_parallel_args {
|
|
43
|
+
VALUE *context;
|
|
44
|
+
VALUE *params;
|
|
45
|
+
float *samples;
|
|
46
|
+
int n_samples;
|
|
47
|
+
int n_processors;
|
|
48
|
+
} full_parallel_args;
|
|
49
|
+
|
|
30
50
|
static void
|
|
31
51
|
ruby_whisper_free(ruby_whisper *rw)
|
|
32
52
|
{
|
|
@@ -124,16 +144,25 @@ ruby_whisper_initialize(int argc, VALUE *argv, VALUE self)
|
|
|
124
144
|
{
|
|
125
145
|
ruby_whisper *rw;
|
|
126
146
|
VALUE whisper_model_file_path;
|
|
147
|
+
VALUE context_params;
|
|
148
|
+
struct whisper_context_params params;
|
|
127
149
|
|
|
128
150
|
// TODO: we can support init from buffer here too maybe another ruby object to expose
|
|
129
|
-
rb_scan_args(argc, argv, "
|
|
151
|
+
rb_scan_args(argc, argv, "11", &whisper_model_file_path, &context_params);
|
|
130
152
|
TypedData_Get_Struct(self, ruby_whisper, &ruby_whisper_type, rw);
|
|
131
153
|
|
|
132
154
|
whisper_model_file_path = ruby_whisper_normalize_model_path(whisper_model_file_path);
|
|
133
155
|
if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
|
|
134
156
|
rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
|
|
135
157
|
}
|
|
136
|
-
|
|
158
|
+
if (NIL_P(context_params)) {
|
|
159
|
+
params = whisper_context_default_params();
|
|
160
|
+
} else {
|
|
161
|
+
ruby_whisper_context_params *rwcp;
|
|
162
|
+
GetContextParams(context_params, rwcp);
|
|
163
|
+
params = rwcp->params;
|
|
164
|
+
}
|
|
165
|
+
rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), params);
|
|
137
166
|
if (rw->context == NULL) {
|
|
138
167
|
rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
|
|
139
168
|
}
|
|
@@ -272,82 +301,175 @@ VALUE ruby_whisper_model_type(VALUE self)
|
|
|
272
301
|
return rb_str_new2(whisper_model_type_readable(rw->context));
|
|
273
302
|
}
|
|
274
303
|
|
|
275
|
-
|
|
276
|
-
*
|
|
277
|
-
* Not thread safe for same context
|
|
278
|
-
* Uses the specified decoding strategy to obtain the text.
|
|
279
|
-
*
|
|
280
|
-
* call-seq:
|
|
281
|
-
* full(params, samples, n_samples) -> nil
|
|
282
|
-
* full(params, samples) -> nil
|
|
283
|
-
*
|
|
284
|
-
* The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
|
|
285
|
-
*/
|
|
286
|
-
VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
|
|
304
|
+
static bool
|
|
305
|
+
check_memory_view(rb_memory_view_t *memview)
|
|
287
306
|
{
|
|
288
|
-
if (
|
|
289
|
-
|
|
307
|
+
if (memview->format != NULL && strcmp(memview->format, "f") != 0) {
|
|
308
|
+
rb_warn("currently only format \"f\" is supported for MemoryView, but given: %s", memview->format);
|
|
309
|
+
return false;
|
|
310
|
+
}
|
|
311
|
+
if (memview->format != NULL && memview->ndim != 1) {
|
|
312
|
+
rb_warn("currently only 1 dimensional MemoryView is supported, but given: %zd", memview->ndim);
|
|
313
|
+
return false;
|
|
290
314
|
}
|
|
291
315
|
|
|
292
|
-
|
|
293
|
-
|
|
294
|
-
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
|
|
298
|
-
|
|
299
|
-
|
|
300
|
-
|
|
301
|
-
|
|
302
|
-
|
|
303
|
-
|
|
304
|
-
|
|
305
|
-
|
|
316
|
+
return true;
|
|
317
|
+
}
|
|
318
|
+
|
|
319
|
+
static VALUE
|
|
320
|
+
fill_samples(VALUE rb_args)
|
|
321
|
+
{
|
|
322
|
+
fill_samples_args *args = (fill_samples_args *)rb_args;
|
|
323
|
+
|
|
324
|
+
if (RB_TYPE_P(*args->src, T_ARRAY)) {
|
|
325
|
+
for (int i = 0; i < args->n_samples; i++) {
|
|
326
|
+
args->dest[i] = RFLOAT_VALUE(rb_ary_entry(*args->src, i));
|
|
327
|
+
}
|
|
328
|
+
} else {
|
|
329
|
+
// TODO: use rb_block_call
|
|
330
|
+
VALUE iter = rb_funcall(*args->src, id_to_enum, 1, rb_str_new2("each"));
|
|
331
|
+
for (int i = 0; i < args->n_samples; i++) {
|
|
332
|
+
// TODO: check if iter is exhausted and raise ArgumentError appropriately
|
|
333
|
+
VALUE sample = rb_funcall(iter, id_next, 0);
|
|
334
|
+
args->dest[i] = RFLOAT_VALUE(sample);
|
|
335
|
+
}
|
|
336
|
+
}
|
|
337
|
+
|
|
338
|
+
return Qnil;
|
|
339
|
+
}
|
|
340
|
+
|
|
341
|
+
struct parsed_samples_t
|
|
342
|
+
parse_samples(VALUE *samples, VALUE *n_samples)
|
|
343
|
+
{
|
|
344
|
+
bool memview_available = rb_memory_view_available_p(*samples);
|
|
345
|
+
struct parsed_samples_t parsed = {0};
|
|
346
|
+
parsed.memview_exported = false;
|
|
347
|
+
const bool is_array = RB_TYPE_P(*samples, T_ARRAY);
|
|
348
|
+
|
|
349
|
+
if (!NIL_P(*n_samples)) {
|
|
350
|
+
parsed.n_samples = NUM2INT(*n_samples);
|
|
351
|
+
if (is_array) {
|
|
352
|
+
if (RARRAY_LEN(*samples) < parsed.n_samples) {
|
|
353
|
+
rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(*samples), parsed.n_samples);
|
|
306
354
|
}
|
|
307
355
|
}
|
|
308
356
|
// Should check when samples.respond_to?(:length)?
|
|
309
357
|
} else {
|
|
310
|
-
if (
|
|
311
|
-
if (RARRAY_LEN(samples) > INT_MAX) {
|
|
358
|
+
if (is_array) {
|
|
359
|
+
if (RARRAY_LEN(*samples) > INT_MAX) {
|
|
312
360
|
rb_raise(rb_eArgError, "samples are too long");
|
|
313
361
|
}
|
|
314
|
-
n_samples = (int)RARRAY_LEN(samples);
|
|
315
|
-
} else if (
|
|
316
|
-
|
|
317
|
-
|
|
318
|
-
|
|
362
|
+
parsed.n_samples = (int)RARRAY_LEN(*samples);
|
|
363
|
+
} else if (memview_available) {
|
|
364
|
+
bool memview_got = rb_memory_view_get(*samples, &parsed.memview, RUBY_MEMORY_VIEW_SIMPLE);
|
|
365
|
+
if (memview_got) {
|
|
366
|
+
parsed.memview_exported = check_memory_view(&parsed.memview);
|
|
367
|
+
if (!parsed.memview_exported) {
|
|
368
|
+
rb_memory_view_release(&parsed.memview);
|
|
369
|
+
parsed.memview = (rb_memory_view_t){0};
|
|
370
|
+
}
|
|
319
371
|
}
|
|
320
|
-
|
|
321
|
-
|
|
322
|
-
|
|
372
|
+
if (parsed.memview_exported) {
|
|
373
|
+
ssize_t n_samples_size = parsed.memview.byte_size / parsed.memview.item_size;
|
|
374
|
+
if (n_samples_size > INT_MAX) {
|
|
375
|
+
rb_memory_view_release(&parsed.memview);
|
|
376
|
+
rb_raise(rb_eArgError, "samples are too long: %zd", n_samples_size);
|
|
377
|
+
}
|
|
378
|
+
parsed.n_samples = (int)n_samples_size;
|
|
379
|
+
} else {
|
|
380
|
+
rb_warn("unable to get a memory view. falls back to Ruby object");
|
|
381
|
+
if (rb_respond_to(*samples, id_length)) {
|
|
382
|
+
parsed.n_samples = NUM2INT(rb_funcall(*samples, id_length, 0));
|
|
383
|
+
} else {
|
|
384
|
+
rb_raise(rb_eArgError, "samples must respond to :length");
|
|
385
|
+
}
|
|
323
386
|
}
|
|
324
|
-
|
|
325
|
-
|
|
326
|
-
n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
|
|
387
|
+
} else if (rb_respond_to(*samples, id_length)) {
|
|
388
|
+
parsed.n_samples = NUM2INT(rb_funcall(*samples, id_length, 0));
|
|
327
389
|
} else {
|
|
328
|
-
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of
|
|
390
|
+
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of float when n_samples is not given");
|
|
329
391
|
}
|
|
330
392
|
}
|
|
331
|
-
|
|
332
|
-
if (
|
|
333
|
-
|
|
393
|
+
|
|
394
|
+
if (parsed.memview_exported) {
|
|
395
|
+
parsed.samples = (float *)parsed.memview.data;
|
|
334
396
|
} else {
|
|
335
|
-
|
|
336
|
-
|
|
337
|
-
|
|
338
|
-
|
|
339
|
-
|
|
340
|
-
|
|
341
|
-
|
|
342
|
-
|
|
343
|
-
|
|
344
|
-
|
|
345
|
-
|
|
346
|
-
}
|
|
397
|
+
parsed.samples = ALLOC_N(float, parsed.n_samples);
|
|
398
|
+
fill_samples_args args = {
|
|
399
|
+
parsed.samples,
|
|
400
|
+
samples,
|
|
401
|
+
parsed.n_samples,
|
|
402
|
+
};
|
|
403
|
+
int state;
|
|
404
|
+
rb_protect(fill_samples, (VALUE)&args, &state);
|
|
405
|
+
if (state) {
|
|
406
|
+
xfree(parsed.samples);
|
|
407
|
+
rb_jump_tag(state);
|
|
347
408
|
}
|
|
348
409
|
}
|
|
349
|
-
|
|
350
|
-
|
|
410
|
+
|
|
411
|
+
return parsed;
|
|
412
|
+
}
|
|
413
|
+
|
|
414
|
+
VALUE
|
|
415
|
+
release_samples(VALUE rb_parsed_args)
|
|
416
|
+
{
|
|
417
|
+
parsed_samples_t *parsed_args = (parsed_samples_t *)rb_parsed_args;
|
|
418
|
+
|
|
419
|
+
if (parsed_args->memview_exported) {
|
|
420
|
+
rb_memory_view_release(&parsed_args->memview);
|
|
421
|
+
} else {
|
|
422
|
+
xfree(parsed_args->samples);
|
|
423
|
+
}
|
|
424
|
+
*parsed_args = (parsed_samples_t){0};
|
|
425
|
+
|
|
426
|
+
return Qnil;
|
|
427
|
+
}
|
|
428
|
+
|
|
429
|
+
static VALUE
|
|
430
|
+
full_body(VALUE rb_args)
|
|
431
|
+
{
|
|
432
|
+
full_args *args = (full_args *)rb_args;
|
|
433
|
+
|
|
434
|
+
ruby_whisper *rw;
|
|
435
|
+
ruby_whisper_params *rwp;
|
|
436
|
+
GetContext(*args->context, rw);
|
|
437
|
+
TypedData_Get_Struct(*args->params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
|
|
438
|
+
|
|
439
|
+
prepare_transcription(rwp, args->context);
|
|
440
|
+
int result = whisper_full(rw->context, rwp->params, args->samples, args->n_samples);
|
|
441
|
+
|
|
442
|
+
return INT2NUM(result);
|
|
443
|
+
}
|
|
444
|
+
|
|
445
|
+
/*
|
|
446
|
+
* Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
|
|
447
|
+
* Not thread safe for same context
|
|
448
|
+
* Uses the specified decoding strategy to obtain the text.
|
|
449
|
+
*
|
|
450
|
+
* call-seq:
|
|
451
|
+
* full(params, samples, n_samples) -> nil
|
|
452
|
+
* full(params, samples) -> nil
|
|
453
|
+
*
|
|
454
|
+
* The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
|
|
455
|
+
*/
|
|
456
|
+
VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
|
|
457
|
+
{
|
|
458
|
+
if (argc < 2 || argc > 3) {
|
|
459
|
+
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
|
|
460
|
+
}
|
|
461
|
+
|
|
462
|
+
VALUE n_samples = argc == 2 ? Qnil : argv[2];
|
|
463
|
+
|
|
464
|
+
struct parsed_samples_t parsed = parse_samples(&argv[1], &n_samples);
|
|
465
|
+
full_args args = {
|
|
466
|
+
&self,
|
|
467
|
+
&argv[0],
|
|
468
|
+
parsed.samples,
|
|
469
|
+
parsed.n_samples,
|
|
470
|
+
};
|
|
471
|
+
VALUE rb_result = rb_ensure(full_body, (VALUE)&args, release_samples, (VALUE)&parsed);
|
|
472
|
+
const int result = NUM2INT(rb_result);
|
|
351
473
|
if (0 == result) {
|
|
352
474
|
return self;
|
|
353
475
|
} else {
|
|
@@ -355,6 +477,22 @@ VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
|
|
|
355
477
|
}
|
|
356
478
|
}
|
|
357
479
|
|
|
480
|
+
static VALUE
|
|
481
|
+
full_parallel_body(VALUE rb_args)
|
|
482
|
+
{
|
|
483
|
+
full_parallel_args *args = (full_parallel_args *)rb_args;
|
|
484
|
+
|
|
485
|
+
ruby_whisper *rw;
|
|
486
|
+
ruby_whisper_params *rwp;
|
|
487
|
+
GetContext(*args->context, rw);
|
|
488
|
+
TypedData_Get_Struct(*args->params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
|
|
489
|
+
|
|
490
|
+
prepare_transcription(rwp, args->context);
|
|
491
|
+
int result = whisper_full_parallel(rw->context, rwp->params, args->samples, args->n_samples, args->n_processors);
|
|
492
|
+
|
|
493
|
+
return INT2NUM(result);
|
|
494
|
+
}
|
|
495
|
+
|
|
358
496
|
/*
|
|
359
497
|
* Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
|
|
360
498
|
* Result is stored in the default state of the context
|
|
@@ -372,19 +510,11 @@ static VALUE
|
|
|
372
510
|
ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self)
|
|
373
511
|
{
|
|
374
512
|
if (argc < 2 || argc > 4) {
|
|
375
|
-
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..
|
|
513
|
+
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..4)", argc);
|
|
376
514
|
}
|
|
377
515
|
|
|
378
|
-
|
|
379
|
-
ruby_whisper_params *rwp;
|
|
380
|
-
GetContext(self, rw);
|
|
381
|
-
VALUE params = argv[0];
|
|
382
|
-
TypedData_Get_Struct(params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
|
|
383
|
-
VALUE samples = argv[1];
|
|
384
|
-
int n_samples;
|
|
516
|
+
VALUE n_samples = argc == 2 ? Qnil : argv[2];
|
|
385
517
|
int n_processors;
|
|
386
|
-
rb_memory_view_t view;
|
|
387
|
-
const bool memory_view_available_p = rb_memory_view_available_p(samples);
|
|
388
518
|
switch (argc) {
|
|
389
519
|
case 2:
|
|
390
520
|
n_processors = 1;
|
|
@@ -396,56 +526,16 @@ ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self)
|
|
|
396
526
|
n_processors = NUM2INT(argv[3]);
|
|
397
527
|
break;
|
|
398
528
|
}
|
|
399
|
-
|
|
400
|
-
|
|
401
|
-
|
|
402
|
-
|
|
403
|
-
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
|
|
407
|
-
|
|
408
|
-
|
|
409
|
-
view.obj = Qnil;
|
|
410
|
-
rb_raise(rb_eArgError, "unable to get a memory view");
|
|
411
|
-
}
|
|
412
|
-
ssize_t n_samples_size = view.byte_size / view.item_size;
|
|
413
|
-
if (n_samples_size > INT_MAX) {
|
|
414
|
-
rb_raise(rb_eArgError, "samples are too long");
|
|
415
|
-
}
|
|
416
|
-
n_samples = (int)n_samples_size;
|
|
417
|
-
} else {
|
|
418
|
-
if (TYPE(samples) == T_ARRAY) {
|
|
419
|
-
if (RARRAY_LEN(samples) > INT_MAX) {
|
|
420
|
-
rb_raise(rb_eArgError, "samples are too long");
|
|
421
|
-
}
|
|
422
|
-
n_samples = (int)RARRAY_LEN(samples);
|
|
423
|
-
} else if (rb_respond_to(samples, id_length)) {
|
|
424
|
-
n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
|
|
425
|
-
} else {
|
|
426
|
-
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
|
|
427
|
-
}
|
|
428
|
-
}
|
|
429
|
-
float * c_samples = (float *)malloc(n_samples * sizeof(float));
|
|
430
|
-
if (memory_view_available_p) {
|
|
431
|
-
c_samples = (float *)view.data;
|
|
432
|
-
} else {
|
|
433
|
-
if (TYPE(samples) == T_ARRAY) {
|
|
434
|
-
for (int i = 0; i < n_samples; i++) {
|
|
435
|
-
c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
|
|
436
|
-
}
|
|
437
|
-
} else {
|
|
438
|
-
// FIXME: use rb_block_call
|
|
439
|
-
VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
|
|
440
|
-
for (int i = 0; i < n_samples; i++) {
|
|
441
|
-
// TODO: check if iter is exhausted and raise ArgumentError
|
|
442
|
-
VALUE sample = rb_funcall(iter, id_next, 0);
|
|
443
|
-
c_samples[i] = RFLOAT_VALUE(sample);
|
|
444
|
-
}
|
|
445
|
-
}
|
|
446
|
-
}
|
|
447
|
-
prepare_transcription(rwp, &self);
|
|
448
|
-
const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
|
|
529
|
+
struct parsed_samples_t parsed = parse_samples(&argv[1], &n_samples);
|
|
530
|
+
const full_parallel_args args = {
|
|
531
|
+
&self,
|
|
532
|
+
&argv[0],
|
|
533
|
+
parsed.samples,
|
|
534
|
+
parsed.n_samples,
|
|
535
|
+
n_processors,
|
|
536
|
+
};
|
|
537
|
+
const VALUE rb_result = rb_ensure(full_parallel_body, (VALUE)&args, release_samples, (VALUE)&parsed);
|
|
538
|
+
const int result = NUM2INT(rb_result);
|
|
449
539
|
if (0 == result) {
|
|
450
540
|
return self;
|
|
451
541
|
} else {
|
|
@@ -631,7 +721,7 @@ ruby_whisper_get_model(VALUE self)
|
|
|
631
721
|
return rb_whisper_model_s_new(self);
|
|
632
722
|
}
|
|
633
723
|
|
|
634
|
-
|
|
724
|
+
VALUE
|
|
635
725
|
init_ruby_whisper_context(VALUE *mWhisper)
|
|
636
726
|
{
|
|
637
727
|
cContext = rb_define_class_under(*mWhisper, "Context", rb_cObject);
|
|
@@ -669,4 +759,6 @@ init_ruby_whisper_context(VALUE *mWhisper)
|
|
|
669
759
|
rb_define_method(cContext, "each_segment", ruby_whisper_each_segment, 0);
|
|
670
760
|
|
|
671
761
|
rb_define_method(cContext, "model", ruby_whisper_get_model, 0);
|
|
762
|
+
|
|
763
|
+
return cContext;
|
|
672
764
|
}
|
|
@@ -0,0 +1,163 @@
|
|
|
1
|
+
#include "ruby_whisper.h"
|
|
2
|
+
|
|
3
|
+
#define NUM_PARAMS 6
|
|
4
|
+
|
|
5
|
+
#define DEF_BOOLEAN_ATTR_METHOD(name) \
|
|
6
|
+
static VALUE \
|
|
7
|
+
ruby_whisper_context_params_get_ ## name(VALUE self) { \
|
|
8
|
+
ruby_whisper_context_params *rwcp; \
|
|
9
|
+
GetContextParams(self, rwcp); \
|
|
10
|
+
return rwcp->params.name ? Qtrue : Qfalse; \
|
|
11
|
+
} \
|
|
12
|
+
static VALUE \
|
|
13
|
+
ruby_whisper_context_params_set_ ## name(VALUE self, VALUE value) { \
|
|
14
|
+
ruby_whisper_context_params *rwcp; \
|
|
15
|
+
GetContextParams(self, rwcp); \
|
|
16
|
+
rwcp->params.name = RTEST(value); \
|
|
17
|
+
return value; \
|
|
18
|
+
}
|
|
19
|
+
|
|
20
|
+
#define DEF_INT_ATTR_METHOD(name) \
|
|
21
|
+
static VALUE \
|
|
22
|
+
ruby_whisper_context_params_get_ ## name(VALUE self) { \
|
|
23
|
+
ruby_whisper_context_params *rwcp; \
|
|
24
|
+
GetContextParams(self, rwcp); \
|
|
25
|
+
return INT2NUM(rwcp->params.name); \
|
|
26
|
+
} \
|
|
27
|
+
static VALUE \
|
|
28
|
+
ruby_whisper_context_params_set_ ## name(VALUE self, VALUE value) { \
|
|
29
|
+
ruby_whisper_context_params *rwcp; \
|
|
30
|
+
GetContextParams(self, rwcp); \
|
|
31
|
+
rwcp->params.name = NUM2INT(value); \
|
|
32
|
+
return value; \
|
|
33
|
+
}
|
|
34
|
+
|
|
35
|
+
#define DEFINE_PARAM(param_name, nth) \
|
|
36
|
+
id_ ## param_name = rb_intern(#param_name); \
|
|
37
|
+
param_names[nth] = id_ ## param_name; \
|
|
38
|
+
rb_define_method(cContextParams, #param_name, ruby_whisper_context_params_get_ ## param_name, 0); \
|
|
39
|
+
rb_define_method(cContextParams, #param_name "=", ruby_whisper_context_params_set_ ## param_name, 1);
|
|
40
|
+
|
|
41
|
+
VALUE cContextParams;
|
|
42
|
+
|
|
43
|
+
static ID param_names[NUM_PARAMS];
|
|
44
|
+
static ID id_use_gpu;
|
|
45
|
+
static ID id_flash_attn;
|
|
46
|
+
static ID id_gpu_device;
|
|
47
|
+
static ID id_dtw_token_timestamps;
|
|
48
|
+
static ID id_dtw_aheads_preset;
|
|
49
|
+
static ID id_dtw_n_top;
|
|
50
|
+
|
|
51
|
+
static size_t
|
|
52
|
+
ruby_whisper_context_params_memsize(const void *p)
|
|
53
|
+
{
|
|
54
|
+
const ruby_whisper_context_params *rwcp = (ruby_whisper_context_params *)p;
|
|
55
|
+
if (!rwcp) {
|
|
56
|
+
return 0;
|
|
57
|
+
}
|
|
58
|
+
return sizeof(ruby_whisper_context_params);
|
|
59
|
+
}
|
|
60
|
+
|
|
61
|
+
const rb_data_type_t ruby_whisper_context_params_type = {
|
|
62
|
+
"ruby_whisper_context_params",
|
|
63
|
+
{0, RUBY_DEFAULT_FREE, ruby_whisper_context_params_memsize,},
|
|
64
|
+
0, 0,
|
|
65
|
+
0
|
|
66
|
+
};
|
|
67
|
+
|
|
68
|
+
static VALUE
|
|
69
|
+
ruby_whisper_context_params_s_allocate(VALUE klass)
|
|
70
|
+
{
|
|
71
|
+
ruby_whisper_context_params *rwcp;
|
|
72
|
+
return TypedData_Make_Struct(klass, ruby_whisper_context_params, &ruby_whisper_context_params_type, rwcp);
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
DEF_BOOLEAN_ATTR_METHOD(use_gpu);
|
|
76
|
+
DEF_BOOLEAN_ATTR_METHOD(flash_attn);
|
|
77
|
+
DEF_INT_ATTR_METHOD(gpu_device);
|
|
78
|
+
DEF_BOOLEAN_ATTR_METHOD(dtw_token_timestamps);
|
|
79
|
+
DEF_INT_ATTR_METHOD(dtw_aheads_preset);
|
|
80
|
+
|
|
81
|
+
static VALUE
|
|
82
|
+
ruby_whisper_context_params_get_dtw_n_top(VALUE self) {
|
|
83
|
+
ruby_whisper_context_params *rwcp;
|
|
84
|
+
GetContextParams(self, rwcp);
|
|
85
|
+
|
|
86
|
+
int dtw_n_top = rwcp->params.dtw_n_top;
|
|
87
|
+
|
|
88
|
+
return dtw_n_top == -1 ? Qnil : INT2NUM(dtw_n_top);
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
static VALUE
|
|
92
|
+
ruby_whisper_context_params_set_dtw_n_top(VALUE self, VALUE value) {
|
|
93
|
+
ruby_whisper_context_params *rwcp;
|
|
94
|
+
GetContextParams(self, rwcp);
|
|
95
|
+
|
|
96
|
+
rwcp->params.dtw_n_top = NIL_P(value) ? -1 : NUM2INT(value);
|
|
97
|
+
|
|
98
|
+
return value;
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
#define SET_PARAM_IF_SAME(param_name) \
|
|
102
|
+
if (id == id_ ## param_name) { \
|
|
103
|
+
ruby_whisper_context_params_set_ ## param_name(self, value); \
|
|
104
|
+
continue; \
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
static VALUE
|
|
108
|
+
ruby_whisper_context_params_initialize(int argc, VALUE *argv, VALUE self)
|
|
109
|
+
{
|
|
110
|
+
ruby_whisper_context_params *rwcp;
|
|
111
|
+
TypedData_Get_Struct(self, ruby_whisper_context_params, &ruby_whisper_context_params_type, rwcp);
|
|
112
|
+
rwcp->params = whisper_context_default_params();
|
|
113
|
+
|
|
114
|
+
VALUE kw_hash;
|
|
115
|
+
rb_scan_args_kw(RB_SCAN_ARGS_KEYWORDS, argc, argv, ":", &kw_hash);
|
|
116
|
+
if (NIL_P(kw_hash)) {
|
|
117
|
+
return Qnil;
|
|
118
|
+
}
|
|
119
|
+
|
|
120
|
+
VALUE values[NUM_PARAMS] = {Qundef};
|
|
121
|
+
rb_get_kwargs(kw_hash, param_names, 0, NUM_PARAMS, values);
|
|
122
|
+
|
|
123
|
+
ID id;
|
|
124
|
+
VALUE value;
|
|
125
|
+
for (int i = 0; i < NUM_PARAMS; i++) {
|
|
126
|
+
id = param_names[i];
|
|
127
|
+
value = values[i];
|
|
128
|
+
if (value == Qundef) {
|
|
129
|
+
continue;
|
|
130
|
+
}
|
|
131
|
+
SET_PARAM_IF_SAME(use_gpu)
|
|
132
|
+
SET_PARAM_IF_SAME(flash_attn)
|
|
133
|
+
SET_PARAM_IF_SAME(gpu_device)
|
|
134
|
+
SET_PARAM_IF_SAME(dtw_token_timestamps)
|
|
135
|
+
SET_PARAM_IF_SAME(dtw_aheads_preset)
|
|
136
|
+
SET_PARAM_IF_SAME(dtw_n_top)
|
|
137
|
+
}
|
|
138
|
+
|
|
139
|
+
return Qnil;
|
|
140
|
+
}
|
|
141
|
+
|
|
142
|
+
#undef SET_PARAM_IF_SAME
|
|
143
|
+
|
|
144
|
+
void
|
|
145
|
+
init_ruby_whisper_context_params(VALUE *cContext)
|
|
146
|
+
{
|
|
147
|
+
cContextParams = rb_define_class_under(*cContext, "Params", rb_cObject);
|
|
148
|
+
|
|
149
|
+
rb_define_alloc_func(cContextParams, ruby_whisper_context_params_s_allocate);
|
|
150
|
+
rb_define_method(cContextParams, "initialize", ruby_whisper_context_params_initialize, -1);
|
|
151
|
+
|
|
152
|
+
DEFINE_PARAM(use_gpu, 0)
|
|
153
|
+
DEFINE_PARAM(flash_attn, 1)
|
|
154
|
+
DEFINE_PARAM(gpu_device, 2)
|
|
155
|
+
DEFINE_PARAM(dtw_token_timestamps, 3)
|
|
156
|
+
DEFINE_PARAM(dtw_aheads_preset, 4)
|
|
157
|
+
DEFINE_PARAM(dtw_n_top, 5)
|
|
158
|
+
}
|
|
159
|
+
|
|
160
|
+
#undef DEFINE_PARAM
|
|
161
|
+
#undef DEF_INT_ATTR_METHOD
|
|
162
|
+
#undef DEF_BOOLEAN_ATTR_METHOD
|
|
163
|
+
#undef NUM_PARAMS
|
data/ext/ruby_whisper_model.c
CHANGED
data/ext/ruby_whisper_params.c
CHANGED
data/ext/ruby_whisper_segment.c
CHANGED
data/ext/ruby_whisper_token.c
CHANGED
|
@@ -1,4 +1,3 @@
|
|
|
1
|
-
#include <ruby.h>
|
|
2
1
|
#include "ruby_whisper.h"
|
|
3
2
|
|
|
4
3
|
#define N_KEY_NAMES 11
|
|
@@ -25,12 +24,34 @@ ruby_whisper_token_memsize(const void *p)
|
|
|
25
24
|
if (!rwt) {
|
|
26
25
|
return 0;
|
|
27
26
|
}
|
|
28
|
-
|
|
27
|
+
size_t size = sizeof(*rwt);
|
|
28
|
+
if (rwt->token_data) {
|
|
29
|
+
size += sizeof(*rwt->token_data);
|
|
30
|
+
}
|
|
31
|
+
return size;
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
static void
|
|
35
|
+
ruby_whisper_token_mark(void *p)
|
|
36
|
+
{
|
|
37
|
+
ruby_whisper_token *rwt = (ruby_whisper_token *)p;
|
|
38
|
+
rb_gc_mark(rwt->text);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
static void
|
|
42
|
+
ruby_whisper_token_free(void *p)
|
|
43
|
+
{
|
|
44
|
+
ruby_whisper_token *rwt = (ruby_whisper_token *)p;
|
|
45
|
+
if (rwt->token_data) {
|
|
46
|
+
xfree(rwt->token_data);
|
|
47
|
+
rwt->token_data = NULL;
|
|
48
|
+
}
|
|
49
|
+
xfree(rwt);
|
|
29
50
|
}
|
|
30
51
|
|
|
31
52
|
static const rb_data_type_t ruby_whisper_token_type = {
|
|
32
53
|
"ruby_whisper_token",
|
|
33
|
-
{
|
|
54
|
+
{ruby_whisper_token_mark, ruby_whisper_token_free, ruby_whisper_token_memsize,},
|
|
34
55
|
0, 0,
|
|
35
56
|
0
|
|
36
57
|
};
|
|
@@ -41,19 +62,19 @@ ruby_whisper_token_allocate(VALUE klass)
|
|
|
41
62
|
ruby_whisper_token *rwt;
|
|
42
63
|
VALUE token = TypedData_Make_Struct(klass, ruby_whisper_token, &ruby_whisper_token_type, rwt);
|
|
43
64
|
rwt->token_data = NULL;
|
|
44
|
-
rwt->text =
|
|
65
|
+
rwt->text = Qnil;
|
|
45
66
|
return token;
|
|
46
67
|
}
|
|
47
68
|
|
|
48
69
|
VALUE
|
|
49
70
|
ruby_whisper_token_s_init(struct whisper_context *context, int i_segment, int i_token)
|
|
50
71
|
{
|
|
51
|
-
whisper_token_data token_data = whisper_full_get_token_data(context, i_segment, i_token);
|
|
52
72
|
const VALUE token = ruby_whisper_token_allocate(cToken);
|
|
53
73
|
ruby_whisper_token *rwt;
|
|
54
74
|
TypedData_Get_Struct(token, ruby_whisper_token, &ruby_whisper_token_type, rwt);
|
|
55
|
-
rwt->token_data =
|
|
56
|
-
rwt->
|
|
75
|
+
rwt->token_data = ALLOC(whisper_token_data);
|
|
76
|
+
*(rwt->token_data) = whisper_full_get_token_data(context, i_segment, i_token);
|
|
77
|
+
rwt->text = rb_str_new2(whisper_full_get_token_text(context, i_segment, i_token));
|
|
57
78
|
return token;
|
|
58
79
|
}
|
|
59
80
|
|
|
@@ -183,10 +204,9 @@ ruby_whisper_token_get_text(VALUE self)
|
|
|
183
204
|
{
|
|
184
205
|
ruby_whisper_token *rwt;
|
|
185
206
|
GetToken(self, rwt);
|
|
186
|
-
return
|
|
207
|
+
return rwt->text;
|
|
187
208
|
}
|
|
188
209
|
|
|
189
|
-
|
|
190
210
|
/*
|
|
191
211
|
* Start time of the token.
|
|
192
212
|
*
|