diff --git a/CMakeLists.txt b/CMakeLists.txt index 6102364f2..b3f782aaf 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -519,6 +519,8 @@ add_library(sdtype_adapter otherarch/sdcpp/src/upscaler.cpp otherarch/sdcpp/src/runtime/guidance.cpp otherarch/sdcpp/src/runtime/guidance.h + otherarch/sdcpp/src/runtime/imatrix.cpp + otherarch/sdcpp/src/runtime/imatrix.h otherarch/sdcpp/src/stable-diffusion.cpp otherarch/sdcpp/thirdparty/zip.c otherarch/sdcpp/src/model_io/gguf_io.cpp @@ -535,6 +537,8 @@ add_library(sdtype_adapter otherarch/sdcpp/src/tokenizers/gpt_oss_tokenizer.cpp otherarch/sdcpp/src/tokenizers/tokenizer.cpp otherarch/sdcpp/src/tokenizers/tokenize_util.cpp + otherarch/sdcpp/src/core/backend_fit.cpp + otherarch/sdcpp/src/core/layer_split_partition.cpp otherarch/sdcpp/src/core/ggml_extend_backend.cpp) target_include_directories(sdtype_adapter PUBLIC . ./ggml/include ./ggml/src ./ggml/src/ggml-cpu ./include ./otherarch ./otherarch/tools ./vendor/stb ./vendor/nlohmann ./vendor ./otherarch/sdcpp ./otherarch/sdcpp/include ./otherarch/sdcpp/src ./otherarch/sdcpp/examples ./tools ./common) target_compile_features(sdtype_adapter PUBLIC cxx_std_17) # don't bump @@ -570,7 +574,10 @@ target_link_libraries(embeddings_adapter PRIVATE common2 ggml ${LLAMA_EXTRA_LIBS set_target_properties(embeddings_adapter PROPERTIES POSITION_INDEPENDENT_CODE ON) add_library(gpttype_adapter - gpttype_adapter.cpp) + gpttype_adapter.cpp + src/llama.cpp + common/chat.cpp + src/llama-model.cpp) target_include_directories(gpttype_adapter PUBLIC . ./src ./ggml/include ./ggml/src ./ggml/src/ggml-cpu ./include ./otherarch ./otherarch/tools ./vendor/stb ./vendor/nlohmann ./vendor ./otherarch/sdcpp ./otherarch/sdcpp/thirdparty ./tools ./common) target_compile_features(gpttype_adapter PUBLIC cxx_std_17) # don't bump target_link_libraries(gpttype_adapter PRIVATE common2 ggml ggml_v1 ggml_v2 ggml_v3 ${LLAMA_EXTRA_LIBS}) diff --git a/Makefile b/Makefile index d0d2f04a7..717c81de9 100644 --- a/Makefile +++ b/Makefile @@ -101,9 +101,9 @@ NONECFLAGS = LLAMA_USE_BUNDLED_GLSLC := 1 FAILSAFE_FLAGS = -DUSE_FAILSAFE -VULKAN_FLAGS = -DGGML_USE_VULKAN -DSD_USE_VULKAN +VULKAN_FLAGS = -DGGML_USE_VULKAN ifdef LLAMA_CUBLAS -CUBLAS_FLAGS = -DGGML_USE_CUDA -DSD_USE_CUDA +CUBLAS_FLAGS = -DGGML_USE_CUDA else CUBLAS_FLAGS = endif @@ -215,7 +215,7 @@ OBJS_CUDA_TEMP_INST += \ ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.o ifdef LLAMA_CUBLAS -CUBLAS_FLAGS = -DGGML_USE_CUDA -DSD_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include +CUBLAS_FLAGS = -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include CUBLASLD_FLAGS = -lcuda -lcublas -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/local/cuda/targets/sbsa-linux/lib -L/usr/lib/wsl/lib CUBLAS_OBJS = ggml-cuda.o ggml_v3-cuda.o ggml_v2-cuda.o ggml_v2-cuda-legacy.o CUBLAS_OBJS += $(patsubst %.cu,%.o,$(filter-out ggml/src/ggml-cuda/ggml-cuda.cu, $(wildcard ggml/src/ggml-cuda/*.cu))) @@ -315,7 +315,7 @@ HIPFLAGS += -DGGML_HIP_NO_ROCWMMA_FATTN endif endif -HIPFLAGS += -DGGML_USE_HIP -DGGML_HIP_NO_VMM -DGGML_USE_CUDA -DSD_USE_CUDA $(shell $(ROCM_PATH)/bin/hipconfig -C) +HIPFLAGS += -DGGML_USE_HIP -DGGML_HIP_NO_VMM -DGGML_USE_CUDA $(shell $(ROCM_PATH)/bin/hipconfig -C) HIPLDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib HIPLDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64 HIPLDFLAGS += -lhipblas -lamdhip64 -lrocblas @@ -339,8 +339,8 @@ endif # LLAMA_HIPBLAS ifdef LLAMA_METAL -CFLAGS += -DGGML_USE_METAL -DGGML_METAL_NDEBUG -DSD_USE_METAL -CXXFLAGS += -DGGML_USE_METAL -DSD_USE_METAL +CFLAGS += -DGGML_USE_METAL -DGGML_METAL_NDEBUG +CXXFLAGS += -DGGML_USE_METAL LDFLAGS += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders OBJS += ggml-metal.o ggml-metal-device.o ggml-metal-device-m.o ggml-metal-context-m.o ggml-metal-common.o ggml-metal-ops.o @@ -682,7 +682,9 @@ ggml-vulkan-shaders-noext.o: ggml/src/ggml-vulkan-shaders-noext.cpp ggml/include $(CXX) $(CXXFLAGS) $(VKGEN_NOEXT_FORCE) $(VULKAN_FLAGS) -c $< -o $@ # intermediate objects -llama.o: src/llama.cpp ggml/include/ggml.h ggml/include/ggml-alloc.h ggml/include/ggml-backend.h ggml/include/ggml-cuda.h ggml/include/ggml-metal.h include/llama.h otherarch/llama-util.h +llama.o: src/llama.cpp ggml/include/ggml.h ggml/include/ggml-alloc.h ggml/include/ggml-backend.h ggml/include/ggml-cuda.h ggml/include/ggml-metal.h include/llama.h otherarch/llama-util.h src/llama-chat.cpp src/llama-mmap.cpp src/llama-context.cpp src/llama-adapter.cpp src/llama-arch.cpp src/llama-batch.cpp src/llama-vocab.cpp src/llama-grammar.cpp src/llama-sampler.cpp src/llama-kv-cache.cpp src/llama-kv-cache-dsa.cpp src/llama-kv-cache-dsv4.cpp src/llama-kv-cache-iswa.cpp src/llama-memory-hybrid.cpp src/llama-memory-hybrid-iswa.cpp src/llama-memory-recurrent.cpp src/llama-model-loader.cpp src/llama-model-saver.cpp src/llama-quant.cpp src/llama-hparams.cpp src/llama-graph.cpp src/llama-io.cpp src/llama-memory.cpp common/fit.cpp ggml/include/ggml.h ggml/include/ggml-cpu.h ggml/include/ggml-cuda.h include/llama.h otherarch/llama-util.h + $(CXX) $(CXXFLAGS) -c $< -o $@ +llama-model.o: src/llama-model.cpp src/llama-model.h src/models/models.h ggml/include/ggml.h include/llama.h $(CXX) $(CXXFLAGS) -c $< -o $@ common.o: common/common.cpp common/common.h common/log.h $(CXX) $(CXXFLAGS) -c $< -o $@ @@ -698,8 +700,10 @@ llama-impl.o: src/llama-impl.cpp src/llama-impl.h $(CXX) $(CXXFLAGS) -c $< -o $@ budget.o: common/reasoning-budget.cpp common/reasoning-budget.h $(CXX) $(CXXFLAGS) -c $< -o $@ +chat.o: common/chat.cpp common/chat.h + $(CXX) $(CXXFLAGS) -c $< -o $@ -SDCPP_COMMON_BASENAMES := include/stable-diffusion.h src/conditioning/conditioner.hpp src/core/ggml_extend_backend.cpp src/core/ggml_extend_backend.h src/core/ggml_extend.hpp src/core/ggml_graph_cut.cpp src/core/ggml_graph_cut.h src/core/ordered_map.hpp src/core/rng.hpp src/core/rng_mt19937.hpp src/core/rng_philox.hpp src/core/tensor_ggml.hpp src/core/tensor.hpp src/core/util.cpp src/core/util.h src/extensions/generation_extension.h src/extensions/photomaker_extension.cpp src/extensions/pulid_extension.cpp src/kcpp_sd_extensions.h src/model/adapter/lora.hpp src/model/adapter/pmid.hpp src/model/adapter/pulid.hpp src/model/common/block.hpp src/model/common/rope.hpp src/model/diffusion/anima.hpp src/model/diffusion/boogu.hpp src/model/diffusion/control.hpp src/model/diffusion/dit.hpp src/model/diffusion/ernie_image.hpp src/model/diffusion/flux.hpp src/model/diffusion/hidream_o1.hpp src/model/diffusion/ideogram4.hpp src/model/diffusion/krea2.hpp src/model/diffusion/lens.hpp src/model/diffusion/ltxv.hpp src/model/diffusion/mmdit.hpp src/model/diffusion/model.hpp src/model/diffusion/pid.hpp src/model/diffusion/qwen_image.hpp src/model/diffusion/unet.hpp src/model/diffusion/wan.hpp src/model/diffusion/z_image.hpp src/model.h src/model_io/binary_io.h src/model_io/gguf_io.cpp src/model_io/gguf_io.h src/model_io/gguf_reader_ext.h src/model_io/pickle_io.cpp src/model_io/pickle_io.h src/model_io/safetensors_io.cpp src/model_io/safetensors_io.h src/model_io/tensor_storage.h src/model_io/torch_legacy_io.cpp src/model_io/torch_legacy_io.h src/model_io/torch_zip_io.cpp src/model_io/torch_zip_io.h src/model_loader.cpp src/model_loader.h src/model_manager.cpp src/model_manager.h src/model/te/clip.hpp src/model/te/llm.hpp src/model/te/t5.hpp src/model/upscaler/esrgan.hpp src/model/upscaler/ltx_latent_upscaler.hpp src/model/vae/auto_encoder_kl.hpp src/model/vae/ltx_audio_vae.hpp src/model/vae/ltx_vae.hpp src/model/vae/tae.hpp src/model/vae/vae.hpp src/model/vae/wan_vae.hpp src/name_conversion.cpp src/name_conversion.h src/runtime/cache_dit.hpp src/runtime/condition_cache_utils.hpp src/runtime/denoiser.hpp src/runtime/easycache.hpp src/runtime/gits_noise.h src/runtime/guidance.cpp src/runtime/guidance.h src/runtime/latent-preview.h src/runtime/preprocessing.hpp src/runtime/sample-cache.cpp src/runtime/sample-cache.h src/runtime/spectrum.hpp src/runtime/ucache.hpp src/stable-diffusion.cpp src/tokenizers/bpe_tokenizer.cpp src/tokenizers/bpe_tokenizer.h src/tokenizers/clip_tokenizer.cpp src/tokenizers/clip_tokenizer.h src/tokenizers/gemma_tokenizer.cpp src/tokenizers/gemma_tokenizer.h src/tokenizers/gpt_oss_tokenizer.cpp src/tokenizers/gpt_oss_tokenizer.h src/tokenizers/mistral_tokenizer.cpp src/tokenizers/mistral_tokenizer.h src/tokenizers/qwen2_tokenizer.cpp src/tokenizers/qwen2_tokenizer.h src/tokenizers/t5_unigram_tokenizer.cpp src/tokenizers/t5_unigram_tokenizer.h src/tokenizers/tokenizer.cpp src/tokenizers/tokenizer.h src/tokenizers/tokenize_util.cpp src/tokenizers/tokenize_util.h src/tokenizers/vocab/vocab.h src/upscaler.cpp src/upscaler.h src/weight_manager.h +SDCPP_COMMON_BASENAMES := include/stable-diffusion.h src/conditioning/conditioner.hpp src/core/backend_fit.cpp src/core/backend_fit.h src/core/ggml_extend_backend.cpp src/core/ggml_extend_backend.h src/core/ggml_extend.hpp src/core/ggml_graph_cut.cpp src/core/ggml_graph_cut.h src/core/layer_split_partition.cpp src/core/layer_split_partition.h src/core/ordered_map.hpp src/core/rng.hpp src/core/rng_mt19937.hpp src/core/rng_philox.hpp src/core/tensor_ggml.hpp src/core/tensor.hpp src/core/util.cpp src/core/util.h src/extensions/generation_extension.h src/extensions/photomaker_extension.cpp src/extensions/pulid_extension.cpp src/kcpp_sd_extensions.h src/model/adapter/lora.hpp src/model/adapter/pmid.hpp src/model/adapter/pulid.hpp src/model/common/block.hpp src/model/common/rope.hpp src/model/diffusion/anima.hpp src/model/diffusion/boogu.hpp src/model/diffusion/control.hpp src/model/diffusion/dit.hpp src/model/diffusion/ernie_image.hpp src/model/diffusion/flux.hpp src/model/diffusion/hidream_o1.hpp src/model/diffusion/ideogram4.hpp src/model/diffusion/krea2.hpp src/model/diffusion/lens.hpp src/model/diffusion/ltxv.hpp src/model/diffusion/minit2i.hpp src/model/diffusion/mmdit.hpp src/model/diffusion/model.hpp src/model/diffusion/pid.hpp src/model/diffusion/qwen_image.hpp src/model/diffusion/sefi_image.hpp src/model/diffusion/unet.hpp src/model/diffusion/wan.hpp src/model/diffusion/z_image.hpp src/model.h src/model_io/binary_io.h src/model_io/gguf_io.cpp src/model_io/gguf_io.h src/model_io/gguf_reader_ext.h src/model_io/pickle_io.cpp src/model_io/pickle_io.h src/model_io/safetensors_io.cpp src/model_io/safetensors_io.h src/model_io/streaming_writer.h src/model_io/tensor_storage.h src/model_io/torch_legacy_io.cpp src/model_io/torch_legacy_io.h src/model_io/torch_zip_io.cpp src/model_io/torch_zip_io.h src/model_loader.cpp src/model_loader.h src/model_manager.cpp src/model_manager.h src/model/te/clip.hpp src/model/te/llm.hpp src/model/te/t5.hpp src/model/upscaler/esrgan.hpp src/model/upscaler/ltx_latent_upscaler.hpp src/model/vae/auto_encoder_kl.hpp src/model/vae/ltx_audio_vae.hpp src/model/vae/ltx_vae.hpp src/model/vae/tae.hpp src/model/vae/vae.hpp src/model/vae/wan_vae.hpp src/name_conversion.cpp src/name_conversion.h src/runtime/cache_dit.hpp src/runtime/condition_cache_utils.hpp src/runtime/denoiser.hpp src/runtime/easycache.hpp src/runtime/gits_noise.h src/runtime/guidance.cpp src/runtime/guidance.h src/runtime/imatrix.cpp src/runtime/imatrix.h src/runtime/latent-preview.h src/runtime/preprocessing.hpp src/runtime/sample-cache.cpp src/runtime/sample-cache.h src/runtime/spectrum.hpp src/runtime/ucache.hpp src/stable-diffusion.cpp src/tokenizers/bpe_tokenizer.cpp src/tokenizers/bpe_tokenizer.h src/tokenizers/clip_tokenizer.cpp src/tokenizers/clip_tokenizer.h src/tokenizers/gemma_tokenizer.cpp src/tokenizers/gemma_tokenizer.h src/tokenizers/gpt_oss_tokenizer.cpp src/tokenizers/gpt_oss_tokenizer.h src/tokenizers/mistral_tokenizer.cpp src/tokenizers/mistral_tokenizer.h src/tokenizers/qwen2_tokenizer.cpp src/tokenizers/qwen2_tokenizer.h src/tokenizers/t5_unigram_tokenizer.cpp src/tokenizers/t5_unigram_tokenizer.h src/tokenizers/tokenizer.cpp src/tokenizers/tokenizer.h src/tokenizers/tokenize_util.cpp src/tokenizers/tokenize_util.h src/tokenizers/vocab/vocab.h src/upscaler.cpp src/upscaler.h src/weight_manager.h SDCPP_MAIN_BASENAMES := examples/cli/image_metadata.cpp examples/cli/image_metadata.h examples/cli/main.cpp examples/cli/msf_gif.h examples/common/common.cpp examples/common/common.h examples/common/log.cpp examples/common/log.h examples/common/media_io.cpp examples/common/media_io.h examples/common/resource_owners.hpp src/tokenizers/vocab/clip_merges.hpp src/tokenizers/vocab/gemma2_merges.hpp src/tokenizers/vocab/gemma2_vocab.hpp src/tokenizers/vocab/gemma_merges.hpp src/tokenizers/vocab/gemma_vocab.hpp src/tokenizers/vocab/gpt_oss_merges.hpp src/tokenizers/vocab/gpt_oss_vocab.hpp src/tokenizers/vocab/mistral_merges.hpp src/tokenizers/vocab/mistral_vocab.hpp src/tokenizers/vocab/qwen_merges.hpp src/tokenizers/vocab/t5.hpp src/tokenizers/vocab/umt5.hpp src/tokenizers/vocab/vocab.cpp src/convert.cpp src/version.cpp @@ -732,7 +736,7 @@ OBJS_SDTYPE := otherarch/sdcpp/sdtype_adapter.o $(OBJS_SDCOMMON) LLAMASERVER_SRCS := tools/server/main.cpp tools/server/server.cpp tools/server/server-schema.cpp tools/server/server-chat.cpp tools/server/server-common.cpp tools/server/server-context.cpp tools/server/server-http.cpp tools/server/server-models.cpp tools/server/server-queue.cpp tools/server/server-task.cpp tools/server/server-tools.cpp tools/server/ui.cpp COMMON_DOWNLOAD_SRCS := common/download.cpp common/hf-cache.cpp vendor/cpp-httplib/httplib.cpp -LLAMASERVER_COMMON_SRCS := common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) +LLAMASERVER_COMMON_SRCS := common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) LLAMASERVER_CXXFLAGS := -I./tools/mtmd @@ -755,7 +759,7 @@ music_default.o: otherarch/acestep/music_adapter.cpp $(CXX) $(CXXFLAGS) -c $< -o $@ # idiotic "for easier compilation" -GPTTYPE_ADAPTER = gpttype_adapter.cpp common/chat.cpp otherarch/llama_v2.cpp otherarch/llama_v3.cpp src/llama.cpp src/llama-chat.cpp src/llama-mmap.cpp src/llama-context.cpp src/llama-adapter.cpp src/llama-arch.cpp src/llama-batch.cpp src/llama-vocab.cpp src/llama-grammar.cpp src/llama-sampler.cpp src/llama-kv-cache.cpp src/llama-kv-cache-iswa.cpp src/llama-memory-hybrid.cpp src/llama-memory-hybrid-iswa.cpp src/llama-memory-recurrent.cpp src/llama-model-loader.cpp src/llama-model.cpp src/llama-quant.cpp src/llama-hparams.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp ggml/include/ggml.h ggml/include/ggml-cpu.h ggml/include/ggml-cuda.h include/llama.h otherarch/llama-util.h +GPTTYPE_ADAPTER = gpttype_adapter.cpp model_adapter.h otherarch/otherarch.h include/llama.h otherarch/llama_v2.cpp otherarch/llama_v3.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp gpttype_adapter_failsafe.o: $(GPTTYPE_ADAPTER) $(CXX) $(CXXFLAGS) $(FAILSAFE_FLAGS) -c $< -o $@ gpttype_adapter.o: $(GPTTYPE_ADAPTER) @@ -775,36 +779,36 @@ clean: rm -vf otherarch/sdcpp/*.o otherarch/sdcpp/*/*.o otherarch/sdcpp/*/*/*.o otherarch/sdcpp/*/*/*/*.o # useful tools -main: tools/completion/completion.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +main: tools/completion/main.cpp tools/completion/completion.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -mainvk: tools/completion/completion.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib - $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) -fitparams: tools/fit-params/main.cpp tools/fit-params/fit-params.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib - $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) -sdmain: $(OBJS_SDCOMMON) $(OBJS_SDMAIN) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +mainvk: tools/completion/main.cpp tools/completion/completion.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib + $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) +fitparams: tools/fit-params/main.cpp tools/fit-params/fit-params.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib + $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) +sdmain: $(OBJS_SDCOMMON) $(OBJS_SDMAIN) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -whispermain: otherarch/whispercpp/main.cpp otherarch/whispercpp/whisper.cpp build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +whispermain: otherarch/whispercpp/main.cpp otherarch/whispercpp/whisper.cpp build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -ttsmain: tools/tts/tts.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +ttsmain: tools/tts/tts.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -gguf-split: tools/gguf-split/gguf-split.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o build-info.h clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +gguf-split: tools/gguf-split/gguf-split.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o build-info.h clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -mtmd-cli: tools/mtmd/mtmd-cli.cpp tools/mtmd/clip.cpp common/debug.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h mtmd.o mtmd-helper.o mtmd-image.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +mtmd-cli: tools/mtmd/mtmd-cli.cpp tools/mtmd/clip.cpp common/debug.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h mtmd.o mtmd-helper.o mtmd-image.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -embedding: examples/embedding/embedding.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) src/llama-cparams.cpp build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +embedding: examples/embedding/embedding.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) src/llama-cparams.cpp build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -embeddingvk: examples/embedding/embedding.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) src/llama-cparams.cpp build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib - $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) -ttscppmain: otherarch/ttscpp/cli/cli.cpp otherarch/ttscpp/cli/playback.cpp otherarch/ttscpp/cli/playback.h otherarch/ttscpp/cli/write_file.cpp otherarch/ttscpp/cli/write_file.h otherarch/ttscpp/cli/vad.cpp otherarch/ttscpp/cli/vad.h otherarch/ttscpp/src/ttscpp.cpp otherarch/ttscpp/src/ttstokenizer.cpp otherarch/ttscpp/src/ttssampler.cpp otherarch/ttscpp/src/parler_model.cpp otherarch/ttscpp/src/dac_model.cpp otherarch/ttscpp/src/ttsutil.cpp otherarch/ttscpp/src/ttsargs.cpp otherarch/ttscpp/src/ttst5_encoder_model.cpp otherarch/ttscpp/src/phonemizer.cpp otherarch/ttscpp/src/tts_model.cpp otherarch/ttscpp/src/kokoro_model.cpp otherarch/ttscpp/src/dia_model.cpp otherarch/ttscpp/src/orpheus_model.cpp otherarch/ttscpp/src/snac_model.cpp otherarch/ttscpp/src/general_neural_audio_codec.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +embeddingvk: examples/embedding/embedding.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) src/llama-cparams.cpp build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib + $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) +ttscppmain: otherarch/ttscpp/cli/cli.cpp otherarch/ttscpp/cli/playback.cpp otherarch/ttscpp/cli/playback.h otherarch/ttscpp/cli/write_file.cpp otherarch/ttscpp/cli/write_file.h otherarch/ttscpp/cli/vad.cpp otherarch/ttscpp/cli/vad.h otherarch/ttscpp/src/ttscpp.cpp otherarch/ttscpp/src/ttstokenizer.cpp otherarch/ttscpp/src/ttssampler.cpp otherarch/ttscpp/src/parler_model.cpp otherarch/ttscpp/src/dac_model.cpp otherarch/ttscpp/src/ttsutil.cpp otherarch/ttscpp/src/ttsargs.cpp otherarch/ttscpp/src/ttst5_encoder_model.cpp otherarch/ttscpp/src/phonemizer.cpp otherarch/ttscpp/src/tts_model.cpp otherarch/ttscpp/src/kokoro_model.cpp otherarch/ttscpp/src/dia_model.cpp otherarch/ttscpp/src/orpheus_model.cpp otherarch/ttscpp/src/snac_model.cpp otherarch/ttscpp/src/general_neural_audio_codec.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -qwen3tts: otherarch/qwen3tts/q3ttsmain.cpp otherarch/qwen3tts/qwen3_tts.cpp otherarch/qwen3tts/text_tokenizer.cpp otherarch/qwen3tts/gguf_loader.cpp otherarch/qwen3tts/tts_transformer.cpp otherarch/qwen3tts/audio_tokenizer_decoder.cpp otherarch/qwen3tts/audio_tokenizer_encoder.cpp otherarch/qwen3tts/coreml_code_predictor_stub.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +qwen3tts: otherarch/qwen3tts/q3ttsmain.cpp otherarch/qwen3tts/qwen3_tts.cpp otherarch/qwen3tts/text_tokenizer.cpp otherarch/qwen3tts/gguf_loader.cpp otherarch/qwen3tts/tts_transformer.cpp otherarch/qwen3tts/audio_tokenizer_decoder.cpp otherarch/qwen3tts/audio_tokenizer_encoder.cpp otherarch/qwen3tts/coreml_code_predictor_stub.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -rpcserver: tools/rpc/rpc-server.cpp common/arg.cpp common/chat.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib - $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) -llamaserver: $(LLAMASERVER_SRCS) $(LLAMASERVER_COMMON_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +rpcserver: tools/rpc/rpc-server.cpp common/arg.cpp common/preset.cpp $(COMMON_DOWNLOAD_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib + $(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) +llamaserver: $(LLAMASERVER_SRCS) $(LLAMASERVER_COMMON_SRCS) build-info.h ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $(LLAMASERVER_CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -llamaservervk: $(LLAMASERVER_SRCS) $(LLAMASERVER_COMMON_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib - $(CXX) $(CXXFLAGS) $(LLAMASERVER_CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) +llamaservervk: $(LLAMASERVER_SRCS) $(LLAMASERVER_COMMON_SRCS) build-info.h ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o console.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o ggml-repack.o $(OBJS_FULL) $(OBJS) lib/vulkan-1.lib + $(CXX) $(CXXFLAGS) $(LLAMASERVER_CXXFLAGS) -DGGML_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS) ggml/src/ggml-vulkan-shaders.cpp: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp ifdef VULKAN_BUILD @@ -904,14 +908,14 @@ else endif #generated libraries -koboldcpp_default: ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +koboldcpp_default: ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(DEFAULT_BUILD) -koboldcpp_macos_failsafe: ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +koboldcpp_macos_failsafe: ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(DEFAULT_BUILD) ifdef FAILSAFE_BUILD -koboldcpp_failsafe: ggml_v4_failsafe.o ggml-cpu_v4_failsafe.o ggml-ops-failsafe.o ggml-vec-failsafe.o ggml-binops.o ggml-unops.o ggml_v3_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o gpttype_adapter_failsafe.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FAILSAFE) $(OBJS) +koboldcpp_failsafe: ggml_v4_failsafe.o ggml-cpu_v4_failsafe.o ggml-ops-failsafe.o ggml-vec-failsafe.o ggml-binops.o ggml-unops.o ggml_v3_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o gpttype_adapter_failsafe.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FAILSAFE) $(OBJS) $(FAILSAFE_BUILD) else koboldcpp_failsafe: @@ -919,7 +923,7 @@ koboldcpp_failsafe: endif ifdef NOAVX2_BUILD -koboldcpp_noavx2: ggml_v4_noavx2.o ggml-cpu_v4_noavx2.o ggml-ops-noavx2.o ggml-vec-noavx2.o ggml-binops.o ggml-unops.o ggml_v3_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o gpttype_adapter_failsafe.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_SIMPLE) $(OBJS) +koboldcpp_noavx2: ggml_v4_noavx2.o ggml-cpu_v4_noavx2.o ggml-ops-noavx2.o ggml-vec-noavx2.o ggml-binops.o ggml-unops.o ggml_v3_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o gpttype_adapter_failsafe.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_default.o tts_default.o music_default.o embeddings_default.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_SIMPLE) $(OBJS) $(NOAVX2_BUILD) else koboldcpp_noavx2: @@ -927,7 +931,7 @@ koboldcpp_noavx2: endif ifdef CUBLAS_BUILD -koboldcpp_cublas: ggml_v4_cublas.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o gpttype_adapter_cublas.o $(OBJS_SDTYPE) whispercpp_cublas.o tts_default.o music_default.o embeddings_default.o clip_cublas.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_cublas.o ggml-repack.o $(CUBLAS_OBJS) $(OBJS_FULL) $(OBJS) +koboldcpp_cublas: ggml_v4_cublas.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o gpttype_adapter_cublas.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_cublas.o tts_default.o music_default.o embeddings_default.o clip_cublas.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_cublas.o ggml-repack.o $(CUBLAS_OBJS) $(OBJS_FULL) $(OBJS) $(CUBLAS_BUILD) else koboldcpp_cublas: @@ -935,7 +939,7 @@ koboldcpp_cublas: endif ifdef HIPBLAS_BUILD -koboldcpp_hipblas: ggml_v4_cublas.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o gpttype_adapter_cublas.o $(OBJS_SDTYPE) whispercpp_cublas.o tts_default.o music_default.o embeddings_default.o clip_cublas.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_cublas.o ggml-repack.o $(HIP_OBJS) $(OBJS_FULL) $(OBJS) +koboldcpp_hipblas: ggml_v4_cublas.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o gpttype_adapter_cublas.o llama.o chat.o llama-model.o $(OBJS_SDTYPE) whispercpp_cublas.o tts_default.o music_default.o embeddings_default.o clip_cublas.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_cublas.o ggml-repack.o $(HIP_OBJS) $(OBJS_FULL) $(OBJS) $(HIPBLAS_BUILD) else koboldcpp_hipblas: @@ -943,12 +947,12 @@ koboldcpp_hipblas: endif ifdef VULKAN_BUILD -koboldcpp_vulkan: ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter_vulkan.o ggml-vulkan.o ggml-vulkan-shaders.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_FULL) $(OBJS) +koboldcpp_vulkan: ggml_v4_vulkan.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o ggml_v3.o ggml_v2.o ggml_v1.o expose.o gpttype_adapter_vulkan.o llama.o chat.o llama-model.o ggml-vulkan.o ggml-vulkan-shaders.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(VULKAN_BUILD) ifdef NOAVX2_BUILD -koboldcpp_vulkan_noavx2: ggml_v4_vulkan_noavx2.o ggml-cpu_v4_noavx2.o ggml-ops-noavx2.o ggml-vec-noavx2.o ggml-binops.o ggml-unops.o ggml_v3_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o gpttype_adapter_vulkan_noavx2.o ggml-vulkan-noext.o ggml-vulkan-shaders-noext.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_SIMPLE) $(OBJS) +koboldcpp_vulkan_noavx2: ggml_v4_vulkan_noavx2.o ggml-cpu_v4_noavx2.o ggml-ops-noavx2.o ggml-vec-noavx2.o ggml-binops.o ggml-unops.o ggml_v3_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o gpttype_adapter_vulkan_noavx2.o llama.o chat.o llama-model.o ggml-vulkan-noext.o ggml-vulkan-shaders-noext.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_SIMPLE) $(OBJS) $(VULKAN_BUILD) -koboldcpp_vulkan_failsafe: ggml_v4_vulkan_failsafe.o ggml-cpu_v4_failsafe.o ggml-ops-failsafe.o ggml-vec-failsafe.o ggml-binops.o ggml-unops.o ggml_v3_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o gpttype_adapter_vulkan_noavx2.o ggml-vulkan-noext.o ggml-vulkan-shaders-noext.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_SIMPLER) $(OBJS) +koboldcpp_vulkan_failsafe: ggml_v4_vulkan_failsafe.o ggml-cpu_v4_failsafe.o ggml-ops-failsafe.o ggml-vec-failsafe.o ggml-binops.o ggml-unops.o ggml_v3_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o gpttype_adapter_vulkan_noavx2.o llama.o chat.o llama-model.o ggml-vulkan-noext.o ggml-vulkan-shaders-noext.o $(OBJS_SDTYPE) whispercpp_vulkan.o tts_default.o music_default.o embeddings_default.o clip_vulkan.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_vulkan.o ggml-repack.o $(OBJS_SIMPLER) $(OBJS) $(VULKAN_BUILD) else koboldcpp_vulkan_noavx2: @@ -966,17 +970,17 @@ koboldcpp_vulkan_failsafe: endif # tools -quantize_gguf: tools/quantize/main.cpp tools/quantize/quantize.cpp common/imatrix-loader.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_gguf: tools/quantize/main.cpp tools/quantize/quantize.cpp common/imatrix-loader.cpp ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -quantize_gptj: otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_gptj: otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -quantize_gpt2: otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_gpt2: otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -quantize_neox: otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_neox: otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -quantize_mpt: otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_mpt: otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o clip_default.o mtmd.o mtmd-helper.o mtmd-image.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -quantize_ace: otherarch/acestep/quantize-acestep.cpp tools/mtmd/clip.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) +quantize_ace: otherarch/acestep/quantize-acestep.cpp tools/mtmd/clip.cpp ggml_v3.o ggml.o ggml-cpu.o ggml-ops.o ggml-vec.o ggml-binops.o ggml-unops.o llama.o chat.o llama-model.o ggml-backend.o ggml-backend-meta.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) diff --git a/colab.ipynb b/colab.ipynb index e5194a430..b8778a411 100644 --- a/colab.ipynb +++ b/colab.ipynb @@ -62,7 +62,7 @@ "Model = \"https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive/resolve/main/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q5_K_M.gguf\" #@param [\"https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive/resolve/main/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q5_K_M.gguf\",\"https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf\",\"https://huggingface.co/KoboldAI/LLaMA2-13B-Estopia-GGUF/resolve/main/LLaMA2-13B-Estopia.Q4_K_S.gguf\",\"https://huggingface.co/KoboldAI/Llama-3.1-8B-BookAdventures-GGUF/resolve/main/Llama-3.1-8B-BookAdventures.Q6_K.gguf\",\"https://huggingface.co/bartowski/TheDrummer_Cydonia-24B-v4.2.0-GGUF/resolve/main/TheDrummer_Cydonia-24B-v4.2.0-Q4_K_S.gguf\",\"https://huggingface.co/mradermacher/Broken-Tutu-24B-GGUF/resolve/main/Broken-Tutu-24B.Q4_K_S.gguf\",\"https://huggingface.co/bartowski/PocketDoc_Dans-PersonalityEngine-V1.3.0-24b-GGUF/resolve/main/PocketDoc_Dans-PersonalityEngine-V1.3.0-24b-Q4_K_S.gguf\",\"https://huggingface.co/LatitudeGames/Harbinger-24B-GGUF/resolve/main/Harbinger-24B-Q4_K_S.gguf\",\"https://huggingface.co/LatitudeGames/Muse-12B-GGUF/resolve/main/Muse-12B-Q4_K_S.gguf\",\"https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/resolve/main/Qwen3-VL-8B-Instruct-Q6_K.gguf\",\"https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/resolve/main/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_S.gguf\",\"https://huggingface.co/ggml-org/gpt-oss-20b-GGUF/resolve/main/gpt-oss-20b-mxfp4.gguf\",\"https://huggingface.co/KoboldAI/Llama-3.1-8B-BookAdventures-GGUF/resolve/main/Llama-3.1-8B-BookAdventures.Q6_K.gguf\",\"https://huggingface.co/bartowski/google_gemma-3-12b-it-GGUF/resolve/main/google_gemma-3-12b-it-Q4_K_S.gguf\",\"https://huggingface.co/unsloth/gemma-3n-E4B-it-GGUF/resolve/main/gemma-3n-E4B-it-Q6_K.gguf\",\"https://huggingface.co/unsloth/GLM-4-9B-0414-GGUF/resolve/main/GLM-4-9B-0414-Q6_K.gguf\",\"https://huggingface.co/mradermacher/Fimbulvetr-11B-v2-GGUF/resolve/main/Fimbulvetr-11B-v2.Q4_K_S.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-13B-GGUF/resolve/main/mythomax-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/ReMM-SLERP-L2-13B-GGUF/resolve/main/remm-slerp-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/Xwin-LM-13B-v0.2-GGUF/resolve/main/xwin-lm-13b-v0.2.Q4_K_M.gguf\",\"https://huggingface.co/mradermacher/mini-magnum-12b-v1.1-GGUF/resolve/main/mini-magnum-12b-v1.1.Q4_K_S.gguf\",\"https://huggingface.co/TheBloke/Stheno-L2-13B-GGUF/resolve/main/stheno-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-Kimiko-v2-13B-GGUF/resolve/main/mythomax-l2-kimiko-v2-13b.Q4_K_M.gguf\",\"https://huggingface.co/bartowski/Rocinante-12B-v1.1-GGUF/resolve/main/Rocinante-12B-v1.1-Q4_K_S.gguf\",\"https://huggingface.co/TheBloke/MistRP-Airoboros-7B-GGUF/resolve/main/mistrp-airoboros-7b.Q4_K_S.gguf\",\"https://huggingface.co/TheBloke/airoboros-mistral2.2-7B-GGUF/resolve/main/airoboros-mistral2.2-7b.Q4_K_S.gguf\",\"https://huggingface.co/concedo/KobbleTinyV2-1.1B-GGUF/resolve/main/KobbleTiny-Q4_K.gguf\",\"https://huggingface.co/grimjim/kukulemon-7B-GGUF/resolve/main/kukulemon-7B.Q8_0.gguf\",\"https://huggingface.co/mradermacher/LemonKunoichiWizardV3-GGUF/resolve/main/LemonKunoichiWizardV3.Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Kunoichi-DPO-v2-7B-GGUF-Imatrix/resolve/main/Kunoichi-DPO-v2-7B-Q4_K_M-imatrix.gguf\",\"https://huggingface.co/mradermacher/L3-8B-Stheno-v3.2-i1-GGUF/resolve/main/L3-8B-Stheno-v3.2.i1-Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Llama-3-Lumimaid-8B-v0.1-OAS-GGUF-IQ-Imatrix/resolve/main/v2-Llama-3-Lumimaid-8B-v0.1-OAS-Q4_K_M-imat.gguf\",\"https://huggingface.co/bartowski/NeuralDaredevil-8B-abliterated-GGUF/resolve/main/NeuralDaredevil-8B-abliterated-Q4_K_M.gguf\",\"https://huggingface.co/bartowski/L3-8B-Lunaris-v1-GGUF/resolve/main/L3-8B-Lunaris-v1-Q4_K_M.gguf\",\"https://huggingface.co/mradermacher/L3-Umbral-Mind-RP-v2.0-8B-GGUF/resolve/main/L3-Umbral-Mind-RP-v2.0-8B.Q4_K_M.gguf\",\"https://huggingface.co/bartowski/TheDrummer_Cydonia-24B-v2-GGUF/resolve/main/TheDrummer_Cydonia-24B-v2-Q4_K_S.gguf\",\"https://huggingface.co/bartowski/PocketDoc_Dans-PersonalityEngine-V1.2.0-24b-GGUF/resolve/main/PocketDoc_Dans-PersonalityEngine-V1.2.0-24b-IQ4_XS.gguf\",\"https://huggingface.co/mradermacher/Tlacuilo-12B-GGUF/resolve/main/Tlacuilo-12B.Q4_K_S.gguf\"] {\"allow-input\":true}\n", "MdCommand = \"\" #@markdown
\n", "Layers = \"Auto\" #@param [\"Auto\",\"999\"]{allow-input: true}\n", - "ContextSize = \"4096\" #@param [\"4096\",\"8192\",\"12288\",\"16384\"] {allow-input: true}\n", + "ContextSize = \"4096\" #@param [\"4096\",\"8192\",\"12288\",\"16384\",\"24576\",\"32768\",\"40960\"] {allow-input: true}\n", "\n", "#@markdown
\n", "LoadVisionMMProjector = False #@param {type:\"boolean\"}\n", @@ -130,7 +130,7 @@ " if Template == \"Gemma4 E4B Uncensored (General)\":\n", " Customized = True\n", " Model = \"https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive/resolve/main/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q5_K_M.gguf\"\n", - " CustomCtxSize = \"16384\"\n", + " CustomCtxSize = \"40960\"\n", " CustomMmproj = \"https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/resolve/main/mmproj-BF16.gguf\"\n", " if Template == \"Tiefighter 13B (General)\":\n", " Customized = True\n", diff --git a/common/arg.cpp b/common/arg.cpp index 53c207bbc..bf2b62460 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -468,7 +468,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params // the first part is what gets loaded, so point params.model.path at it if (!url_tasks.empty()) { std::string first_path = url_tasks.front().local_path; - url_tasks.front().on_done = [&]() { params.model.path = first_path; }; + url_tasks.front().on_done = [&, first_path]() { params.model.path = first_path; }; } for (auto & task : url_tasks) { tasks.push_back(std::move(task)); @@ -497,13 +497,15 @@ void common_models_handler_apply(common_models_handler & handler, common_params } // handle hf_plan tasks - auto add_tasks = [&opts, &tasks](const hf_cache::hf_files & model_files, common_params_model & model) { + auto add_tasks = [&opts, &tasks](const hf_cache::hf_files & model_files, + const hf_cache::hf_file & primary, + common_params_model & model) { for (size_t i = 0; i < model_files.size(); ++i) { auto & model_file = model_files[i]; - bool is_first = (i == 0); - tasks.emplace_back(model_file, opts, [&, is_first]() { - if (is_first) { - // only use first part as model path + bool is_primary = (model_file.path == primary.path); + tasks.emplace_back(model_file, opts, [&, is_primary]() { + if (is_primary) { + // the primary file is the first split (00001-of), use it as model path model.path = hf_cache::finalize_file(model_file); } else { hf_cache::finalize_file(model_file); @@ -512,7 +514,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params } }; if (!plan.model_files.empty()) { - add_tasks(plan.model_files, params.model); + add_tasks(plan.model_files, plan.primary, params.model); } if (!plan.mmproj.local_path.empty()) { tasks.emplace_back(plan.mmproj, opts, [&]() { @@ -540,12 +542,12 @@ void common_models_handler_apply(common_models_handler & handler, common_params // handle plan_spec (e.g. --spec-draft-hf) if (!plan_spec.model_files.empty()) { - add_tasks(plan_spec.model_files, params.speculative.draft.mparams); + add_tasks(plan_spec.model_files, plan_spec.primary, params.speculative.draft.mparams); } // handle vocoder plan (e.g. --hf-repo-v) if (!plan_voc.model_files.empty()) { - add_tasks(plan_voc.model_files, params.vocoder.model); + add_tasks(plan_voc.model_files, plan_voc.primary, params.vocoder.model); } // run all tasks in parallel @@ -3297,6 +3299,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sampling.reasoning_budget_message = value; } ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK_BUDGET_MESSAGE")); + add_opt(common_arg( + {"--reasoning-preserve"}, + {"--no-reasoning-preserve"}, + "preserve reasoning trace in the full history, not just the last assistant message (default: template default)\n" + "compatible with certain templates having 'supports_preserve_reasoning' capability\n" + "example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking", + [](common_params & params, bool value) { + if (value) { + params.default_template_kwargs["preserve_reasoning"] = "true"; + } else { + params.default_template_kwargs["preserve_reasoning"] = "false"; + } + } + ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_PRESERVE")); add_opt(common_arg( {"--chat-template"}, "JINJA_TEMPLATE", string_format( @@ -3472,7 +3488,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.offline = true; } - ).set_env("LLAMA_ARG_OFFLINE")); + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_OFFLINE")); add_opt(common_arg( {"-lv", "--verbosity", "--log-verbosity"}, "N", string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n" diff --git a/common/chat.cpp b/common/chat.cpp index 6db04e1ff..28fba2ff7 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -7,7 +7,6 @@ #include "ggml.h" #include "json-schema-to-grammar.h" #include "log.h" -#include "regex-partial.cpp" #include "reasoning-budget.h" #include "chat-auto-parser-generator.cpp" #include "chat-auto-parser-helpers.cpp" @@ -926,6 +925,10 @@ static std::string common_chat_template_direct_apply_impl( if (inputs.add_generation_prompt) { inp["add_generation_prompt"] = true; } + if (inp.contains("preserve_reasoning") && inp["preserve_reasoning"].is_boolean()) { + bool enabled = inp["preserve_reasoning"].get(); + jinja::caps_apply_preserve_reasoning(ctx, enabled); + } jinja::global_from_json(ctx, inp, inputs.mark_input); @@ -2388,6 +2391,166 @@ static void func_args_not_string(json & messages) { } } +// Trim leading/trailing whitespace from message contents before rendering. This +// has to run on the messages (not on the rendered JSON) because templates with +// string-only content caps concatenate typed content parts into a single string +// during rendering, after which the per-part whitespace can no longer be reached. +// Both the plain string content and the text of typed content parts are trimmed. +static void trim_all_content(std::vector & messages) { + for (auto & message : messages) { + message.content = trim_whitespace(message.content); + message.reasoning_content = trim_whitespace(message.reasoning_content); + for (auto & part : message.content_parts) { + if (part.type == "text") { + part.text = trim_whitespace(part.text); + } + } + } +} + +} + +// MiniCPM5 format: +// - Reasoning: {reasoning} (optional) +// - Tool calls: value +static common_chat_params common_chat_params_init_minicpm5(const common_chat_template & tmpl, + const autoparser::generation_params & inputs) { + common_chat_params data; + + data.prompt = common_chat_template_direct_apply_impl(tmpl, inputs); + data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs); + data.format = COMMON_CHAT_FORMAT_PEG_NATIVE; + data.supports_thinking = true; + data.preserved_tokens = { + "", + "", + "", + "", + }; + + data.thinking_start_tag = ""; + data.thinking_end_tag = ""; + + data.message_delimiters = { + { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant" }, + { COMMON_CHAT_ROLE_TOOL, "<|im_start|>user\n" }, + { COMMON_CHAT_ROLE_USER, "<|im_start|>user" }, + { COMMON_CHAT_ROLE_SYSTEM, "<|im_start|>system" }, + }; + + auto has_tools = inputs.tools.is_array() && !inputs.tools.empty(); + auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty(); + auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE; + auto include_grammar = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE); + + if (inputs.has_continuation()) { + const auto & msg = inputs.continue_msg; + + data.generation_prompt = "<|im_start|>assistant\n\n" + msg.reasoning_content; + if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) { + data.generation_prompt += "\n\n\n" + msg.render_content(); + } + + data.prompt += data.generation_prompt; + } + + auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) { + auto generation_prompt = p.literal("<|im_start|>assistant\n"); + + auto reasoning = p.eps(); + if (extract_reasoning) { + reasoning = ("" << p.reasoning(p.until("")) << "") + p.space(); + } + + // Response format parser + if (has_response_format) { + return generation_prompt + reasoning + p.content(p.schema(p.json(), "response-format", inputs.json_schema)); + } + + if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) { + // CDATA lets a value carry characters that would otherwise close the tag (e.g. + // ); capture the inner text only, excluding the CDATA markers. + auto string_value = p.choice({ + p.literal("")) + p.literal("]]>"), "]]>") + p.tool_arg_close(p.literal("")), + p.negate(p.literal("")) + p.tool_arg_close(p.literal("")), "") + }); + + auto tool_choice = p.choice(); + foreach_function(inputs.tools, [&](const json & tool) { + const auto & function = tool.at("function"); + const std::string name = function.at("name"); + auto params = function.contains("parameters") ? function.at("parameters") : json::object(); + + auto args = p.eps(); + if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) { + auto schema_info = common_schema_info(); + schema_info.resolve_refs(params); + + auto arg_choice = p.choice(); + for (const auto & [prop_name, prop_schema] : params.at("properties").items()) { + auto value_parser = p.eps(); + if (schema_info.resolves_to_string(prop_schema)) { + value_parser = string_value; + } else { + value_parser = p.tool_arg_json_value( + p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false) + ) + p.tool_arg_close(p.literal("")); + } + + auto arg_rule = p.tool_arg( + p.tool_arg_open(p.literal("")) + + value_parser + ); + + arg_choice |= arg_rule; + } + args = p.zero_or_more(arg_choice + p.space()); + } + + auto tool_parser = p.tool( + p.tool_open(p.literal("")) + << p.tool_args(args) + << p.tool_close(p.literal(""))); + + tool_choice |= p.rule("tool-" + name, tool_parser); + }); + + auto max_calls = inputs.parallel_tool_calls ? -1 : 1; + auto tool_calls = p.trigger_rule("tool-call", p.repeat(tool_choice + p.space(), 1, max_calls)); + + auto content = p.content(p.until(" common_chat_try_specialized_template( return common_chat_params_init_gemma4(tmpl, params); } + // MiniCPM5 - XML tool calls with ... + if (src.find("Tool usage guidelines:") != std::string::npos && + src.find("template_tool_use ? *tmpls->template_tool_use : *tmpls->template_default; const auto & src = tmpl.source(); const auto & caps = tmpl.original_caps(); - params.messages = render_message_to_json(inputs.messages, tmpl.original_caps()); + std::vector trimmed_messages; + const std::vector * messages_to_render = &inputs.messages; + if (src.find("You have access to the following functions in JSONSchema format") != std::string::npos) { + // StepFun: trim message contents (including typed content parts) before rendering, + // otherwise leftover whitespace drives the model into reasoning loops (issue #24181) + trimmed_messages = inputs.messages; + workaround::trim_all_content(trimmed_messages); + messages_to_render = &trimmed_messages; + } + params.messages = render_message_to_json(*messages_to_render, tmpl.original_caps()); params.tool_choice = inputs.tool_choice; params.reasoning_format = inputs.reasoning_format; params.enable_thinking = inputs.enable_thinking; diff --git a/common/common.cpp b/common/common.cpp index 44c6f8ab1..14f1a7892 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -231,7 +231,7 @@ bool set_process_priority(enum ggml_sched_priority prio) { } if (!SetPriorityClass(GetCurrentProcess(), p)) { - LOG_WRN("failed to set process priority class %d : (%d)\n", prio, (int) GetLastError()); + COM_WRN("failed to set process priority class %d : (%d)\n", prio, (int) GetLastError()); return false; } @@ -257,7 +257,7 @@ bool set_process_priority(enum ggml_sched_priority prio) { } if (setpriority(PRIO_PROCESS, 0, p) != 0) { - LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno); + COM_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno); return false; } return true; @@ -290,14 +290,14 @@ void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_para if (n_set && n_set < cpuparams.n_threads) { // Not enough set bits, may experience performance issues. - LOG_WRN("Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads); + COM_WRN("Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads); } } bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) { size_t dash_loc = range.find('-'); if (dash_loc == std::string::npos) { - LOG_ERR("Format of CPU range is invalid! Expected []-[].\n"); + COM_ERR("%s", "Format of CPU range is invalid! Expected []-[].\n"); return false; } @@ -309,7 +309,7 @@ bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THRE } else { start_i = std::stoull(range.substr(0, dash_loc)); if (start_i >= GGML_MAX_N_THREADS) { - LOG_ERR("Start index out of bounds!\n"); + COM_ERR("%s", "Start index out of bounds!\n"); return false; } } @@ -319,7 +319,7 @@ bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THRE } else { end_i = std::stoull(range.substr(dash_loc + 1)); if (end_i >= GGML_MAX_N_THREADS) { - LOG_ERR("End index out of bounds!\n"); + COM_ERR("%s", "End index out of bounds!\n"); return false; } } @@ -339,7 +339,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD } size_t num_digits = mask.length() - start_i; - if (num_digits > 128) num_digits = 128; + num_digits = std::min(num_digits, 128); size_t end_i = num_digits + start_i; @@ -354,7 +354,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD } else if (c >= 'A' && c <= 'F') { id -= 'A' - 10; } else { - LOG_ERR("Invalid hex character '%c' at position %d\n", c, int32_t(i)); + COM_ERR("Invalid hex character '%c' at position %d\n", c, int32_t(i)); return false; } @@ -385,21 +385,21 @@ void common_params_print_info(const common_params & params, bool print_devices) #else const char * build_type = " (debug)"; #endif - LOG_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type); + COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type); - LOG_INF("log_info: verbosity = %d (adjust with the `-lv N` CLI arg)\n", common_log_get_verbosity_thold()); + COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold()); // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device if (print_devices) { - LOG_INF("device_info:\n"); + COM_TRC("%s", "device_info:\n"); for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { auto * dev = ggml_backend_dev_get(i); size_t free, total; ggml_backend_dev_memory(dev, &free, &total); - LOG_INF(" - %-8s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024); + COM_TRC(" - %-8s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024); } } - LOG_INF("%s\n", common_params_get_system_info(params).c_str()); + COM_TRC("%s\n", common_params_get_system_info(params).c_str()); } std::string common_params_get_system_info(const common_params & params) { @@ -666,7 +666,7 @@ void string_process_escapes(std::string & input) { bool string_parse_kv_override(const char * data, std::vector & overrides) { const char * sep = strchr(data, '='); if (sep == nullptr || sep - data >= 128) { - LOG_ERR("%s: malformed KV override '%s'\n", __func__, data); + COM_ERR("%s: malformed KV override '%s'\n", __func__, data); return false; } llama_model_kv_override kvo; @@ -689,20 +689,20 @@ bool string_parse_kv_override(const char * data, std::vector 127) { - LOG_ERR("%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data); + COM_ERR("%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data); return false; } strncpy(kvo.val_str, sep, 127); kvo.val_str[127] = '\0'; } else { - LOG_ERR("%s: invalid type for KV override '%s'\n", __func__, data); + COM_ERR("%s: invalid type for KV override '%s'\n", __func__, data); return false; } overrides.emplace_back(std::move(kvo)); @@ -1205,8 +1205,8 @@ common_init_result::common_init_result(common_params & params, bool model_only) auto cparams = common_context_params_to_llama(params); if (params.fit_params) { - LOG_INF("%s: fitting params to device memory ...\n", __func__); - LOG_INF("%s: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n", __func__); + COM_TRC("%s", "fitting params to device memory ...\n"); + COM_TRC("%s", "(for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n"); common_fit_params(params.model.path.c_str(), &mparams, &cparams, params.tensor_split, params.tensor_buft_overrides.data(), @@ -1233,7 +1233,7 @@ common_init_result::common_init_result(common_params & params, bool model_only) llama_adapter_lora_ptr lora; lora.reset(llama_adapter_lora_init(model, la.path.c_str())); if (lora == nullptr) { - LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str()); + COM_ERR("failed to load lora adapter '%s'\n", la.path.c_str()); pimpl->model.reset(model); return; } @@ -1252,14 +1252,14 @@ common_init_result::common_init_result(common_params & params, bool model_only) common_init_sampler_from_model(model, params.sampling); if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) { - LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__); + COM_WRN("%s", "vocab does not have an EOS token, ignoring --ignore-eos\n"); params.sampling.ignore_eos = false; } // initialize once for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) { if (llama_vocab_is_eog(vocab, i)) { - LOG_TRC("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY); + COM_TRC("added %s logit bias = %f\n", common_token_to_piece(vocab, i).c_str(), -INFINITY); params.sampling.logit_bias_eog.push_back({i, -INFINITY}); } } @@ -1297,7 +1297,7 @@ common_init_result::common_init_result(common_params & params, bool model_only) llama_context * lctx = llama_init_from_model(model, cparams); if (lctx == NULL) { - LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str()); + COM_ERR("failed to create context with model '%s'\n", params.model.path.c_str()); return; } @@ -1334,7 +1334,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode llama_model * model = res->model(); if (model == NULL) { - LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str()); + COM_ERR("failed to load model '%s'\n", params.model.path.c_str()); return res; } @@ -1344,14 +1344,14 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode llama_context * lctx = res->context(); if (lctx == NULL) { - LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str()); + COM_ERR("failed to create context with model '%s'\n", params.model.path.c_str()); return res; } const llama_vocab * vocab = llama_model_get_vocab(model); if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) { - LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__); + COM_WRN("%s", "KV cache shifting is not supported for this context, disabling KV cache shifting\n"); params.ctx_shift = false; } @@ -1380,7 +1380,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode bool ok = true; if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) { - LOG_WRN("%s: warning: vocab does not have a BOS token, reranking will not work\n", __func__); + COM_WRN("%s", "vocab does not have a BOS token, reranking will not work\n"); ok = false; } @@ -1389,10 +1389,10 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL; if (!has_eos && !has_sep && !has_rerank_prompt) { - LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__); + COM_WRN("%s", "vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n"); ok = false; } else if (!has_eos) { - LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__); + COM_WRN("%s", "vocab does not have an EOS token, using SEP token as fallback\n"); } if (!ok) { @@ -1405,7 +1405,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode } if (params.warmup) { - LOG_INF("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__); + COM_TRC("%s", "warming up the model with an empty run - please wait ... (--no-warmup to disable)\n"); std::vector tmp; llama_token bos = llama_vocab_bos(vocab); @@ -1479,20 +1479,20 @@ common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) { int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size())); if (ret != 0) { - LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret); + COM_ERR("llama_decode() failed: %d\n", ret); res = COMMON_CONTEXT_SEQ_RM_TYPE_NO; goto done; } if (llama_n_rs_seq(ctx) > 0) { - LOG_INF("%s: the context supports bounded partial sequence removal\n", __func__); + COM_TRC("%s", "the context supports bounded partial sequence removal\n"); res = COMMON_CONTEXT_SEQ_RM_TYPE_RS; goto done; } // try to remove the last tokens if (!llama_memory_seq_rm(mem, 0, 1, -1)) { - LOG_TRC("%s: the context does not support partial sequence removal\n", __func__); + COM_TRC("%s", "the context does not support partial sequence removal\n"); res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL; goto done; } @@ -1809,13 +1809,13 @@ static common_control_vector_data common_control_vector_load_one(const common_co }; struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params); if (!ctx_gguf) { - LOG_ERR("%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str()); + COM_ERR("failed to load control vector file from %s\n", load_info.fname.c_str()); return result; } int32_t n_tensors = gguf_get_n_tensors(ctx_gguf); if (n_tensors == 0) { - LOG_WRN("%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str()); + COM_WRN("no direction tensors found in %s\n", load_info.fname.c_str()); } for (int i = 0; i < n_tensors; i++) { @@ -1833,23 +1833,23 @@ static common_control_vector_data common_control_vector_load_one(const common_co } } if (layer_idx < 0) { - LOG_ERR("%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str()); + COM_ERR("invalid/unparsable direction tensor layer index in %s\n", load_info.fname.c_str()); result.n_embd = -1; break; } else if (layer_idx == 0) { - LOG_ERR("%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str()); + COM_ERR("invalid (zero) direction tensor layer index in %s\n", load_info.fname.c_str()); result.n_embd = -1; break; } struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str()); if (tensor->type != GGML_TYPE_F32) { - LOG_ERR("%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str()); + COM_ERR("invalid (non-F32) direction tensor type in %s\n", load_info.fname.c_str()); result.n_embd = -1; break; } if (ggml_n_dims(tensor) != 1) { - LOG_ERR("%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str()); + COM_ERR("invalid (non-1D) direction tensor shape in %s\n", load_info.fname.c_str()); result.n_embd = -1; break; } @@ -1857,7 +1857,7 @@ static common_control_vector_data common_control_vector_load_one(const common_co if (result.n_embd == -1) { result.n_embd = ggml_nelements(tensor); } else if (ggml_nelements(tensor) != result.n_embd) { - LOG_ERR("%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str()); + COM_ERR("direction tensor in %s does not match previous dimensions\n", load_info.fname.c_str()); result.n_embd = -1; break; } @@ -1874,7 +1874,7 @@ static common_control_vector_data common_control_vector_load_one(const common_co } if (result.n_embd == -1) { - LOG_WRN("%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str()); + COM_WRN("skipping %s due to invalid direction tensors\n", load_info.fname.c_str()); result.data.clear(); } @@ -1895,7 +1895,7 @@ common_control_vector_data common_control_vector_load(const std::vector(all_tokens.data() + offset), n_tokens_before_last))) { - LOG_ERR("%s : failed to eval\n", __func__); + COM_ERR("%s", "failed to eval\n"); return false; } n_past += n_tokens_before_last; llama_state_save_file(ctx, state_path.data(), all_tokens.data(), all_tokens.size()); - LOG_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size()); + COM_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size()); llama_token last_token = all_tokens.back(); llama_batch batch = llama_batch_get_one(&last_token, 1); @@ -2036,13 +2036,13 @@ bool common_prompt_batch_decode( batch.pos = &pos; if (llama_decode(ctx, batch)) { - LOG_ERR("%s : failed to eval last token\n", __func__); + COM_ERR("%s", "failed to eval last token\n"); return false; } n_past++; } else { if (llama_decode(ctx, llama_batch_get_one(const_cast(all_tokens.data() + offset), n_new))) { - LOG_ERR("%s : failed to eval\n", __func__); + COM_ERR("%s", "failed to eval\n"); return false; } n_past += n_new; diff --git a/common/common.h b/common/common.h index 819548103..7939d3b07 100644 --- a/common/common.h +++ b/common/common.h @@ -26,6 +26,13 @@ #define DIRECTORY_SEPARATOR '/' #endif // _WIN32 +#define COM_DBG(fmt, ...) LOG_DBG("cmn %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define COM_TRC(fmt, ...) LOG_TRC("cmn %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define COM_INF(fmt, ...) LOG_INF("cmn %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define COM_WRN(fmt, ...) LOG_WRN("cmn %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define COM_ERR(fmt, ...) LOG_ERR("cmn %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define COM_CNT(fmt, ...) LOG_CNT("" fmt, __VA_ARGS__) + #define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0) #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0) @@ -163,6 +170,7 @@ enum common_speculative_type { COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, // standalone draft model speculative decoding COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, // Eagle3 speculative decoding COMMON_SPECULATIVE_TYPE_DRAFT_MTP, // Multi-token prediction + COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, // DFlash speculative decoding COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, // simple self-speculative decoding based on n-grams COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, // self-speculative decoding with n-gram keys only COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values @@ -378,7 +386,7 @@ struct common_params_speculative { uint32_t need_n_rs_seq() const { bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) { - return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3; + return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH; }); return needs_rs_seq ? draft.n_max : 0u; diff --git a/common/fit.cpp b/common/fit.cpp index a8565bfc9..afbf0b10f 100644 --- a/common/fit.cpp +++ b/common/fit.cpp @@ -233,7 +233,7 @@ static void common_params_fit_impl( sum_projected_used = dmds_full.back().mb.total(); sum_free = dmds_full.back().total; sum_projected_free = sum_free - sum_projected_used; - LOG_INF("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n", + LOG_TRC("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n", __func__, sum_projected_used/MiB, sum_free/MiB); if (sum_projected_free >= margins[0]) { LOG_TRC("%s: will leave %" PRId64 " >= %" PRId64 " MiB of system memory, no changes needed\n", diff --git a/common/http.h b/common/http.h index d3daccd6b..e88bc6a5e 100644 --- a/common/http.h +++ b/common/http.h @@ -11,6 +11,11 @@ struct common_http_url { std::string path; }; +// bracket an IPv6 literal host for a URL authority (RFC 3986) +static std::string common_http_format_host(const std::string & host) { + return host.find(':') != std::string::npos ? "[" + host + "]" : host; +} + static common_http_url common_http_parse_url(const std::string & url) { common_http_url parts; auto scheme_end = url.find("://"); @@ -49,11 +54,28 @@ static common_http_url common_http_parse_url(const std::string & url) { parts.path = "/"; } - auto colon_pos = parts.host.find(':'); + // split the authority into host and optional port, a bracketed IPv6 literal keeps its inner colons (RFC 3986) + std::string port_str; + if (!parts.host.empty() && parts.host.front() == '[') { + auto close = parts.host.find(']'); + if (close == std::string::npos) { + throw std::runtime_error("invalid IPv6 URL authority: " + parts.host); + } + auto after = parts.host.substr(close + 1); + if (!after.empty() && after.front() == ':') { + port_str = after.substr(1); + } + parts.host = parts.host.substr(1, close - 1); + } else { + auto colon_pos = parts.host.find(':'); + if (colon_pos != std::string::npos) { + port_str = parts.host.substr(colon_pos + 1); + parts.host = parts.host.substr(0, colon_pos); + } + } - if (colon_pos != std::string::npos) { - parts.port = std::stoi(parts.host.substr(colon_pos + 1)); - parts.host = parts.host.substr(0, colon_pos); + if (!port_str.empty()) { + parts.port = std::stoi(port_str); } else if (parts.scheme == "http") { parts.port = 80; } else if (parts.scheme == "https") { @@ -83,7 +105,7 @@ static std::pair common_http_client(const std: } #endif - httplib::Client cli(parts.scheme + "://" + parts.host + ":" + std::to_string(parts.port)); + httplib::Client cli(parts.scheme + "://" + common_http_format_host(parts.host) + ":" + std::to_string(parts.port)); if (!parts.user.empty()) { cli.set_basic_auth(parts.user, parts.password); @@ -95,5 +117,5 @@ static std::pair common_http_client(const std: } static std::string common_http_show_masked_url(const common_http_url & parts) { - return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path; + return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + common_http_format_host(parts.host) + parts.path; } diff --git a/common/jinja/caps.cpp b/common/jinja/caps.cpp index ead864763..f7de8b806 100644 --- a/common/jinja/caps.cpp +++ b/common/jinja/caps.cpp @@ -9,6 +9,9 @@ #include #include +#ifdef FILENAME +#undef FILENAME +#endif #define FILENAME "jinja-caps" using json = nlohmann::ordered_json; @@ -16,22 +19,34 @@ using json = nlohmann::ordered_json; namespace jinja { using caps_json_fn = std::function; -using caps_analyze_fn = std::function; +using caps_ctx_fn = std::function; +using caps_analyze_fn = std::function; + +void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) { + ctx.set_val("preserve_thinking", mk_val(enabled)); + ctx.set_val("clear_thinking", mk_val(!enabled)); + ctx.set_val("truncate_history_thinking", mk_val(!enabled)); +} static void caps_try_execute(jinja::program & prog, const caps_json_fn & messages_fn, + const caps_ctx_fn & ctx_fn, const caps_json_fn & tools_fn, const caps_analyze_fn & analyze_fn) { context ctx; ctx.is_get_stats = true; jinja::global_from_json(ctx, json{ {"messages", messages_fn()}, - {"tools", tools_fn()}, + {"tools", tools_fn ? tools_fn() : json::array()}, {"bos_token", ""}, {"eos_token", ""}, {"add_generation_prompt", true} }, true); + if (ctx_fn) { + ctx_fn(ctx); + } + auto messages = ctx.get_val("messages"); auto tools = ctx.get_val("tools"); @@ -49,7 +64,7 @@ static void caps_try_execute(jinja::program & prog, // ignore exceptions during capability analysis } - analyze_fn(success, messages, tools); + analyze_fn(success, messages, tools, result); } // for debugging only @@ -109,11 +124,9 @@ caps caps_get(jinja::program & prog) { } }); }, - [&]() { - // tools - return json{nullptr}; - }, - [&](bool success, value & messages, value &) { + nullptr, // ctx_fn + nullptr, // tools_fn + [&](bool success, value & messages, value &, const std::string &) { auto & content = messages->at(0)->at("content"); caps_print_stats(content, "messages[0].content"); if (has_op(content, "selectattr") || has_op(content, "array_access")) { @@ -145,11 +158,9 @@ caps caps_get(jinja::program & prog) { }, }); }, - [&]() { - // tools - return json::array(); - }, - [&](bool, value & messages, value &) { + nullptr, // ctx_fn + nullptr, // tools_fn + [&](bool, value & messages, value &, const std::string &) { auto & content = messages->at(0)->at("content"); caps_print_stats(content, "messages[0].content"); if (!content->stats.used) { @@ -201,6 +212,7 @@ caps caps_get(jinja::program & prog) { }, }); }, + nullptr, // ctx_fn [&]() { // tools return json::array({ @@ -224,7 +236,7 @@ caps caps_get(jinja::program & prog) { }, }); }, - [&](bool success, value & messages, value & tools) { + [&](bool success, value & messages, value & tools, const std::string &) { if (!success) { return; // Nothing can be inferred } @@ -293,6 +305,7 @@ caps caps_get(jinja::program & prog) { }, }); }, + nullptr, // ctx_fn [&]() { // tools return json::array({ @@ -316,7 +329,7 @@ caps caps_get(jinja::program & prog) { }, }); }, - [&](bool success, value & messages, value & tools) { + [&](bool success, value & messages, value & tools, const std::string &) { if (!success) { result.supports_tool_calls = false; result.supports_tools = false; @@ -394,6 +407,7 @@ caps caps_get(jinja::program & prog) { }, }); }, + nullptr, // ctx_fn [&]() { // tools return json::array({ @@ -417,7 +431,7 @@ caps caps_get(jinja::program & prog) { }, }); }, - [&](bool success, value & messages, value & /*tools*/) { + [&](bool success, value & messages, value &, const std::string &) { if (!success) { result.supports_parallel_tool_calls = false; return; @@ -438,11 +452,22 @@ caps caps_get(jinja::program & prog) { JJ_DEBUG("%s\n", ">>> Running capability check: preserve reasoning"); // case: preserve reasoning content in chat history + const std::string reasoning_placeholder = ""; caps_try_execute( prog, [&]() { // messages return json::array({ + { + {"role", "user"}, + {"content", "User message"} + }, + { + {"role", "assistant"}, + {"content", "Assistant message"}, + // check of reasoning_content deeper in the history, not just the last assistant message + {"reasoning_content", reasoning_placeholder} + }, { {"role", "user"}, {"content", "User message"} @@ -458,14 +483,13 @@ caps caps_get(jinja::program & prog) { }, }); }, - [&]() { - // tools - return json::array(); + [&](context & ctx) { + caps_apply_preserve_reasoning(ctx, true); }, - [&](bool, value & messages, value &) { - auto & content = messages->at(1)->at("reasoning_content"); - caps_print_stats(content, "messages[1].reasoning_content"); - if (content->stats.used) { + nullptr, // tools_fn + [&](bool, value &, value &, const std::string & output) { + // note: we cannot use stats here because the reasoning_content may be used for "if" condition test, but not actually outputted in the final result + if (output.find(reasoning_placeholder) != std::string::npos) { result.supports_preserve_reasoning = true; } } diff --git a/common/jinja/caps.h b/common/jinja/caps.h index 93a7fe092..a290cd7da 100644 --- a/common/jinja/caps.h +++ b/common/jinja/caps.h @@ -12,7 +12,9 @@ struct caps { bool supports_tool_calls = true; bool supports_system_role = true; bool supports_parallel_tool_calls = true; - bool supports_preserve_reasoning = false; // support assistant message with reasoning_content + + // supports preserve reasoning trace in the full history, not just the last assistant message + bool supports_preserve_reasoning = false; // one of the 2 content capabilities must be true bool supports_string_content = true; @@ -29,4 +31,6 @@ struct caps { caps caps_get(jinja::program & prog); +void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled); + } // namespace jinja diff --git a/common/jinja/lexer.cpp b/common/jinja/lexer.cpp index 598982c2f..63740dff6 100644 --- a/common/jinja/lexer.cpp +++ b/common/jinja/lexer.cpp @@ -7,6 +7,9 @@ #include #include +#ifdef FILENAME +#undef FILENAME +#endif #define FILENAME "jinja-lexer" namespace jinja { diff --git a/common/jinja/parser.cpp b/common/jinja/parser.cpp index 2b25654a7..e9e6dcb7f 100644 --- a/common/jinja/parser.cpp +++ b/common/jinja/parser.cpp @@ -8,6 +8,9 @@ #include #include +#ifdef FILENAME +#undef FILENAME +#endif #define FILENAME "jinja-parser" namespace jinja { diff --git a/common/jinja/runtime.cpp b/common/jinja/runtime.cpp index f98cb0876..afba0025f 100644 --- a/common/jinja/runtime.cpp +++ b/common/jinja/runtime.cpp @@ -8,6 +8,9 @@ #include #include +#ifdef FILENAME +#undef FILENAME +#endif #define FILENAME "jinja-runtime" bool g_jinja_debug = false; @@ -954,4 +957,50 @@ value keyword_argument_expression::execute_impl(context & ctx) { return mk_val(k, v); } +std::string runtime::debug_dump_program(const program & prog, const std::string & src) { + std::ostringstream oss; + size_t lvl = 0; + context ctx; + ctx.src.reset(new std::string(src)); + + auto indent = [](size_t lvl) -> std::string { + return std::string(lvl * 2, ' '); + }; + + ctx.visitor = [&](bool is_leaf, statement * node, std::vector children) { + oss << indent(lvl) << node->type() << ":\n"; + lvl++; + if (is_leaf) { + const auto & pos = node->pos; + oss << indent(lvl) << "(leaf) at " << get_line_col(src, pos) << " in source:\n"; + std::string snippet = peak_source(src, pos); + string_replace_all(snippet, "\n", "\n" + indent(lvl)); + oss << indent(lvl) << snippet << "\n"; + } else { + for (auto & [label, children_vec] : children) { + oss << indent(lvl) << label << ":\n"; + lvl++; + if (children_vec.empty()) { + oss << indent(lvl) << "\n\n"; + } else { + for (auto * child : children_vec) { + if (!child) { + continue; + } + child->visit(ctx); + } + } + lvl--; + } + } + lvl--; + }; + + for (const auto & stmt : prog.body) { + stmt->visit(ctx); + } + + return oss.str(); +} + } // namespace jinja diff --git a/common/jinja/runtime.h b/common/jinja/runtime.h index 37b4c35ca..0884a1592 100644 --- a/common/jinja/runtime.h +++ b/common/jinja/runtime.h @@ -47,12 +47,19 @@ const T * cast_stmt(const statement_ptr & ptr) { // not thread-safe void enable_debug(bool enable); +// for visiting AST nodes +// function signature: void(bool is_leaf, statement * node, pair of ) +using visitor_pair = std::pair>; +using visitor_fn = std::function)>; + struct context { std::shared_ptr src; // for debugging; use shared_ptr to avoid copying on scope creation std::time_t current_time; // for functions that need current time bool is_get_stats = false; // whether to collect stats + visitor_fn visitor; + // src is optional, used for error reporting context(std::string src = "") : src(std::make_shared(std::move(src))) { env = mk_val(); @@ -99,6 +106,15 @@ private: value_object env; }; +// utils for visiting AST nodes +static std::vector stmts_to_ptr(const statements & stmts) { + std::vector children; + for (const auto & stmt : stmts) { + children.push_back(stmt.get()); + } + return children; +} + /** * Base class for all nodes in the AST. */ @@ -106,6 +122,7 @@ struct statement { size_t pos; // position in source, for debugging virtual ~statement() = default; virtual std::string type() const { return "Statement"; } + virtual void visit(context & ctx) { ctx.visitor(true, this, {}); } // execute_impl must be overridden by derived classes virtual value execute_impl(context &) { throw_exec_error(); } @@ -166,6 +183,13 @@ struct if_statement : public statement { std::string type() const override { return "If"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"test", {test.get()}}, + {"body", stmts_to_ptr(body)}, + {"alternate", stmts_to_ptr(alternate)} + }); + } }; struct identifier; @@ -190,6 +214,14 @@ struct for_statement : public statement { std::string type() const override { return "For"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"loopvar", {loopvar.get()}}, + {"iterable", {iterable.get()}}, + {"body", stmts_to_ptr(body)}, + {"default_block", stmts_to_ptr(default_block)} + }); + } }; struct break_statement : public statement { @@ -241,6 +273,13 @@ struct set_statement : public statement { std::string type() const override { return "Set"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"assignee", {assignee.get()}}, + {"value", {val.get()}}, + {"body", stmts_to_ptr(body)} + }); + } }; struct macro_statement : public statement { @@ -256,6 +295,13 @@ struct macro_statement : public statement { std::string type() const override { return "Macro"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"name", {name.get()}}, + {"args", stmts_to_ptr(args)}, + {"body", stmts_to_ptr(body)} + }); + } }; struct comment_statement : public statement { @@ -289,6 +335,12 @@ struct member_expression : public expression { } std::string type() const override { return "MemberExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"object", {object.get()}}, + {"property", {property.get()}} + }); + } }; struct call_expression : public expression { @@ -302,6 +354,12 @@ struct call_expression : public expression { } std::string type() const override { return "CallExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"callee", {callee.get()}}, + {"args", stmts_to_ptr(args)} + }); + } }; /** @@ -405,6 +463,12 @@ struct binary_expression : public expression { } std::string type() const override { return "BinaryExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"left", {left.get()}}, + {"right", {right.get()}} + }); + } }; /** @@ -431,6 +495,12 @@ struct filter_expression : public expression { std::string type() const override { return "FilterExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"operand", {operand.get()}}, + {"filter", {filter.get()}} + }); + } }; struct filter_statement : public statement { @@ -443,6 +513,12 @@ struct filter_statement : public statement { } std::string type() const override { return "FilterStatement"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"filter", {filter.get()}}, + {"body", stmts_to_ptr(body)} + }); + } }; /** @@ -468,6 +544,12 @@ struct select_expression : public expression { } return lhs->execute_impl(ctx); } + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"lhs", {lhs.get()}}, + {"test", {test.get()}} + }); + } }; /** @@ -486,6 +568,12 @@ struct test_expression : public expression { } std::string type() const override { return "TestExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"operand", {operand.get()}}, + {"test", {test.get()}} + }); + } }; /** @@ -501,6 +589,11 @@ struct unary_expression : public expression { } std::string type() const override { return "UnaryExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"argument", {argument.get()}} + }); + } }; struct slice_expression : public expression { @@ -518,6 +611,13 @@ struct slice_expression : public expression { [[noreturn]] value execute_impl(context &) override { throw std::runtime_error("must be handled by MemberExpression"); } + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"start_expr", {start_expr.get()}}, + {"stop_expr", {stop_expr.get()}}, + {"step_expr", {step_expr.get()}} + }); + } }; struct keyword_argument_expression : public expression { @@ -531,6 +631,12 @@ struct keyword_argument_expression : public expression { } std::string type() const override { return "KeywordArgumentExpression"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"key", {key.get()}}, + {"val", {val.get()}} + }); + } }; struct spread_expression : public expression { @@ -539,6 +645,11 @@ struct spread_expression : public expression { chk_type(this->argument); } std::string type() const override { return "SpreadExpression"; } + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"argument", {argument.get()}} + }); + } }; struct call_statement : public statement { @@ -553,6 +664,13 @@ struct call_statement : public statement { } std::string type() const override { return "CallStatement"; } value execute_impl(context & ctx) override; + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"call", {call.get()}}, + {"caller_args", stmts_to_ptr(caller_args)}, + {"body", stmts_to_ptr(body)} + }); + } }; struct ternary_expression : public expression { @@ -575,6 +693,13 @@ struct ternary_expression : public expression { return false_expr->execute(ctx); } } + void visit(context & ctx) override { + ctx.visitor(false, this, { + {"condition", {condition.get()}}, + {"true_expr", {true_expr.get()}}, + {"false_expr", {false_expr.get()}} + }); + } }; struct raised_exception : public std::exception { @@ -648,6 +773,8 @@ struct runtime { } return parts; } + + static std::string debug_dump_program(const program & prog, const std::string & src); }; } // namespace jinja diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp index 189451620..f36bbbc0c 100644 --- a/common/jinja/value.cpp +++ b/common/jinja/value.cpp @@ -12,6 +12,9 @@ #include #include +#ifdef FILENAME +#undef FILENAME +#endif #define FILENAME "jinja-value" namespace jinja { @@ -1108,6 +1111,50 @@ const func_builtins & value_array_t::get_builtins() const { std::reverse(arr.begin(), arr.end()); return is_val(val) ? mk_val(std::move(arr)) : mk_val(std::move(arr)); }}, + {"min", [](const func_args & args) -> value { + args.ensure_count(1, 4); + args.ensure_vals(); + value val_case = args.get_kwarg_or_pos("case_sensitive", 1); + value attribute = args.get_kwarg_or_pos("attribute", 2); + if (!attribute->is_undefined()) { + throw not_implemented_exception("min: attribute not implemented"); + } + // FIXME: min is currently always case sensitive + (void) val_case; + const auto & arr = args.get_pos(0)->as_array(); + if (arr.empty()) { + return mk_val(); + } + value result = arr[0]; + for (size_t i = 1; i < arr.size(); ++i) { + if (value_compare(arr[i], result, value_compare_op::lt)) { + result = arr[i]; + } + } + return result; + }}, + {"max", [](const func_args & args) -> value { + args.ensure_count(1, 4); + args.ensure_vals(); + value val_case = args.get_kwarg_or_pos("case_sensitive", 1); + value attribute = args.get_kwarg_or_pos("attribute", 2); + if (!attribute->is_undefined()) { + throw not_implemented_exception("max: attribute not implemented"); + } + // FIXME: max is currently always case sensitive + (void) val_case; + const auto & arr = args.get_pos(0)->as_array(); + if (arr.empty()) { + return mk_val(); + } + value result = arr[0]; + for (size_t i = 1; i < arr.size(); ++i) { + if (value_compare(arr[i], result, value_compare_op::gt)) { + result = arr[i]; + } + } + return result; + }}, {"unique", array_unique_not_implemented}, }; return builtins; diff --git a/common/preset.cpp b/common/preset.cpp index f0cc1fa1a..4362c0621 100644 --- a/common/preset.cpp +++ b/common/preset.cpp @@ -7,6 +7,7 @@ #include #include #include +#include static std::string rm_leading_dashes(const std::string & str) { size_t pos = 0; @@ -16,6 +17,23 @@ static std::string rm_leading_dashes(const std::string & str) { return str.substr(pos); } +static std::string canonical_tag(const std::string & tag) { + static const std::regex re_tag("[-.]([A-Z0-9_]+)$", std::regex::icase); + std::smatch m; + if (std::regex_search(tag, m, re_tag)) { + std::string canon = m[1].str(); + for (char & c : canon) { + c = (char) std::toupper((unsigned char) c); + } + return canon; + } + std::string upper = tag; + for (char & c : upper) { + c = (char) std::toupper((unsigned char) c); + } + return upper; +} + std::vector common_preset::to_args(const std::string & bin_path) const { std::vector args; @@ -270,11 +288,18 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co for (auto section : ini_data) { common_preset preset; - if (section.first.empty()) { - preset.name = COMMON_PRESET_DEFAULT_NAME; - } else { - preset.name = section.first; + std::string section_name = section.first.empty() ? std::string(COMMON_PRESET_DEFAULT_NAME) : section.first; + if (section_name != "*" && section_name != COMMON_PRESET_DEFAULT_NAME) { + auto colon_idx = section_name.rfind(':'); + if (colon_idx != std::string::npos) { + std::string tag = section_name.substr(colon_idx + 1); + std::string canon_tag = canonical_tag(tag); + if (canon_tag != tag) { + section_name = section_name.substr(0, colon_idx + 1) + canon_tag; + } + } } + preset.name = section_name; LOG_DBG("loading preset: %s\n", preset.name.c_str()); for (const auto & [key, value] : section.second) { if (key == "version") { diff --git a/common/reasoning-budget.cpp b/common/reasoning-budget.cpp index ce41d029b..7da0bb1c5 100644 --- a/common/reasoning-budget.cpp +++ b/common/reasoning-budget.cpp @@ -65,12 +65,12 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to if (ctx->start_matcher.advance(token)) { ctx->state = REASONING_BUDGET_COUNTING; ctx->remaining = ctx->budget; - LOG_INF("reasoning-budget: activated, budget=%d tokens\n", ctx->budget); + COM_TRC("activated, budget=%d tokens\n", ctx->budget); if (ctx->remaining <= 0) { ctx->state = REASONING_BUDGET_FORCING; ctx->force_pos = 0; - LOG_INF("reasoning-budget: budget=0, forcing immediately\n"); + COM_TRC("%s", "budget=0, forcing immediately\n"); } } break; @@ -80,7 +80,7 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to { if (ctx->end_matcher.advance(token)) { ctx->state = REASONING_BUDGET_DONE; - LOG_INF("reasoning-budget: deactivated (natural end)\n"); + COM_TRC("%s", "deactivated (natural end)\n"); break; } @@ -95,7 +95,7 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to ctx->state = REASONING_BUDGET_FORCING; ctx->force_pos = 0; ctx->end_matcher.reset(); - LOG_INF("reasoning-budget: UTF-8 complete, now forcing end sequence\n"); + COM_TRC("%s", "UTF-8 complete, now forcing end sequence\n"); } } else if (ctx->state == REASONING_BUDGET_COUNTING) { ctx->remaining--; @@ -104,11 +104,11 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to ctx->state = REASONING_BUDGET_FORCING; ctx->force_pos = 0; ctx->end_matcher.reset(); - LOG_INF("reasoning-budget: budget exhausted, forcing end sequence\n"); + COM_TRC("%s", "budget exhausted, forcing end sequence\n"); } else { ctx->state = REASONING_BUDGET_WAITING_UTF8; ctx->end_matcher.reset(); - LOG_INF("reasoning-budget: budget exhausted, waiting for UTF-8 completion\n"); + COM_TRC("%s", "budget exhausted, waiting for UTF-8 completion\n"); } } } @@ -118,7 +118,7 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to ctx->force_pos++; if (ctx->force_pos >= ctx->forced_tokens.size()) { ctx->state = REASONING_BUDGET_DONE; - LOG_INF("reasoning-budget: forced sequence complete, done\n"); + COM_TRC("%s", "forced sequence complete, done\n"); } break; case REASONING_BUDGET_DONE: @@ -128,12 +128,12 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to ctx->state = REASONING_BUDGET_COUNTING; ctx->remaining = ctx->budget; ctx->end_matcher.reset(); - LOG_INF("reasoning-budget: re-activated on new start tag, budget=%d tokens\n", ctx->budget); + COM_TRC("re-activated on new start tag, budget=%d tokens\n", ctx->budget); if (ctx->remaining <= 0) { ctx->state = REASONING_BUDGET_FORCING; ctx->force_pos = 0; - LOG_INF("reasoning-budget: budget=0, forcing immediately\n"); + COM_TRC("%s", "budget=0, forcing immediately\n"); } } break; @@ -264,7 +264,7 @@ bool common_reasoning_budget_force(struct llama_sampler * smpl) { ctx->state = REASONING_BUDGET_FORCING; ctx->force_pos = 0; ctx->end_matcher.reset(); - LOG_INF("reasoning-budget: forced into forcing state (manual transition)\n"); + COM_TRC("%s", "forced into forcing state (manual transition)\n"); return true; } diff --git a/common/regex-partial.cpp b/common/regex-partial.cpp deleted file mode 100644 index bd9034e93..000000000 --- a/common/regex-partial.cpp +++ /dev/null @@ -1,204 +0,0 @@ -#include "regex-partial.h" -#include "common.h" -#include -#include - -common_regex::common_regex(const std::string & pattern) : - pattern(pattern), - rx(pattern), - rx_reversed_partial(regex_to_reversed_partial_regex(pattern)) {} - -common_regex_match common_regex::search(const std::string & input, size_t pos, bool as_match) const { - std::smatch match; - if (pos > input.size()) { - throw std::runtime_error("Position out of bounds"); - } - auto start = input.begin() + pos; - auto found = as_match - ? std::regex_match(start, input.end(), match, rx) - : std::regex_search(start, input.end(), match, rx); - if (found) { - common_regex_match res; - res.type = COMMON_REGEX_MATCH_TYPE_FULL; - for (size_t i = 0; i < match.size(); ++i) { - auto begin = pos + match.position(i); - res.groups.emplace_back(begin, begin + match.length(i)); - } - return res; - } - std::match_results srmatch; - if (std::regex_search(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial, std::regex_constants::match_continuous)) { - auto group = srmatch[1].str(); - if (group.length() != 0) { - auto it = srmatch[1].second.base(); - // auto position = static_cast(std::distance(input.begin(), it)); - if ((!as_match) || it == input.begin()) { - common_regex_match res; - res.type = COMMON_REGEX_MATCH_TYPE_PARTIAL; - const size_t begin = std::distance(input.begin(), it); - const size_t end = input.size(); - if (begin == std::string::npos || end == std::string::npos || begin > end) { - throw std::runtime_error("Invalid range"); - } - res.groups.push_back({begin, end}); - return res; - } - } - } - return {}; -} - -/* - Transforms a regex pattern to a partial match pattern that operates on a reversed input string to find partial final matches of the original pattern. - - Ideally we'd like to use boost::match_partial (https://beta.boost.org/doc/libs/1_59_0/libs/regex/doc/html/boost_regex/partial_matches.html) - to see if a string ends with a partial regex match, but but it's not in std::regex yet. - Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input. - - - /abcd/ -> ^(dcba|cba|ba|a) -> ^((?:(?:(?:(?:d)?c)?b)?a) - - /a|b/ -> ^(a|b) - - /a*?/ -> error, could match "" - - /a*b/ -> ^((?:b)?a*+) (final repetitions become eager) - - /.*?ab/ -> ^((?:b)?a) (omit .*) - - /a.*?b/ -> ^((?:b)?.*?a) (keep reluctant matches) - - /a(bc)d/ -> ^((?:(?:d)?(?:(?:c)?b))?a) - - /a(bc|de)/ -> ^((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a) - - /ab{2,4}c/ -> ^cbbb?b?a -> ^((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a) - - The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern. - All other groups are turned into non-capturing groups, and reluctant quantifiers are ignored. -*/ -std::string regex_to_reversed_partial_regex(const std::string & pattern) { - auto it = pattern.begin(); - const auto end = pattern.end(); - - std::function process = [&]() { - std::vector> alternatives(1); - std::vector * sequence = &alternatives.back(); - - while (it != end) { - if (*it == '[') { - auto start = it; - ++it; - while (it != end) { - if ((*it == '\\') && (++it != end)) { - ++it; - } else if ((it != end) && (*it == ']')) { - break; - } else { - ++it; - } - } - if (it == end) { - throw std::runtime_error("Unmatched '[' in pattern"); - } - ++it; - sequence->push_back(std::string(start, it)); - } else if (*it == '*' || *it == '?' || *it == '+') { - if (sequence->empty()) { - throw std::runtime_error("Quantifier without preceding element"); - } - sequence->back() += *it; - auto is_star = *it == '*'; - ++it; - if (is_star) { - if (it != end && *it == '?') { - ++it; - } - } - } else if (*it == '{') { - if (sequence->empty()) { - throw std::runtime_error("Repetition without preceding element"); - } - ++it; - auto start = it; - while (it != end && *it != '}') { - ++it; - } - if (it == end) { - throw std::runtime_error("Unmatched '{' in pattern"); - } - auto parts = string_split(std::string(start, it), ","); - ++it; - if (parts.size() > 2) { - throw std::runtime_error("Invalid repetition range in pattern"); - } - - auto parseOptInt = [&](const std::string & s, const std::optional & def = std::nullopt) -> std::optional { - if (s.empty()) { - return def; - } - return std::stoi(s); - }; - auto min = parseOptInt(parts[0], 0); - auto max = parts.size() == 1 ? min : parseOptInt(parts[1]); - if (min && max && *max < *min) { - throw std::runtime_error("Invalid repetition range in pattern"); - } - // Brutal but... let's repeat at least min times, then ? for the delta between min & max (or * for unbounded) - auto part = sequence->back(); - sequence->pop_back(); - for (int i = 0; i < *min; i++) { - sequence->push_back(part); - } - if (max) { - for (int i = *min; i < *max; i++) { - sequence->push_back(part + "?"); - } - } else { - sequence->push_back(part + "*"); - } - } else if (*it == '(') { - ++it; - if (it != end && *it == '?' && (it + 1 != end) && *(it + 1) == ':') { - it += 2; - } - auto sub = process(); - if (*it != ')') { - throw std::runtime_error("Unmatched '(' in pattern"); - } - ++it; - auto & part = sequence->emplace_back("(?:"); - part += sub; - part += ")"; - } else if (*it == ')') { - break; - } else if (*it == '|') { - ++it; - alternatives.emplace_back(); - sequence = &alternatives.back(); - } else if (*it == '\\' && (++it != end)) { - auto str = std::string("\\") + *it; - sequence->push_back(str); - ++it; - } else if (it != end) { - sequence->push_back(std::string(1, *it)); - ++it; - } - } - - // /abcd/ -> ^(dcba|cba|ba|a) -> ^((?:(?:(?:d)?c)?b)?a) - // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group - // We'll do the outermost capturing group and final .* in the enclosing function. - std::vector res_alts; - for (const auto & parts : alternatives) { - auto & res = res_alts.emplace_back(); - for (size_t i = 0; i < parts.size() - 1; i++) { - res += "(?:"; - } - for (auto it = parts.rbegin(); it != parts.rend(); ++it) { - res += *it; - if (it != parts.rend() - 1) { - res += ")?"; - } - } - } - return string_join(res_alts, "|"); - }; - auto res = process(); - if (it != end) { - throw std::runtime_error("Unmatched '(' in pattern"); - } - - return "^(" + res + ")"; -} diff --git a/common/regex-partial.h b/common/regex-partial.h deleted file mode 100644 index 634cb4022..000000000 --- a/common/regex-partial.h +++ /dev/null @@ -1,56 +0,0 @@ -#pragma once - -#include -#include - -enum common_regex_match_type { - COMMON_REGEX_MATCH_TYPE_NONE, - COMMON_REGEX_MATCH_TYPE_PARTIAL, - COMMON_REGEX_MATCH_TYPE_FULL, -}; - -struct common_string_range { - size_t begin; - size_t end; - common_string_range(size_t begin, size_t end) : begin(begin), end(end) { - if (begin > end) { - throw std::runtime_error("Invalid range"); - } - } - // prevent default ctor - common_string_range() = delete; - bool empty() const { - return begin == end; - } - bool operator==(const common_string_range & other) const { - return begin == other.begin && end == other.end; - } -}; - -struct common_regex_match { - common_regex_match_type type = COMMON_REGEX_MATCH_TYPE_NONE; - std::vector groups; - - bool operator==(const common_regex_match & other) const { - return type == other.type && groups == other.groups; - } - bool operator!=(const common_regex_match & other) const { - return !(*this == other); - } -}; - -class common_regex { - std::string pattern; - std::regex rx; - std::regex rx_reversed_partial; - - public: - explicit common_regex(const std::string & pattern); - - common_regex_match search(const std::string & input, size_t pos, bool as_match = false) const; - - const std::string & str() const { return pattern; } -}; - -// For testing only (pretty print of failures). -std::string regex_to_reversed_partial_regex(const std::string & pattern); diff --git a/common/speculative.cpp b/common/speculative.cpp index 0a293edeb..4e0439216 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -18,6 +18,13 @@ #include #include +#define SPC_DBG(fmt, ...) LOG_DBG("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define SPC_TRC(fmt, ...) LOG_TRC("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define SPC_INF(fmt, ...) LOG_INF("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define SPC_WRN(fmt, ...) LOG_WRN("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define SPC_ERR(fmt, ...) LOG_ERR("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__) +#define SPC_CNT(fmt, ...) LOG_CNT("" fmt, __VA_ARGS__) + #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5 @@ -26,6 +33,7 @@ const std::map common_speculative_type_fro {"draft-simple", COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE}, {"draft-eagle3", COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3}, {"draft-mtp", COMMON_SPECULATIVE_TYPE_DRAFT_MTP}, + {"draft-dflash", COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH}, {"ngram-simple", COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE}, {"ngram-map-k", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K}, {"ngram-map-k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V}, @@ -60,21 +68,20 @@ static bool common_speculative_are_compatible( const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft); const auto vocab_type_tgt = llama_vocab_type(vocab_tgt); - LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt); + SPC_DBG("vocab_type tgt: %d\n", vocab_type_tgt); const auto vocab_type_dft = llama_vocab_type(vocab_dft); - LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft); + SPC_DBG("vocab_type dft: %d\n", vocab_type_dft); if (vocab_type_tgt != vocab_type_dft) { - LOG_WRN("%s: draft model vocab type must match target model to use speculation but " - "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt); + SPC_WRN("draft model vocab type must match target model to use speculation but " + "vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt); return false; } if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) || (llama_vocab_get_add_bos(vocab_tgt) && llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft))) { - LOG_WRN("%s: draft model bos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n", - __func__, + SPC_WRN("draft model bos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n", llama_vocab_get_add_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_dft), llama_vocab_bos(vocab_tgt), llama_vocab_bos(vocab_dft)); return false; @@ -82,8 +89,7 @@ static bool common_speculative_are_compatible( if (llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) || (llama_vocab_get_add_eos(vocab_tgt) && llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft))) { - LOG_WRN("%s: draft model eos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n", - __func__, + SPC_WRN("draft model eos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n", llama_vocab_get_add_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_dft), llama_vocab_eos(vocab_tgt), llama_vocab_eos(vocab_dft)); return false; @@ -97,8 +103,8 @@ static bool common_speculative_are_compatible( : n_vocab_dft - n_vocab_tgt; if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) { - LOG_DBG("%s: draft model vocab must closely match target model to use speculation but ", __func__); - LOG_DBG("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n", + SPC_DBG("draft model vocab must closely match target model to use speculation but " + "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n", n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE); return false; } @@ -108,8 +114,8 @@ static bool common_speculative_are_compatible( const char * token_text_dft = llama_vocab_get_text(vocab_dft, i); if (std::strcmp(token_text_tgt, token_text_dft) != 0) { - LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__); - LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i, + SPC_DBG("draft model vocab must match target model to use speculation but " + "token %d content differs - target '%s', draft '%s'\n", i, common_token_to_piece(vocab_tgt, i).c_str(), common_token_to_piece(vocab_dft, i).c_str()); return false; @@ -186,9 +192,9 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { auto * ctx_dft = this->params.ctx_dft; auto * ctx_tgt = this->params.ctx_tgt; - LOG_INF("%s: adding speculative implementation 'draft-simple'\n", __func__); - LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min); - LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__, + SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n"); + SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min); + SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", this->params.n_gpu_layers, ggml_type_name(this->params.cache_type_k), ggml_type_name(this->params.cache_type_v), @@ -228,16 +234,16 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { } const bool vocab_cmpt = common_speculative_are_compatible(llama_get_model(ctx_tgt), llama_get_model(ctx_dft)); - LOG_DBG("%s: vocab_cmpt = %d\n", __func__, vocab_cmpt); + SPC_DBG("vocab_cmpt = %d\n", vocab_cmpt); if (!vocab_cmpt) { - LOG_ERR("%s: the target and draft vocabs are not compatible\n", __func__); + SPC_ERR("%s", "the target and draft vocabs are not compatible\n"); throw std::runtime_error("draft model vocab type must match target model to use speculation"); } if (n_seq != llama_n_seq_max(ctx_dft)) { - LOG_ERR("%s: n_seq mismatch: %d != %d\n", __func__, n_seq, llama_n_seq_max(ctx_dft)); + SPC_ERR("n_seq mismatch: %d != %d\n", n_seq, llama_n_seq_max(ctx_dft)); throw std::runtime_error("the draft model number of sequences is incompatible with the speculative n_seq"); } @@ -257,7 +263,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { const int ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_ERR("%s: failed to decode draft batch, ret = %d\n", __func__, ret); + SPC_ERR("failed to decode draft batch, ret = %d\n", ret); return false; } @@ -290,7 +296,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { int ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_WRN("%s: llama_decode returned %d\n", __func__, ret); + SPC_ERR("llama_decode returned %d\n", ret); return; } @@ -314,7 +320,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { const auto * cur_p = common_sampler_get_candidates(smpl, true); for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { - LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", + SPC_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); } @@ -354,7 +360,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { // evaluate the drafted tokens on the draft model ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret); + SPC_ERR("llama_decode[%d] returned %d\n", i, ret); break; } @@ -449,8 +455,8 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq) , params(params.draft) { - LOG_INF("%s: adding speculative implementation 'draft-eagle3'\n", __func__); - LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f, backend_sampling=%d\n", __func__, params.draft.n_max, params.draft.n_min, params.draft.p_min, (int) params.draft.backend_sampling); + SPC_TRC("%s", "adding speculative implementation 'draft-eagle3'\n"); + SPC_TRC("- n_max=%d, n_min=%d, p_min=%f, backend_sampling=%d\n", params.draft.n_max, params.draft.n_min, params.draft.p_min, (int) params.draft.backend_sampling); auto * ctx_tgt = this->params.ctx_tgt; auto * ctx_dft = this->params.ctx_dft; @@ -493,7 +499,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { llama_sampler_chain_add(chain, llama_sampler_init_top_k(10)); if (!llama_set_sampler(ctx_dft, seq_id, chain)) { - LOG_WRN("%s: backend offload failed for seq_id=%d; using CPU sampler\n", __func__, (int) seq_id); + SPC_WRN("backend offload failed for seq_id=%d; using CPU sampler\n", (int) seq_id); llama_sampler_free(chain); chain = nullptr; } @@ -548,9 +554,9 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { auto * ctx_dft = this->params.ctx_dft; const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id); if (pos_max < N - 2) { - LOG_WRN("%s: ctx_dft pos_max=%d < N-2=%d — process() did not run on every prefill ubatch. " + SPC_WRN("ctx_dft pos_max=%d < N-2=%d — process() did not run on every prefill ubatch. " "Drafts may degrade.\n", - __func__, (int) pos_max, N - 2); + (int) pos_max, N - 2); } } @@ -621,8 +627,8 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { }; const int32_t rc = llama_encode(ctx_dft, enc_batch); if (rc != 0) { - LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n", - __func__, rc, (int) n_chunk, (int) i); + SPC_ERR("llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n", + rc, (int) n_chunk, (int) i); return false; } @@ -692,8 +698,8 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { if (batch.n_tokens > 0) { const int32_t rc = llama_decode(ctx_dft, batch); if (rc != 0) { - LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, ubatch_pos[0]=%d)\n", - __func__, rc, (int) batch.n_tokens, (int) batch_in.pos[0]); + SPC_ERR("llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, ubatch_pos[0]=%d)\n", + rc, (int) batch.n_tokens, (int) batch_in.pos[0]); return false; } } @@ -744,7 +750,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { int ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_WRN("%s: llama_decode returned %d\n", __func__, ret); + SPC_ERR("llama_decode returned %d\n", ret); return; } @@ -770,7 +776,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { const auto * cur_p = common_sampler_get_candidates(smpl, true); for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { - LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", + SPC_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); } @@ -809,7 +815,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret); + SPC_ERR("llama_decode[%d] returned %d\n", i, ret); break; } @@ -893,6 +899,305 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { } }; +// DFlash: block-diffusion drafting with a draft-side KV cache injection +struct common_speculative_impl_draft_dflash : public common_speculative_impl { + common_params_speculative_draft params; + + llama_batch batch; // noise tokens + llama_batch batch_inject; // target features for KV cache injection + + std::vector smpls; + + int32_t n_embd_dec = 0; // draft hidden size + int32_t n_embd_enc = 0; // target_layer_ids_n * target_hidden_size + int32_t n_embd_tgt = 0; // target model hidden size + + int32_t block_size = 0; + llama_token mask_token_id = 0; + + const int32_t * target_layer_ids = nullptr; // model_dft's extract layer indices + uint32_t target_layer_ids_n = 0; + + // scratch buffer for concatenated target features [n_tokens, n_embd_enc] + std::vector features_buf; + + common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, n_seq) + , params(params.draft) + { + auto * ctx_tgt = this->params.ctx_tgt; + auto * ctx_dft = this->params.ctx_dft; + GGML_ASSERT(ctx_tgt && ctx_dft && "DFlash requires ctx_tgt and ctx_dft to be set"); + + const llama_model * model_dft = llama_get_model(ctx_dft); + const llama_model * model_tgt = llama_get_model(ctx_tgt); + + target_layer_ids = llama_model_target_layer_ids (model_dft); + target_layer_ids_n = llama_model_target_layer_ids_n(model_dft); + GGML_ASSERT(target_layer_ids_n > 0 && "DFlash model has no target_layer_ids"); + + n_embd_tgt = llama_model_n_embd(model_tgt); + n_embd_dec = llama_model_n_embd(model_dft); + n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt; + + // read the trained block size from the dflash.block_size metadata key + block_size = 16; + { + char buf[32] = {}; + if (llama_model_meta_val_str(model_dft, "dflash.block_size", buf, sizeof(buf)) >= 0) { + block_size = std::atoi(buf); + } + } + mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft)); + + LOG_INF("%s: adding speculative implementation 'draft-dflash'\n", __func__); + LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min); + LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n); + + // DFlash input is [id_last, * (block_size-1)], so it can draft at most block_size-1 tokens per step + if (this->params.n_max > block_size - 1 || this->params.n_min > block_size - 1) { + LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained DFlash block size %d -- clamping to %d\n", + __func__, this->params.n_max, this->params.n_min, block_size, block_size - 1); + this->params.n_max = std::min(this->params.n_max, block_size - 1); + this->params.n_min = std::min(this->params.n_min, block_size - 1); + } + + batch = llama_batch_init(llama_n_batch(ctx_dft), 0, n_seq); + batch_inject = llama_batch_init(llama_n_batch(ctx_dft), n_embd_dec, n_seq); + + smpls.resize(n_seq); + for (auto & s : smpls) { + common_params_sampling sparams; + sparams.no_perf = false; + sparams.top_k = 10; + sparams.samplers = { COMMON_SAMPLER_TYPE_TOP_K }; + s.reset(common_sampler_init(model_dft, sparams)); + } + + // turn on extraction of the target layers' input embeddings + for (uint32_t k = 0; k < target_layer_ids_n; ++k) { + llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true); + } + + llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true); + llama_set_causal_attn(ctx_dft, false); // DFlash needs non-causal attention + } + + ~common_speculative_impl_draft_dflash() override { + llama_batch_free(batch); + llama_batch_free(batch_inject); + } + + void begin(llama_seq_id seq_id, const llama_tokens & prompt) override { + if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) { + return; + } + + const int32_t N = (int32_t) prompt.size(); + if (N <= 0) { + return; + } + + const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(params.ctx_dft), seq_id); + if (pos_max < N - 1) { + LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - process() did not run on every prefill ubatch. " + "Drafts may degrade.\n", + __func__, (int) pos_max, N - 1); + } + } + + bool process(const llama_batch & batch_in) override { + if (batch_in.n_tokens <= 0) { + return true; + } + + if (batch_in.token == nullptr || batch_in.embd != nullptr) { + return true; + } + + const int32_t n_tokens = batch_in.n_tokens; + + // per-seq inclusive batch range (assumes each seq's tokens are contiguous in the batch) + std::vector i_batch_beg(n_seq, -1); + std::vector i_batch_end(n_seq, -1); + for (int32_t k = 0; k < n_tokens; ++k) { + GGML_ASSERT(batch_in.n_seq_id[k] == 1); + const llama_seq_id seq_id = batch_in.seq_id[k][0]; + if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) { + continue; + } + i_batch_end[seq_id] = k; + if (i_batch_beg[seq_id] < 0) { + i_batch_beg[seq_id] = k; + } + } + + auto * ctx_tgt = this->params.ctx_tgt; + auto * ctx_dft = this->params.ctx_dft; + + const int32_t n_ubatch = (int32_t) llama_n_ubatch(ctx_dft); + + for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) { + if (i_batch_beg[seq_id] < 0) { + continue; + } + const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1; + + for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) { + const int32_t n_chunk = std::min(n_ubatch, n_rows - offset); + + // gather this chunk's target features, interleaved by extract layer + features_buf.resize((size_t) n_chunk * n_embd_enc); + for (uint32_t k = 0; k < target_layer_ids_n; ++k) { + const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]); + if (!layer) { + GGML_ABORT("DFlash: target layer %d input not extracted.", target_layer_ids[k]); + } + for (int32_t i = 0; i < n_chunk; ++i) { + float * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt; + const float * src = layer + (size_t) (i_batch_beg[seq_id] + offset + i) * n_embd_tgt; + std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float)); + } + } + + // fuse extracted features through DFlash encoder + llama_batch enc_batch = { + /*.n_tokens =*/ n_chunk, + /*.token =*/ nullptr, + /*.embd =*/ features_buf.data(), + /*.pos =*/ nullptr, + /*.n_seq_id =*/ nullptr, + /*.seq_id =*/ nullptr, + /*.logits =*/ nullptr, + }; + + int32_t rc = llama_encode(ctx_dft, enc_batch); + if (rc != 0) { + LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n", + __func__, rc, (int) n_chunk, (int) offset); + return false; + } + + const float * inp_g = llama_get_embeddings_nextn(ctx_dft); + GGML_ASSERT(inp_g && "DFlash encoder produced no output."); + + // inject the DFlash decoder K/V cache at the tokens' target positions + batch_inject.n_tokens = n_chunk; + std::memcpy(batch_inject.embd, inp_g, (size_t) n_chunk * n_embd_dec * sizeof(float)); + + for (int32_t i = 0; i < n_chunk; ++i) { + batch_inject.pos[i] = batch_in.pos[i_batch_beg[seq_id] + offset + i]; + batch_inject.n_seq_id[i] = 1; + batch_inject.seq_id[i][0] = seq_id; + batch_inject.logits[i] = false; + } + rc = llama_decode(ctx_dft, batch_inject); + if (rc != 0) { + LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n", + __func__, rc, (int) n_chunk, (int) offset); + return false; + } + } + } + + return true; + } + + void draft(common_speculative_draft_params_vec & dparams) override { + auto & ctx_dft = params.ctx_dft; + + common_batch_clear(batch); + + // build one batch holding every drafting sequence's noise block into a single decode) + // record where each block starts and its size + std::vector i_block_beg(n_seq, -1); + std::vector n_block (n_seq, 0); + + for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) { + auto & dp = dparams[seq_id]; + if (!dp.drafting) { + continue; + } + + common_sampler_reset(smpls[seq_id].get()); + + const int32_t n = (int32_t) dp.n_past; + + int32_t n_draft = params.n_max; + if (dp.n_max > 0) { + n_draft = std::min(n_draft, dp.n_max); + } + + const int32_t n_block_tokens = n_draft + 1; // id_last + n_draft * + i_block_beg[seq_id] = batch.n_tokens; + n_block [seq_id] = n_block_tokens; + for (int32_t i = 0; i < n_block_tokens; ++i) { + common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, true); + } + } + + if (batch.n_tokens == 0) { + return; + } + + // decode all sequence's noise block in a single batch + int ret = llama_decode(ctx_dft, batch); + if (ret != 0) { + LOG_WRN("%s: llama_decode returned %d\n", __func__, ret); + return; + } + + for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) { + if (i_block_beg[seq_id] < 0) { + continue; + } + auto & dp = dparams[seq_id]; + + const int32_t beg = i_block_beg[seq_id]; + const int32_t n_block_tokens = n_block[seq_id]; + + auto * smpl = smpls[seq_id].get(); + + auto & result = *dp.result; + + // greedily read the predicted block at this sequence's noise positions 1..n_block_tokens-1 + for (int32_t i = 1; i < n_block_tokens; ++i) { + common_sampler_sample(smpl, ctx_dft, beg + i, true); + + const auto * cur_p = common_sampler_get_candidates(smpl, true); + + for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { + LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", + seq_id, k, i - 1, cur_p->data[k].id, cur_p->data[k].p, + common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); + } + + const llama_token id = cur_p->data[0].id; + + if (cur_p->data[0].p < params.p_min) { + break; + } + + common_sampler_accept(smpl, id, true); + + result.push_back(id); + } + + if (result.size() < (size_t) params.n_min) { + result.clear(); + } + } + } + + void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override { + // noop + } + + bool need_embd() const override { + return false; + } +}; + struct common_speculative_impl_draft_mtp : public common_speculative_impl { common_params_speculative_draft params; // reuses the draft-model params slot (ctx_tgt/ctx_dft) @@ -942,9 +1247,9 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { "MTP input row width must match the target h_nextn width"); n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft))); - LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__); - LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling); - LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__, + SPC_TRC("%s", "adding speculative implementation 'draft-mtp'\n"); + SPC_TRC("- n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling); + SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", this->params.n_gpu_layers, ggml_type_name(this->params.cache_type_k), ggml_type_name(this->params.cache_type_v), @@ -975,7 +1280,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { llama_sampler_chain_add(chain, llama_sampler_init_top_k(10)); if (!llama_set_sampler(ctx_dft, seq_id, chain)) { - LOG_WRN("%s: backend offload failed for seq_id=%d; using CPU sampler\n", __func__, (int) seq_id); + SPC_WRN("backend offload failed for seq_id=%d; using CPU sampler\n", (int) seq_id); llama_sampler_free(chain); chain = nullptr; } @@ -1038,11 +1343,11 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id); if (pos_max < N - 1 && !is_mem_shared) { - LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - " + SPC_WRN("ctx_dft pos_max=%d < N-1=%d - " "process() hook may not have run on every prefill ubatch " "(need_embd / logits=1 on every prompt position?). " "Drafts may degrade.\n", - __func__, (int) pos_max, N - 1); + (int) pos_max, N - 1); } } @@ -1128,8 +1433,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { const int32_t rc = llama_decode(ctx_dft, batch); if (rc != 0) { - LOG_ERR("%s: llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n", - __func__, head, (int) rc, (int) batch_in.pos[0]); + SPC_ERR("llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n", + head, (int) rc, (int) batch_in.pos[0]); ok = false; break; } @@ -1217,7 +1522,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { int ret = llama_decode(ctx_dft, batch); if (ret != 0) { - LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret); + SPC_ERR("llama_decode[%d] returned %d\n", i, ret); break; } @@ -1239,7 +1544,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { const auto * cur_p = common_sampler_get_candidates(smpl, true); for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { - LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", + SPC_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); } @@ -1353,8 +1658,8 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl { , params(params.ngram_simple) , config(config) { - LOG_INF("%s: adding speculative implementation 'ngram-simple'\n", __func__); - LOG_INF("%s: - size_n=%d, size_m=%d, min_hits=%d\n", __func__, + SPC_TRC("%s", "adding speculative implementation 'ngram-simple'\n"); + SPC_TRC("- size_n=%d, size_m=%d, min_hits=%d\n", this->params.size_n, this->params.size_m, this->params.min_hits); } @@ -1403,8 +1708,8 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl { this->config.push_back(config); } - LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(this->type).c_str()); - LOG_INF("%s: - size_key=%d, size_value=%d, key_only=%d, min_hits=%d\n", __func__, + SPC_TRC("adding speculative implementation '%s'\n", common_speculative_type_to_str(this->type).c_str()); + SPC_TRC("- size_key=%d, size_value=%d, key_only=%d, min_hits=%d\n", config.size_key, config.size_value, config.key_only, config.min_hits); } @@ -1478,15 +1783,15 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl { , verbose(std::getenv("LLAMA_TRACE") != nullptr) { static_assert(sizeof(llama_token) == sizeof(common_ngram_mod::entry_t)); - LOG_INF("%s: adding speculative implementation 'ngram-mod'\n", __func__); - LOG_INF("%s: - n_match=%d, n_max=%d, n_min=%d\n", __func__, + SPC_TRC("%s", "adding speculative implementation 'ngram-mod'\n"); + SPC_TRC("- n_match=%d, n_max=%d, n_min=%d\n", this->params.n_match, this->params.n_max, this->params.n_min); - LOG_INF("%s: - mod size=%zu (%.3f MB)\n", __func__, + SPC_TRC("- mod size=%zu (%.3f MB)\n", mod.size(), (float)(mod.size_bytes())/1024/1024); if (this->params.n_match < 16) { - LOG_WRN("%s: ngram_mod n_match=%d is too small - poor quality is possible, " - "see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, this->params.n_match); + SPC_WRN("ngram_mod n_match=%d is too small - poor quality is possible, " + "see: https://github.com/ggml-org/llama.cpp/pull/19164\n", this->params.n_match); } sinfos.resize(n_seq); @@ -1510,11 +1815,11 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl { sinfo.i_last = prompt.size() - n; const double f = (double)mod.get_used() / (double)mod.size(); - LOG_INF("%s: ngram_mod occupancy = %zu/%zu (%.2f)\n", __func__, mod.get_used(), mod.size(), f); + SPC_TRC("ngram_mod occupancy = %zu/%zu (%.2f)\n", mod.get_used(), mod.size(), f); constexpr double f_thold = 0.25; if (f > f_thold) { - LOG_WRN("%s: ngram_mod occupancy %.2f exceeds threshold (%.2f) - resetting\n", __func__, f, f_thold); + SPC_WRN("ngram_mod occupancy %.2f exceeds threshold (%.2f) - resetting\n", f, f_thold); mod.reset(); } @@ -1608,7 +1913,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl { sinfo.n_low++; if (sinfo.n_low >= 5) { if (verbose) { - LOG_WRN("%s: low acceptance streak (%d) - resetting ngram_mod\n", __func__, sinfo.n_low); + SPC_TRC("low acceptance streak (%d) - resetting ngram_mod\n", sinfo.n_low); } mod.reset(); @@ -1658,8 +1963,8 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl { , save_dynamic(save_dynamic) , save_static(save_static) { - LOG_INF("%s: adding speculative implementation 'ngram-cache'\n", __func__); - LOG_INF("%s: - n_draft=%d, cache_static=%s, cache_dynamic=%s\n", __func__, + SPC_TRC("%s", "adding speculative implementation 'ngram-cache'\n"); + SPC_TRC("- n_draft=%d, cache_static=%s, cache_dynamic=%s\n", n_draft, path_static.empty() ? "none" : path_static.c_str(), path_dynamic.empty() ? "none" : path_dynamic.c_str()); @@ -1674,7 +1979,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl { sinfo.ngram_cache_static = ngram_cache_static; } } catch (...) { - LOG_ERR("failed to open static lookup cache: %s", path_static.c_str()); + SPC_ERR("failed to open static lookup cache: %s", path_static.c_str()); GGML_ABORT("Couldn't read static lookup cache"); } } @@ -1687,7 +1992,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl { sinfo.ngram_cache_dynamic = ngram_cache_dynamic; } } catch (...) { - LOG_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str()); + SPC_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str()); GGML_ABORT("Couldn't read dynamic lookup cache"); } } @@ -1836,6 +2141,7 @@ std::string common_speculative_type_to_str(common_speculative_type type) { case COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE: return "draft-simple"; case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3: return "draft-eagle3"; case COMMON_SPECULATIVE_TYPE_DRAFT_MTP: return "draft-mtp"; + case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH: return "draft-dflash"; case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: return "ngram-simple"; case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: return "ngram-map-k"; case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram-map-k4v"; @@ -1888,6 +2194,7 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) { case COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE: case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3: case COMMON_SPECULATIVE_TYPE_DRAFT_MTP: + case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH: n_max = std::max(n_max, std::max(0, spec->draft.n_max)); break; case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: @@ -1925,6 +2232,7 @@ common_speculative * common_speculative_init(common_params_speculative & params, bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE)); bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr; bool has_draft_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr; + bool has_draft_dflash = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)) && params.draft.ctx_dft != nullptr; @@ -1935,7 +2243,7 @@ common_speculative * common_speculative_init(common_params_speculative & params, bool has_ngram_mod = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD)); // when adding a new type - update here the logic above - static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 9); + static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 10); // this list here defines the priority of the speculators // the one with highest priority are listed first @@ -1965,6 +2273,9 @@ common_speculative * common_speculative_init(common_params_speculative & params, if (has_draft_mtp) { configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params)); } + if (has_draft_dflash) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params)); + } } std::vector> impls = {}; @@ -1985,6 +2296,10 @@ common_speculative * common_speculative_init(common_params_speculative & params, impls.push_back(std::make_unique(config.params, n_seq)); break; } + case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH: { + impls.push_back(std::make_unique(config.params, n_seq)); + break; + } case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: { common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple); @@ -2034,7 +2349,7 @@ common_speculative * common_speculative_init(common_params_speculative & params, } if (impls.empty()) { - LOG_WRN("%s: no implementations specified for speculative decoding\n", __func__); + SPC_TRC("%s", "no implementations specified for speculative decoding\n"); return nullptr; } @@ -2161,13 +2476,13 @@ void common_speculative_draft(common_speculative * spec) { if (dp.n_max > 0) { if (!result.empty() && (int) result.size() > dp.n_max) { - LOG_DBG("%s: truncating draft to %d tokens\n", __func__, dp.n_max); + SPC_DBG("truncating draft to %d tokens\n", dp.n_max); result.resize(dp.n_max); } } if (!result.empty()) { - LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__, + SPC_DBG("called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", common_speculative_type_to_str(impl.get()->type).c_str(), dp.prompt->size(), impl.get()->n_call_draft, result.size()); @@ -2291,7 +2606,7 @@ void common_speculative_print_stats(const common_speculative * spec) { str_stats = ", #mean acc len = " + oss.str() + ", #acc rate/pos = (" + tmp.str() + ")"; } - LOG_INF("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s%s\n", + SPC_TRC("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s%s\n", common_speculative_type_to_str(impl->type).c_str(), impl->n_call_begin, impl->n_call_draft, impl->n_call_accept, impl->n_gen_drafts, diff --git a/conversion/__init__.py b/conversion/__init__.py index 5aad203e5..02ea63852 100644 --- a/conversion/__init__.py +++ b/conversion/__init__.py @@ -50,6 +50,8 @@ TEXT_MODEL_MAP: dict[str, str] = { "DeepseekV2ForCausalLM": "deepseek", "DeepseekV3ForCausalLM": "deepseek", "DeepseekV32ForCausalLM": "deepseek", + "DFlashDraftModel": "qwen", + "DeepseekV4ForCausalLM": "deepseek", "DistilBertForMaskedLM": "bert", "DistilBertForSequenceClassification": "bert", "DistilBertModel": "bert", diff --git a/conversion/base.py b/conversion/base.py index 08fd3747c..0421aa4bc 100644 --- a/conversion/base.py +++ b/conversion/base.py @@ -1273,7 +1273,7 @@ class TextModel(ModelBase): if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None: self.gguf_writer.add_layer_norm_eps(f_norm_eps) logger.info(f"gguf: layer norm epsilon = {f_norm_eps}") - if (n_experts := self.find_hparam(["num_local_experts", "num_experts"], optional=True)) is not None: + if (n_experts := self.find_hparam(["num_local_experts", "num_experts", "n_routed_experts"], optional=True)) is not None: self.gguf_writer.add_expert_count(n_experts) logger.info(f"gguf: expert count = {n_experts}") if (n_experts_used := self.find_hparam(["num_experts_per_tok", "num_experts_per_token", "top_k_experts"], optional=True)) is not None: @@ -1291,6 +1291,8 @@ class TextModel(ModelBase): self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID) elif score_func == "softmax": self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX) + elif score_func == "sqrtsoftplus": + self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SQRTSOFTPLUS) else: raise ValueError(f"Unsupported expert score gating function value: {score_func}") logger.info(f"gguf: expert score gating function = {score_func}") @@ -2600,6 +2602,17 @@ class LazyTorchTensor(gguf.LazyBase): return cls._wrap_fn(func)(*args, **kwargs) +if hasattr(torch, "float8_e8m0fnu"): + _torch_float8_e8m0 = torch.float8_e8m0fnu + LazyTorchTensor._dtype_map[_torch_float8_e8m0] = np.uint8 + LazyTorchTensor._dtype_byteswap_map[_torch_float8_e8m0] = np.uint8 + LazyTorchTensor._dtype_str_map["F8_E8M0"] = _torch_float8_e8m0 +else: + # Older torch builds do not expose F8_E8M0. Keep the raw bytes so callers + # that know the format can decode them explicitly. + LazyTorchTensor._dtype_str_map["F8_E8M0"] = torch.uint8 + + def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str: # TODO @ngxson : this won't work correctly if the model has both audio & vision encoders # maybe we should fallback to text model's arch in that case, since not many models have both diff --git a/conversion/deepseek.py b/conversion/deepseek.py index 4c93fb66d..ea6ae23d5 100644 --- a/conversion/deepseek.py +++ b/conversion/deepseek.py @@ -1,15 +1,18 @@ from __future__ import annotations +import json import re +from pathlib import Path from typing import Any, Callable, Iterable, TYPE_CHECKING +import numpy as np import torch if TYPE_CHECKING: from torch import Tensor -from .base import MmprojModel, ModelBase, TextModel, gguf, logger +from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logger from .qwen import QwenModel @@ -467,3 +470,307 @@ class DeepseekV32Model(DeepseekV2Model): self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"]) self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"]) self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"]) + + +@ModelBase.register("DeepseekV4ForCausalLM") +class DeepseekV4Model(TextModel): + model_arch = gguf.MODEL_ARCH.DEEPSEEK4 + _skipped_mtp_tensors = 0 + + def __init__(self, *args, **kwargs): + type(self)._skipped_mtp_tensors = 0 + super().__init__(*args, **kwargs) + + with open(self.dir_model / "config.json", "r", encoding="utf-8") as f: + raw_hparams = json.load(f) + for key, value in raw_hparams.items(): + self.hparams.setdefault(key, value) + + self.block_count = self.hparams["num_hidden_layers"] + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + + self._dsv4_fp8_dequantized: set[str] = set() + self._dsv4_bf16_tensors: set[str] = set() + self._dsv4_f32_tensors: set[str] = set() + self._dsv4_mxfp4_generated = False + self._collect_source_dtypes() + + if type(self)._skipped_mtp_tensors: + logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors) + + # add a default chat template; if the model has a built-in template, it will be overridden later + template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja" + if template_path.is_file(): + with open(template_path, "r", encoding="utf-8") as f: + self.gguf_writer.add_chat_template(f.read()) + + @classmethod + def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None: + name, _ = item + if name.startswith("mtp."): + cls._skipped_mtp_tensors += 1 + return None + return super().filter_tensors(item) + + @staticmethod + def _float8_dtypes() -> tuple[torch.dtype, ...]: + return tuple( + dtype for dtype in ( + getattr(torch, "float8_e4m3fn", None), + getattr(torch, "float8_e5m2", None), + ) if dtype is not None + ) + + @staticmethod + def _e8m0_to_float(scale: Tensor) -> Tensor: + torch_float8_e8m0 = getattr(torch, "float8_e8m0fnu", None) + if torch_float8_e8m0 is not None and scale.dtype == torch_float8_e8m0: + return scale.float() + + bits = scale.view(torch.uint8).float() + return torch.exp2(bits - 127.0) + + def _collect_source_dtypes(self) -> None: + for name, gen in self.model_tensors.items(): + dtype = gen().dtype + if dtype == torch.bfloat16: + self._dsv4_bf16_tensors.add(name) + elif dtype == torch.float32: + self._dsv4_f32_tensors.add(name) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) + self.gguf_writer.add_sliding_window(hparams["sliding_window"]) + + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"]) + self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"]) + self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"]) + self.gguf_writer.add_swiglu_clamp_exp([hparams["swiglu_limit"]] * self.block_count) + self.gguf_writer.add_swiglu_clamp_shexp([hparams["swiglu_limit"]] * self.block_count) + + self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"]) + self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"]) + self.gguf_writer.add_indexer_top_k(hparams["index_topk"]) + + self.gguf_writer.add_attention_output_group_count(hparams["o_groups"]) + self.gguf_writer.add_attention_output_lora_rank(hparams["o_lora_rank"]) + self.gguf_writer.add_attention_compress_ratios(hparams["compress_ratios"]) + self.gguf_writer.add_attention_compress_rope_freq_base(hparams["compress_rope_theta"]) + self.gguf_writer.add_hyper_connection_count(hparams["hc_mult"]) + self.gguf_writer.add_hyper_connection_sinkhorn_iterations(hparams["hc_sinkhorn_iters"]) + self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"]) + self.gguf_writer.add_hash_layer_count(hparams["num_hash_layers"]) + + def dequant_model(self): + fp8_dtypes = self._float8_dtypes() + tensors_to_remove: list[str] = [] + + def dequant_fp8_weight(weight: Tensor, scale: Tensor) -> Tensor: + out_features, in_features = weight.shape + scale_f = self._e8m0_to_float(scale) + scale_f = scale_f.repeat_interleave(128, 0)[:out_features] + scale_f = scale_f.repeat_interleave(128, 1)[:, :in_features] + return weight.float() * scale_f + + for name in list(self.model_tensors.keys()): + if not name.endswith(".scale"): + continue + weight_name = name.removesuffix(".scale") + ".weight" + if weight_name not in self.model_tensors: + continue + + weight = self.model_tensors[weight_name] + scale = self.model_tensors[name] + if weight().dtype not in fp8_dtypes: + continue + + self.model_tensors[weight_name] = lambda w=weight, s=scale: dequant_fp8_weight(w(), s()) + self._dsv4_fp8_dequantized.add(weight_name) + tensors_to_remove.append(name) + + for name in tensors_to_remove: + del self.model_tensors[name] + + @staticmethod + def _pack_mxfp4_blocks(weight: Tensor, scale: Tensor) -> np.ndarray: + packed = weight.contiguous().view(torch.uint8) + scale_u8 = scale.contiguous().view(torch.uint8) + + out_features, packed_cols = packed.shape + logical_cols = packed_cols * 2 + if logical_cols % 32 != 0: + raise ValueError(f"MXFP4 source row has {logical_cols} values, expected a multiple of 32") + + n_blocks = logical_cols // 32 + if tuple(scale_u8.shape) != (out_features, n_blocks): + raise ValueError(f"MXFP4 scale shape {tuple(scale_u8.shape)} does not match {(out_features, n_blocks)}") + + src = packed.reshape(out_features, n_blocks, 16) + low = src & 0x0F + high = (src >> 4) & 0x0F + + # The safetensors bytes store adjacent values as low/high nibbles. + # ggml MXFP4 blocks store values 0..15 in low nibbles and 16..31 in high nibbles. + vals = torch.stack((low, high), dim=-1).reshape(out_features, n_blocks, 32) + qs = vals[:, :, :16] | (vals[:, :, 16:] << 4) + raw = torch.cat((scale_u8.unsqueeze(-1), qs.to(torch.uint8)), dim=-1) + return raw.reshape(out_features, n_blocks * 17).cpu().numpy() + + def _write_mxfp4_expert_tensor(self, bid: int, proj: str, tensor_key: gguf.MODEL_TENSOR) -> list[str]: + n_experts = self.hparams["n_routed_experts"] + data: np.ndarray | None = None + consumed: list[str] = [] + + for eid in range(n_experts): + weight_name = f"layers.{bid}.ffn.experts.{eid}.{proj}.weight" + scale_name = f"layers.{bid}.ffn.experts.{eid}.{proj}.scale" + if weight_name not in self.model_tensors or scale_name not in self.model_tensors: + raise KeyError(f"Missing routed expert tensors for {weight_name}") + + weight = LazyTorchTensor.to_eager(self.model_tensors[weight_name]()) + scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]()) + packed = self._pack_mxfp4_blocks(weight, scale) + if data is None: + data = np.empty((n_experts, *packed.shape), dtype=packed.dtype) + data[eid] = packed + consumed.extend((weight_name, scale_name)) + + assert data is not None + new_name = self.format_tensor_name(tensor_key, bid) + shape = gguf.quant_shape_from_byte_shape(data.shape, gguf.GGMLQuantizationType.MXFP4) + logger.info(f"{new_name}: repacked routed experts to MXFP4, shape = {{{', '.join(str(n) for n in reversed(shape))}}}") + self.gguf_writer.add_tensor(new_name, data, raw_dtype=gguf.GGMLQuantizationType.MXFP4) + + return consumed + + def _write_hash_routing_tensors(self) -> list[str]: + consumed: list[str] = [] + + for bid in range(self.hparams["num_hash_layers"]): + name = f"layers.{bid}.ffn.gate.tid2eid" + if name not in self.model_tensors: + raise KeyError(f"Missing hash routing tensor {name}") + + data_torch = LazyTorchTensor.to_eager(self.model_tensors[name]()) + data = data_torch.to(torch.int32).cpu().numpy() + new_name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_TID2EID, bid, ".weight") + logger.info(f"{new_name}: converted hash routing table to I32, shape = {{{', '.join(str(n) for n in reversed(data.shape))}}}") + self.gguf_writer.add_tensor(new_name, data) + consumed.append(name) + + return consumed + + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + if self._dsv4_mxfp4_generated: + return () + + consumed: list[str] = self._write_hash_routing_tensors() + for bid in range(self.block_count): + consumed.extend(self._write_mxfp4_expert_tensor(bid, "w1", gguf.MODEL_TENSOR.FFN_GATE_EXP)) + consumed.extend(self._write_mxfp4_expert_tensor(bid, "w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP)) + consumed.extend(self._write_mxfp4_expert_tensor(bid, "w3", gguf.MODEL_TENSOR.FFN_UP_EXP)) + + for name in consumed: + del self.model_tensors[name] + + self._dsv4_mxfp4_generated = True + return () + + def _format_dsv4_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None, suffix: str = ".weight") -> str: + return self.format_tensor_name(key, bid, suffix) + + def _map_dsv4_tensor_name(self, name: str, bid: int | None) -> tuple[gguf.MODEL_TENSOR, str]: + root_map: dict[str, tuple[gguf.MODEL_TENSOR, str]] = { + "embed.weight": (gguf.MODEL_TENSOR.TOKEN_EMBD, ".weight"), + "norm.weight": (gguf.MODEL_TENSOR.OUTPUT_NORM, ".weight"), + "head.weight": (gguf.MODEL_TENSOR.OUTPUT, ".weight"), + "hc_head_fn": (gguf.MODEL_TENSOR.HC_HEAD_FN, ".weight"), + "hc_head_base": (gguf.MODEL_TENSOR.HC_HEAD_BASE, ".weight"), + "hc_head_scale": (gguf.MODEL_TENSOR.HC_HEAD_SCALE, ".weight"), + } + if name in root_map: + return root_map[name] + + match = re.match(r"layers\.(\d+)\.(.+)$", name) + if match is None: + raise ValueError(f"Unsupported DeepSeek-V4 tensor {name!r}") + + layer = int(match.group(1)) + if bid != layer: + raise ValueError(f"Tensor {name!r} parsed bid {bid} but layer name has {layer}") + + layer_map: dict[str, tuple[gguf.MODEL_TENSOR, str]] = { + "hc_attn_fn": (gguf.MODEL_TENSOR.HC_ATTN_FN, ".weight"), + "hc_attn_base": (gguf.MODEL_TENSOR.HC_ATTN_BASE, ".weight"), + "hc_attn_scale": (gguf.MODEL_TENSOR.HC_ATTN_SCALE, ".weight"), + "hc_ffn_fn": (gguf.MODEL_TENSOR.HC_FFN_FN, ".weight"), + "hc_ffn_base": (gguf.MODEL_TENSOR.HC_FFN_BASE, ".weight"), + "hc_ffn_scale": (gguf.MODEL_TENSOR.HC_FFN_SCALE, ".weight"), + "attn.attn_sink": (gguf.MODEL_TENSOR.ATTN_SINKS, ".weight"), + "attn.wq_a.weight": (gguf.MODEL_TENSOR.ATTN_Q_A, ".weight"), + "attn.wq_b.weight": (gguf.MODEL_TENSOR.ATTN_Q_B, ".weight"), + "attn.q_norm.weight": (gguf.MODEL_TENSOR.ATTN_Q_A_NORM, ".weight"), + "attn.wkv.weight": (gguf.MODEL_TENSOR.ATTN_KV, ".weight"), + "attn.kv_norm.weight": (gguf.MODEL_TENSOR.ATTN_KV_NORM, ".weight"), + "attn.wo_a.weight": (gguf.MODEL_TENSOR.ATTN_OUT_A, ".weight"), + "attn.wo_b.weight": (gguf.MODEL_TENSOR.ATTN_OUT_B, ".weight"), + "attn.compressor.ape": (gguf.MODEL_TENSOR.ATTN_COMPRESSOR_APE, ".weight"), + "attn.compressor.wkv.weight": (gguf.MODEL_TENSOR.ATTN_COMPRESSOR_WKV, ".weight"), + "attn.compressor.wgate.weight": (gguf.MODEL_TENSOR.ATTN_COMPRESSOR_WGATE, ".weight"), + "attn.compressor.norm.weight": (gguf.MODEL_TENSOR.ATTN_COMPRESSOR_NORM, ".weight"), + "attn.indexer.wq_b.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_Q_B, ".weight"), + "attn.indexer.weights_proj.weight": (gguf.MODEL_TENSOR.INDEXER_PROJ, ".weight"), + "attn.indexer.compressor.ape": (gguf.MODEL_TENSOR.INDEXER_COMPRESSOR_APE, ".weight"), + "attn.indexer.compressor.wkv.weight": (gguf.MODEL_TENSOR.INDEXER_COMPRESSOR_WKV, ".weight"), + "attn.indexer.compressor.wgate.weight": (gguf.MODEL_TENSOR.INDEXER_COMPRESSOR_WGATE, ".weight"), + "attn.indexer.compressor.norm.weight": (gguf.MODEL_TENSOR.INDEXER_COMPRESSOR_NORM, ".weight"), + "attn_norm.weight": (gguf.MODEL_TENSOR.ATTN_NORM, ".weight"), + "ffn_norm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"), + "ffn.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"), + "ffn.gate.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"), + "ffn.gate.tid2eid": (gguf.MODEL_TENSOR.FFN_GATE_TID2EID, ".weight"), + "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"), + "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"), + "ffn.shared_experts.w3.weight": (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"), + } + + tensor_name = match.group(2) + if tensor_name in layer_map: + return layer_map[tensor_name] + + if re.match(r"ffn\.experts\.\d+\.w[123]\.(weight|scale)$", tensor_name): + return gguf.MODEL_TENSOR.FFN_GATE_EXP, ".weight" + + raise ValueError(f"Unsupported DeepSeek-V4 tensor {name!r}") + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name): + return [] + + tensor_key, suffix = self._map_dsv4_tensor_name(name, bid) + if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID: + return [] + + return [(self._format_dsv4_tensor_name(tensor_key, bid, suffix), data_torch)] + + def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool: + del new_name, bid # unused + + if name in self._dsv4_fp8_dequantized and n_dims >= 2: + return gguf.GGMLQuantizationType.Q8_0 + if name in self._dsv4_f32_tensors: + return gguf.GGMLQuantizationType.F32 + if name in self._dsv4_bf16_tensors and n_dims >= 2: + return gguf.GGMLQuantizationType.BF16 + + return False + + def prepare_tensors(self): + super().prepare_tensors() + self._is_mxfp4 = True + self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE diff --git a/conversion/llama.py b/conversion/llama.py index b43cc994a..315a619c9 100644 --- a/conversion/llama.py +++ b/conversion/llama.py @@ -73,7 +73,7 @@ class LlamaModel(TextModel): target_num_layers = target_config["num_hidden_layers"] target_layers = [2, target_num_layers // 2, target_num_layers - 3] logger.info(f"EAGLE-3: target_layers = {target_layers} (target model has {target_num_layers} layers)") - self.gguf_writer.add_array(f"{self.gguf_writer.arch}.target_layers", target_layers) + self.gguf_writer.add_target_layers(target_layers) # target_hidden_size: prefer eagle3 config, fallback to target config if eagle3_raw_config.get("target_hidden_size") is not None: @@ -83,12 +83,12 @@ class LlamaModel(TextModel): target_hidden_size = target_config["hidden_size"] src = "target model config" logger.info(f"EAGLE-3: target_hidden_size = {target_hidden_size} (from {src})") - self.gguf_writer.add_uint32(f"{self.gguf_writer.arch}.target_hidden_size", target_hidden_size) + self.gguf_writer.add_target_hidden_size(target_hidden_size) # norm_before_residual (RedHat-style eagle3 specific) norm_before_residual = eagle3_raw_config.get("norm_before_residual", False) logger.info(f"EAGLE-3: norm_before_residual = {norm_before_residual}") - self.gguf_writer.add_bool(f"{self.gguf_writer.arch}.norm_before_residual", norm_before_residual) + self.gguf_writer.add_norm_before_residual(norm_before_residual) def set_vocab(self): # eagle3: use tokenizer from target model if provided diff --git a/conversion/qwen.py b/conversion/qwen.py index 6b85eb9aa..0356bd2da 100644 --- a/conversion/qwen.py +++ b/conversion/qwen.py @@ -625,3 +625,51 @@ class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReor @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM") class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase): model_arch = gguf.MODEL_ARCH.QWEN35MOE + + +@ModelBase.register("DFlashDraftModel") +class DFlashModel(Qwen3Model): + model_arch = gguf.MODEL_ARCH.DFLASH + + def set_vocab(self): + if self.target_model_dir is None: + raise ValueError( + "DFlash draft model requires --target-model-dir to be specified. " + "Please provide the path to the target model directory containing the tokenizer." + ) + logger.info(f"DFlash: Using tokenizer from target model: {self.target_model_dir}") + original_dir = self.dir_model + self.dir_model = self.target_model_dir + super().set_vocab() + self.dir_model = original_dir + + mask_token_id = self.hparams.get("dflash_config", {}).get("mask_token_id") + if mask_token_id is not None: + self.gguf_writer.add_mask_token_id(mask_token_id) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + + block_size = self.hparams.get("block_size", 16) + self.gguf_writer.add_block_size(block_size) + dflash_config = self.hparams.get("dflash_config", {}) + + target_layer_ids = dflash_config.get("target_layer_ids", []) + if target_layer_ids: + extract_layer_ids = [i + 1 for i in target_layer_ids] + self.gguf_writer.add_target_layers(extract_layer_ids) + + use_sliding_window = self.hparams.get("use_sliding_window", False) + sliding_window = self.hparams.get("sliding_window") + layer_types = self.hparams.get("layer_types") + if use_sliding_window and sliding_window and layer_types: + is_swa = [lt == "sliding_attention" for lt in layer_types] + self.gguf_writer.add_sliding_window(sliding_window) + self.gguf_writer.add_sliding_window_pattern(is_swa) + + @classmethod + def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None: + name, gen = item + if not name.startswith("model."): + name = "model." + name + return super().filter_tensors((name, gen)) diff --git a/embd_res/kcpp_sdui.embd b/embd_res/kcpp_sdui.embd index 50807c23b..745fec1d3 100644 --- a/embd_res/kcpp_sdui.embd +++ b/embd_res/kcpp_sdui.embd @@ -5,18 +5,18 @@ Stable UI for KoboldCpp - - diff --git a/embd_res/klite.embd b/embd_res/klite.embd index 81d70a8fc..3b3f86b40 100644 --- a/embd_res/klite.embd +++ b/embd_res/klite.embd @@ -12,7 +12,7 @@ Current version indicated by LITEVER below. -->
@@ -153,13 +151,13 @@ MCP Servers - {getEnabledMcpServers().length} server{getEnabledMcpServers().length !== 1 ? 's' : ''} + {visibleMcpServers.length} server{visibleMcpServers.length !== 1 ? 's' : ''}
- {#each getEnabledMcpServers() as server (server.id)} + {#each visibleMcpServers as server (server.id)} {@const healthState = mcpStore.getHealthCheckState(server.id)} {@const hasError = healthState.status === HealthCheckStatus.ERROR} {@const displayName = mcpStore.getServerLabel(server)} @@ -202,7 +200,7 @@ {/each} - {#if getEnabledMcpServers().length === 0} + {#if visibleMcpServers.length === 0}
No MCP servers configured
diff --git a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserPending.svelte b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserPending.svelte index 5c2913202..4be582b39 100644 --- a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserPending.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserPending.svelte @@ -56,7 +56,7 @@
diff --git a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte index e466c84ee..4337bb6a1 100644 --- a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageActions/ChatMessageActionCard/ChatMessageActionCardPermissionRequest.svelte @@ -1,8 +1,9 @@ @@ -60,29 +69,27 @@ Add New Server -
- (newServerUrl = v)} - onHeadersChange={(v) => (newServerHeaders = v)} - urlError={newServerUrl ? newServerUrlError : null} - id="new-server" - /> -
+
+
+ (newServerUrl = v)} + onHeadersChange={(v) => (newServerHeaders = v)} + urlError={newServerUrl ? newServerUrlError : null} + id="new-server" + /> +
- - + + - - + + +
diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerRecommendations.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerRecommendations.svelte new file mode 100644 index 000000000..9b4489b82 --- /dev/null +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerRecommendations.svelte @@ -0,0 +1,180 @@ + + + + + + Do more with MCP + + Power-up your experience by adding tools, resources and more capabilities provided by MCP + servers. + + + +
+

Quickly get started with

+ + {#each RECOMMENDED_MCP_SERVERS as server (server.id)} + (selected[server.id] = enabled)} + /> + {/each} + + {#if addedServers.length > 0} + {#each addedServers as server (server.id)} + + {/each} + {/if} + + {#if showAddForm} + + (newServerUrl = v)} + onHeadersChange={(v) => (newServerHeaders = v)} + urlError={newServerUrl ? newServerUrlError : null} + id="recommendation-new-server" + /> + +
+ + + +
+ + {:else} + + + + {/if} +
+ + + + + + + + diff --git a/tools/ui/src/lib/components/app/dialogs/index.ts b/tools/ui/src/lib/components/app/dialogs/index.ts index 29136308c..73f22c565 100644 --- a/tools/ui/src/lib/components/app/dialogs/index.ts +++ b/tools/ui/src/lib/components/app/dialogs/index.ts @@ -18,6 +18,15 @@ */ export { default as DialogMcpServerAddNew } from './DialogMcpServerAddNew.svelte'; +/** + * **DialogMcpServerRecommendations** - Suggested MCP servers opt-in dialog + * + * Prompts the user to enable pre-defined recommended MCP servers on first launch. + * Shows one switch per suggested server and persists the choice as a per-chat + * override so the selected servers become available in conversations. + */ +export { default as DialogMcpServerRecommendations } from './DialogMcpServerRecommendations.svelte'; + /** * **DialogExportSettings** - Settings export dialog with sensitive data warning * diff --git a/tools/ui/src/lib/components/app/forms/KeyValuePairs.svelte b/tools/ui/src/lib/components/app/forms/KeyValuePairs.svelte index e0bd8d98e..fd6e59a5b 100644 --- a/tools/ui/src/lib/components/app/forms/KeyValuePairs.svelte +++ b/tools/ui/src/lib/components/app/forms/KeyValuePairs.svelte @@ -1,4 +1,5 @@
@@ -103,6 +123,7 @@ {#each pairs as pair, index (index)}
-
+
{#if showSkeleton} {:else if protocolVersion} diff --git a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardCompact.svelte b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardCompact.svelte new file mode 100644 index 000000000..6cb3e18b6 --- /dev/null +++ b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardCompact.svelte @@ -0,0 +1,156 @@ + + + +
+
+ {#if showSkeleton} + + + + + {:else} + + {/if} +
+ + +
+ + {#if isError && errorMessage} +

{errorMessage}

+ {/if} + + {#if showSkeleton} +
+ +
+ +
+ + + + +
+ {:else} + {#if description} + {#if description.lines === 2} +

+ {description.text} +

+ {:else} +

+ {description.text} +

+ {/if} + {/if} + + {#if tools.length > 0} +
+ {#each visibleTools as tool (tool.name)} + + + + {tool.name} + + + + +

+ {tool.description ?? 'No description'} +

+ + + {/each} + + {#if hiddenToolCount > 0} + + + + + {hiddenToolCount} more tools + + + + +

+ {hiddenTools.map((tool) => tool.name).join(', ')} +

+ + + {/if} +
+ {/if} + {/if} + diff --git a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardEditForm.svelte b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardEditForm.svelte index 6727a9000..8ed4ee8b8 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardEditForm.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardEditForm.svelte @@ -1,6 +1,7 @@ -
-

Configure Server

+
+
+

Configure Server

- (editUrl = v)} - onHeadersChange={(v) => (editHeaders = v)} - onUseProxyChange={(v) => (editUseProxy = v)} - urlError={editUrl ? urlError : null} - id={serverId} - /> + (editUrl = v)} + onHeadersChange={(v) => (editHeaders = v)} + onUseProxyChange={(v) => (editUseProxy = v)} + urlError={editUrl ? urlError : null} + id={serverId} + /> -
- +
+ - + +
-
+
diff --git a/tools/ui/src/lib/components/app/mcp/McpServerForm.svelte b/tools/ui/src/lib/components/app/mcp/McpServerForm.svelte index 79738e30d..7f05d5fef 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerForm.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerForm.svelte @@ -38,14 +38,87 @@ let headerPairs = $derived(parseHeadersToArray(headers)); + const AUTHORIZATION_HEADER = 'Authorization'; + const BEARER_PREFIX = 'Bearer '; + + // Heuristic: this dedicated UI only owns Authorization headers that already + // carry a Bearer scheme. Anything else (e.g. Basic, raw tokens) stays in the + // KV section so the user can still edit those values verbatim. + const matchesAuthorizationKey = (key: string): boolean => + key.trim().toLowerCase() === AUTHORIZATION_HEADER.toLowerCase(); + + const isBearerScheme = (value: string): boolean => + value.trim().toLowerCase().startsWith(BEARER_PREFIX.toLowerCase()); + + const ownedByBearerUi = (p: KeyValuePair): boolean => + matchesAuthorizationKey(p.key) && isBearerScheme(p.value); + + let hasAuthorization = $derived(headerPairs.some(ownedByBearerUi)); + + let wantsAuthorization = $state(false); + + let showAuthorization = $derived(hasAuthorization || wantsAuthorization); + + let urlInput: HTMLInputElement | null = $state(null); + let bearerInput: HTMLInputElement | null = $state(null); + + $effect(() => { + urlInput?.focus(); + }); + + $effect(() => { + if (wantsAuthorization && bearerInput) { + bearerInput.focus(); + } + }); + + let bearerToken = $derived.by(() => { + const auth = headerPairs.find(ownedByBearerUi); + if (!auth) return ''; + return auth.value.trim().slice(BEARER_PREFIX.length).trim(); + }); + + $effect(() => { + if (!headers.trim()) { + wantsAuthorization = false; + } + }); + function updateHeaderPairs(newPairs: KeyValuePair[]) { headerPairs = newPairs; onHeadersChange(serializeHeaders(newPairs)); } + + // The dedicated UI owns the Authorization slot end-to-end when the user + // engages it: any prior Authorization row (Bearer or otherwise) is replaced + // by exactly one { Authorization: "Bearer " } entry. JSON's last-key + // behavior would otherwise pick one arbitrarily, so we strip first. + function updateBearerToken(token: string) { + const filtered = headerPairs.filter((p) => !matchesAuthorizationKey(p.key)); + + const trimmed = token.trim(); + + if (trimmed) { + filtered.push({ key: AUTHORIZATION_HEADER, value: `${BEARER_PREFIX}${trimmed}` }); + } + + updateHeaderPairs(filtered); + } + + function setUseAuthorization(checked: boolean) { + wantsAuthorization = checked; + + if (!checked) { + // Only drop the entry this UI owns; a non-Bearer Authorization row + // authored in the KV section must survive a toggle off untouched. + const filtered = headerPairs.filter((p) => !ownedByBearerUi(p)); + updateHeaderPairs(filtered); + } + } -
-
+
+
@@ -57,50 +130,52 @@ value={url} oninput={(e) => onUrlChange(e.currentTarget.value)} class={urlError ? 'border-destructive' : ''} + bind:ref={urlInput} /> {#if urlError}

{urlError}

{/if} - - {#if !isWebSocket && onUseProxyChange} - - {/if}
+ + + {#if showAuthorization} +
+ updateBearerToken(e.currentTarget.value)} + class="pl-16" + bind:ref={bearerInput} + /> + + + Bearer + +
+ {/if} + !ownedByBearerUi(p))} + onPairsChange={(pairs) => { + const auth = headerPairs.find(ownedByBearerUi); + updateHeaderPairs(auth ? [...pairs, auth] : pairs); + }} keyPlaceholder="Header name" valuePlaceholder="Value" addButtonLabel="Add" @@ -108,4 +183,37 @@ sectionLabel="Custom Headers" sectionLabelOptional /> + + {#if !isWebSocket && onUseProxyChange} + + {/if}
diff --git a/tools/ui/src/lib/components/app/mcp/McpServerIdentity.svelte b/tools/ui/src/lib/components/app/mcp/McpServerIdentity.svelte index feafc5d81..3f128e02c 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerIdentity.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerIdentity.svelte @@ -1,6 +1,7 @@ -
{ + if (!e.currentTarget.contains(e.relatedTarget as Node | null)) { + handleMouseLeave(); + } + }} > -
{#if depth > 0} @@ -109,7 +130,7 @@ @@ -125,15 +146,18 @@ {#if isLoading} - +
@@ -145,50 +169,52 @@
-
- { - e.stopPropagation(); - handleTogglePin(); - } - }, - { - icon: Pencil, - label: 'Edit', - onclick: handleEdit, - shortcut: ['shift', 'cmd', 'e'] - }, - { - icon: Download, - label: 'Export', - onclick: (e: Event) => { - e.stopPropagation(); - conversationsStore.downloadConversation(conversation.id); + {#if renderActionsDropdown} +
+ { + e.stopPropagation(); + handleTogglePin(); + } }, - shortcut: ['shift', 'cmd', 's'] - }, - { - icon: Trash2, - label: 'Delete', - onclick: handleDelete, - variant: 'destructive', - shortcut: ['shift', 'cmd', 'd'], - separator: true - } - ]} - /> -
-
+ { + icon: Pencil, + label: 'Edit', + onclick: handleEdit, + shortcut: ['shift', 'cmd', 'e'] + }, + { + icon: Download, + label: 'Export', + onclick: (e: Event) => { + e.stopPropagation(); + conversationsStore.downloadConversation(conversation.id); + }, + shortcut: ['shift', 'cmd', 's'] + }, + { + icon: Trash2, + label: 'Delete', + onclick: handleDelete, + variant: 'destructive', + shortcut: ['shift', 'cmd', 'd'], + separator: true + } + ]} + /> +
+ {/if} +