Merge branch 'concedo_experimental' into kcpp_sd_perf_flags

Author: LostRuins

.github/workflows/kcpp-build-release-arm64.yaml

@@ -87,6 +87,7 @@ jobs:
               --add-data './taesd_xl.embd:.' \
               --add-data './taesd_f.embd:.' \
               --add-data './taesd_3.embd:.' \
+              --add-data './kokoro_ipa.embd:.' \
               --add-data './rwkv_vocab.embd:.' \
               --add-data './rwkv_world_vocab.embd:.' \
               --version-file './version.txt' \

.github/workflows/kcpp-build-release-macos.yaml

@@ -36,7 +36,7 @@ jobs:
           make LLAMA_METAL=1 LLAMA_PORTABLE=1
           chmod +x './create_ver_file.sh'
           . create_ver_file.sh
-          pyinstaller --noconfirm --onefile --collect-all customtkinter --collect-all psutil --add-data './koboldcpp_default.so:.' --add-data './ggml-metal-merged.metal:.' --add-data './kcpp_adapters:./kcpp_adapters' --add-data './koboldcpp.py:.' --add-data './json_to_gbnf.py:.' --add-data './LICENSE.md:.' --add-data './MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md:.' --add-data './klite.embd:.' --add-data './kcpp_docs.embd:.' --add-data './kcpp_sdui.embd:.' --add-data './taesd.embd:.' --add-data './taesd_xl.embd:.' --add-data './taesd_f.embd:.' --add-data './taesd_3.embd:.' --add-data './rwkv_vocab.embd:.' --add-data './rwkv_world_vocab.embd:.' --version-file './version.txt' --clean --console koboldcpp.py -n "koboldcpp-mac-arm64"
+          pyinstaller --noconfirm --onefile --collect-all customtkinter --collect-all psutil --add-data './koboldcpp_default.so:.' --add-data './ggml-metal-merged.metal:.' --add-data './kcpp_adapters:./kcpp_adapters' --add-data './koboldcpp.py:.' --add-data './json_to_gbnf.py:.' --add-data './LICENSE.md:.' --add-data './MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md:.' --add-data './klite.embd:.' --add-data './kcpp_docs.embd:.' --add-data './kcpp_sdui.embd:.' --add-data './taesd.embd:.' --add-data './taesd_xl.embd:.' --add-data './taesd_f.embd:.' --add-data './taesd_3.embd:.' --add-data './kokoro_ipa.embd:.' --add-data './rwkv_vocab.embd:.' --add-data './rwkv_world_vocab.embd:.' --version-file './version.txt' --clean --console koboldcpp.py -n "koboldcpp-mac-arm64"
 
       - name: Test
         id: test

CMakeLists.txt

@@ -474,7 +474,7 @@ set_target_properties(whisper_adapter PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
 add_library(tts_adapter
             otherarch/tts_adapter.cpp)
-target_include_directories(tts_adapter PUBLIC . ./ggml/include ./ggml/src ./ggml/src/ggml-cpu ./include ./otherarch ./otherarch/tools ./vendor/stb ./vendor ./tools ./common)
+target_include_directories(tts_adapter PUBLIC . ./ggml/include ./ggml/src ./ggml/src/ggml-cpu ./include ./otherarch ./otherarch/tools ./vendor/stb ./vendor ./otherarch/ttscpp/include ./otherarch/ttscpp/src ./tools ./common)
 target_compile_features(tts_adapter PUBLIC cxx_std_17) # don't bump
 target_link_libraries(tts_adapter PRIVATE common2 ggml ${LLAMA_EXTRA_LIBS})
 set_target_properties(tts_adapter PROPERTIES POSITION_INDEPENDENT_CODE ON)

Makefile

@@ -55,8 +55,8 @@ ifdef KCPP_SANITIZE
 	CFLAGS += -fsanitize=undefined -fsanitize-undefined-trap-on-error
 	CXXFLAGS += -fsanitize=undefined -fsanitize-undefined-trap-on-error
 endif
-CFLAGS   += -I. -Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -I./common -I./vendor -I./vendor/stb -I./include -I./include/CL -I./otherarch -I./otherarch/tools -I./otherarch/sdcpp -I./otherarch/sdcpp/thirdparty -I./include/vulkan -O3 -fno-finite-math-only -std=c11 -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE -DGGML_USE_CPU -DGGML_USE_CPU_REPACK
-CXXFLAGS += -I. -Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -I./common -I./vendor -I./vendor/stb -I./include -I./include/CL -I./otherarch -I./otherarch/tools -I./otherarch/sdcpp -I./otherarch/sdcpp/thirdparty -I./include/vulkan -O3 -fno-finite-math-only -std=c++17 -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE -DGGML_USE_CPU -DGGML_USE_CPU_REPACK
+CFLAGS   += -I. -Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -I./common -I./vendor -I./vendor/stb -I./include -I./include/CL -I./otherarch -I./otherarch/tools -I./otherarch/sdcpp -I./otherarch/ttscpp/include -I./otherarch/ttscpp/src -I./otherarch/sdcpp/thirdparty -I./include/vulkan -O3 -fno-finite-math-only -std=c11 -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE -DGGML_USE_CPU -DGGML_USE_CPU_REPACK
+CXXFLAGS += -I. -Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -I./common -I./vendor -I./vendor/stb -I./include -I./include/CL -I./otherarch -I./otherarch/tools -I./otherarch/sdcpp -I./otherarch/ttscpp/include -I./otherarch/ttscpp/src -I./otherarch/sdcpp/thirdparty -I./include/vulkan -O3 -fno-finite-math-only -std=c++17 -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE -DGGML_USE_CPU -DGGML_USE_CPU_REPACK
 ifndef KCPP_DEBUG
 	CFLAGS += -DNDEBUG -s
 	CXXFLAGS += -DNDEBUG -s
@@ -684,7 +684,7 @@ whispercpp_cublas.o: otherarch/whispercpp/whisper_adapter.cpp
 	$(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $(HIPFLAGS) -c $< -o $@
 
 #tts objects
-tts_default.o: otherarch/tts_adapter.cpp
+tts_default.o: otherarch/tts_adapter.cpp 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
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 embeddings_default.o: otherarch/embeddings_adapter.cpp
@@ -729,6 +729,8 @@ mainvk: tools/main/main.cpp common/arg.cpp build-info.h ggml_v4_vulkan.o ggml-cp
 	$(CXX) $(CXXFLAGS) -DGGML_USE_VULKAN -DSD_USE_VULKAN $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 embedding: examples/embedding/embedding.cpp common/arg.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 llavaclip_default.o llava.o ggml-backend_default.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -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 llavaclip_default.o llava.o ggml-backend_default.o ggml-backend-reg_default.o ggml-repack.o $(OBJS_FULL) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ggml/src/ggml-vulkan-shaders.cpp:
 ifdef VULKAN_BUILD

README.md

@@ -191,6 +191,7 @@ and it will install everything required. Alternatively, you can download the abo
 - KoboldCpp code and other files are also under the AGPL v3.0 License unless otherwise stated
 - Llama.cpp source repo is at https://github.com/ggml-org/llama.cpp (MIT)
 - Stable-diffusion.cpp source repo is at https://github.com/leejet/stable-diffusion.cpp (MIT)
+- TTS.cpp source repo is at https://github.com/mmwillet/TTS.cpp (MIT)
 - KoboldCpp source repo is at https://github.com/LostRuins/koboldcpp (AGPL)
 - KoboldAI Lite source repo is at https://github.com/LostRuins/lite.koboldai.net (AGPL)
 - For any further enquiries, contact @concedo on discord, or LostRuins on github.

convert_hf_to_gguf.py

@@ -1334,6 +1334,12 @@ def _find_param(self, obj: dict[str, Any], keys: Iterable[str], optional: bool =
             return None
         raise KeyError(f"could not find any of: {keys}")
 
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd.weight" in new_name:
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
+        return False
+
 
 @ModelBase.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(TextModel):
@@ -2305,10 +2311,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_use_gelu(True)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".embeddings." in name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -3296,12 +3301,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -3374,10 +3376,9 @@ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         yield ("audio_tower.embed_positions.weight", pos_embd)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".conv" in name and ".weight" in name:
             return gguf.GGMLQuantizationType.F16
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.startswith("thinker."):
@@ -3423,12 +3424,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def _mapping_interns1_name(self, name):
         names_map = {
@@ -5062,13 +5060,12 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         # related to https://github.com/ggml-org/llama.cpp/issues/13025
         if "input_projection" in name:
             return gguf.GGMLQuantizationType.F16
         if ".embeddings." in name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -7727,10 +7724,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".conv" in name and ".weight" in name:
             return gguf.GGMLQuantizationType.F16
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -8251,8 +8247,7 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling.get("original_max_position_embeddings", 4096))
 
 
-@ModelBase.register("Lfm2ForCausalLM")
-@ModelBase.register("LFM2ForCausalLM")
+@ModelBase.register("Lfm2ForCausalLM", "LFM2ForCausalLM")
 class LFM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.LFM2
 
@@ -8287,13 +8282,55 @@ def set_gguf_parameters(self):
         self._add_feed_forward_length()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+        if is_vision_tensor:
+            # skip vision tensors
+            return []
+
+        name = name.replace("language_model.", "")
+
         # conv op requires 2d tensor
         if 'conv.conv' in name:
             data_torch = data_torch.squeeze(1)
 
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Lfm2VlForConditionalGeneration")
+class LFM2VLModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        # TODO(tarek): for dynamic resolution image_size is not specified, setting here for compatibility
+        self.hparams_vision["image_size"] = 256
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LFM2)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["layer_norm_eps"]))
+        self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("downsample_factor", 2))
+        self.gguf_writer.add_vision_use_gelu(True)
+        # python notation, e.g. for vision_feature_layer == -1, we pick last layer -> vision_feature_layers_to_drop = 0
+        vision_feature_layers_to_drop = -(self.global_config.get("vision_feature_layer", -1) + 1)
+        self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys) - vision_feature_layers_to_drop)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            # remove "model." prefix
+            name = name.replace("model.vision_tower.", "vision_tower.")
+            name = name.replace("model.multi_modal_projector.", "multi_modal_projector.")
+
+            if "patch_embedding.weight" in name:
+                data_torch = data_torch.view(data_torch.shape[0], 16, 16, 3).permute(0, 3, 1, 2)
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
+
 @ModelBase.register("SmallThinkerForCausalLM")
 class SmallThinkerModel(TextModel):
     model_arch = gguf.MODEL_ARCH.SMALLTHINKER

examples/diffusion/CMakeLists.txt

@@ -249,6 +249,7 @@ struct tts_generation_inputs
     const char * prompt = nullptr;
     const int speaker_seed = 0;
     const int audio_seed = 0;
+    const char * custom_speaker_voice = "";
     const char * custom_speaker_text = "";
     const char * custom_speaker_data = "";
 };

expose.h

@@ -553,6 +553,18 @@ extern "C" {
         GGML_OP_GLU,
 
         GGML_OP_COUNT,
+
+        //kcpp: dirtypatch of unofficial ops for ttscpp
+        GGML_OP_UPSCALE_LINEAR, // linear interpolate
+        GGML_OP_RECIPROCAL,
+        GGML_OP_ROUND,
+        GGML_OP_MOD,
+        GGML_OP_CUMSUM,
+        GGML_OP_STFT,
+        GGML_OP_AA_STFT,
+        GGML_OP_ISTFT,
+        GGML_OP_AA_ISTFT,
+        GGML_OP_CONV_TRANSPOSE_1D_TTS,
     };
 
     enum ggml_unary_op {
@@ -2475,6 +2487,76 @@ extern "C" {
     GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
     GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
 
+    //kcpp: dirtypatch of ttscpp additions
+    GGML_API struct ggml_tensor * ggml_round(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a);
+    GGML_API struct ggml_tensor * ggml_round_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a);
+    // This is a floating point mod by the mod_val parameter
+    GGML_API struct ggml_tensor * ggml_mod(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a,
+            float mod_val);
+    GGML_API struct ggml_tensor * ggml_mod_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a,
+            float mod_val);
+    // reciprocal of each value in tensor a
+    GGML_API struct ggml_tensor * ggml_reciprocal(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+    GGML_API struct ggml_tensor * ggml_reciprocal_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+        // cumulative sums along first axis (ne0)
+    GGML_API struct ggml_tensor * ggml_cumsum(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+    GGML_API struct ggml_tensor * ggml_conv_1d_dw_tts(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+    GGML_API struct ggml_tensor * ggml_conv_transpose_1d_tts(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0,  // dilation
+            int                   op0, // output padding
+            int                   g0); // groups
+    GGML_API struct ggml_tensor * ggml_conv_1d_tts(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+    GGML_API struct ggml_tensor * ggml_stft(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a,
+            struct ggml_tensor * w, // the window
+            int filter_length,
+            int hop_length,
+            bool compute_abs_and_angle);
+    GGML_API struct ggml_tensor * ggml_istft(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a, // magnitude + phase
+            struct ggml_tensor * w,
+            int filter_length,
+            int hop_length,
+            bool from_abs_and_angle);
+    GGML_API struct ggml_tensor * ggml_upscale_linear(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   scale_factor);
+    //end kcpp dirtypatch
+
 #ifdef  __cplusplus
 }
 #endif