Transpiler complete

.github/workflows/verify-pr.yml

@@ -106,26 +106,26 @@ jobs:
         run: |
           source venv/bin/activate
           convert_if_missing() {
-            local model_id=$1 output_dir=$2 precision=$3
+            local model_id=$1 output_dir=$2 bits=$3
             if [ ! -f "weights/$output_dir/config.txt" ]; then
               echo "Converting $model_id -> weights/$output_dir"
-              python3 -m src.cli convert "$model_id" "weights/$output_dir" --precision "$precision" ${HF_TOKEN:+--token "$HF_TOKEN"} || echo "WARNING: conversion of $model_id failed (exit $?), checking fallback directory"
+              python3 -m cactus convert "$model_id" "weights/$output_dir" --bits "$bits" ${HF_TOKEN:+--token "$HF_TOKEN"} || echo "WARNING: conversion of $model_id failed (exit $?), checking fallback directory"
               local default_dir="weights/$(echo "$model_id" | sed 's|.*/||' | tr '[:upper:]' '[:lower:]')"
               if [ ! -f "weights/$output_dir/config.txt" ] && [ -f "$default_dir/config.txt" ]; then
                 mv "$default_dir" "weights/$output_dir"
               fi
             fi
           }
-          convert_if_missing "Qwen/Qwen3-0.6B" "qwen3-0.6b" "INT8"
-          convert_if_missing "nomic-ai/nomic-embed-text-v2-moe" "nomic-embed-text-v2-moe" "INT8"
-          convert_if_missing "LiquidAI/LFM2.5-350M" "lfm2-350m" "INT8"
-          convert_if_missing "LiquidAI/LFM2-8B-A1B" "lfm2-8b-a1b" "INT8"
-          convert_if_missing "google/gemma-3-270m-it" "gemma-3-270m-it" "INT8"
-          convert_if_missing "openai/whisper-small" "whisper-small" "INT8"
-          convert_if_missing "UsefulSensors/moonshine-base" "moonshine-base" "INT8"
-          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-int8" "INT8"
-          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-int4" "INT4"
-          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-fp16" "FP16"
+          convert_if_missing "Qwen/Qwen3-0.6B" "qwen3-0.6b" "4"
+          convert_if_missing "nomic-ai/nomic-embed-text-v2-moe" "nomic-embed-text-v2-moe" "4"
+          convert_if_missing "LiquidAI/LFM2.5-350M" "lfm2-350m" "4"
+          convert_if_missing "LiquidAI/LFM2-8B-A1B" "lfm2-8b-a1b" "4"
+          convert_if_missing "google/gemma-3-270m-it" "gemma-3-270m-it" "4"
+          convert_if_missing "openai/whisper-small" "whisper-small" "4"
+          convert_if_missing "UsefulSensors/moonshine-base" "moonshine-base" "4"
+          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-int8" "4"
+          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-int4" "4"
+          convert_if_missing "LiquidAI/LFM2-VL-450M" "lfm2-vl-450m-fp16" "4"
 
       - name: Build library
         run: bash cactus/build.sh

cactus-engine/CMakeLists.txt

@@ -74,29 +74,17 @@ set(ENGINE_SOURCES
     src/npu.cpp
 )
 
-set(MODEL_SOURCES
-    models/model_lfm2.cpp
-    models/model_lfm2vl.cpp
-    models/model_qwen.cpp
-    models/model_siglip2.cpp
-    models/gemma4/model_gemma4.cpp
-    models/gemma4/model_gemma4_mm.cpp
-    models/gemma4/model_gemma4_audio.cpp
-    models/gemma4/model_gemma4_vision.cpp
-)
-
 if(APPLE)
     enable_language(OBJCXX)
     list(APPEND ENGINE_SOURCES src/npu_ane.mm)
     set_source_files_properties(src/npu_ane.mm PROPERTIES COMPILE_FLAGS "-fobjc-arc")
 endif()
 
-add_library(cactus_engine STATIC ${ENGINE_SOURCES} ${MODEL_SOURCES})
+add_library(cactus_engine STATIC ${ENGINE_SOURCES})
 set(CACTUS_KERNELS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../cactus-kernels")
 target_include_directories(cactus_engine
     PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}
     PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src
-            ${CMAKE_CURRENT_SOURCE_DIR}/models
             ${CACTUS_KERNELS_DIR}/src
             ${CACTUS_KERNELS_DIR}/libs
 )

cactus-engine/cactus_engine.h

@@ -292,6 +292,7 @@ CACTUS_FFI_EXPORT int cactus_graph_mean(cactus_graph_t graph, cactus_node_t x, i
 CACTUS_FFI_EXPORT int cactus_graph_variance(cactus_graph_t graph, cactus_node_t x, int32_t axis, cactus_node_t* out);
 CACTUS_FFI_EXPORT int cactus_graph_min(cactus_graph_t graph, cactus_node_t x, int32_t axis, cactus_node_t* out);
 CACTUS_FFI_EXPORT int cactus_graph_max(cactus_graph_t graph, cactus_node_t x, int32_t axis, cactus_node_t* out);
+CACTUS_FFI_EXPORT int cactus_graph_cumsum(cactus_graph_t graph, cactus_node_t x, int32_t axis, cactus_node_t* out);
 
 CACTUS_FFI_EXPORT int cactus_graph_concat(
     cactus_graph_t graph, cactus_node_t a, cactus_node_t b, int32_t axis,
@@ -312,6 +313,8 @@ CACTUS_FFI_EXPORT int cactus_graph_mmap_embeddings(
     cactus_graph_t graph, const char* filename, cactus_node_t* out);
 CACTUS_FFI_EXPORT int cactus_graph_mmap_weights(
     cactus_graph_t graph, const char* filename, cactus_node_t* out);
+CACTUS_FFI_EXPORT int cactus_graph_bind_mmap_weights(
+    cactus_graph_t graph, cactus_node_t node, const char* filename);
 CACTUS_FFI_EXPORT int cactus_graph_bilinear_interpolation(
     cactus_graph_t graph, cactus_node_t pos_embeds, size_t dst_height, size_t dst_width, cactus_node_t* out);
 CACTUS_FFI_EXPORT int cactus_graph_release_weight_pages(cactus_graph_t graph, cactus_node_t node);