starvla: integrate seven Qwen policy variants

Author: JJJYmmm

CMakeLists.txt

@@ -1,9 +1,12 @@
 cmake_minimum_required(VERSION 3.16)
 project(robotcpp VERSION 0.1.0 LANGUAGES C CXX)
 
+include(CTest)
+
 option(ROBOT_CPP_BUILD_ROBOT_SERVER "Build model-server target" ON)
 option(ROBOT_CPP_BUILD_MODEL_CLI "Build model-cli target" OFF)
 option(ROBOT_CPP_BUILD_ROBOT_CLIENT "Build C++ robot client targets" OFF)
+option(ROBOT_CPP_BUILD_STARVLA "Build the StarVLA runtime (requires llama.cpp overlay)" OFF)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
@@ -12,6 +15,27 @@ set(CMAKE_CXX_EXTENSIONS OFF)
 if(NOT EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/CMakeLists.txt")
     message(FATAL_ERROR "third_party/llama.cpp is required; run `git submodule update --init --recursive`")
 endif()
+if(ROBOT_CPP_BUILD_STARVLA)
+    file(READ
+        "${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/tools/mtmd/models/qwen3vl.cpp"
+        ROBOT_CPP_QWEN3VL_MTMD_SOURCE)
+    file(READ
+        "${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/include/llama.h"
+        ROBOT_CPP_LLAMA_PUBLIC_HEADER)
+    string(FIND "${ROBOT_CPP_QWEN3VL_MTMD_SOURCE}"
+        "FFN_GELU_ERF" ROBOT_CPP_QWEN3VL_PARITY_PATCH_INDEX)
+    string(FIND "${ROBOT_CPP_LLAMA_PUBLIC_HEADER}"
+        "llama_set_backend_native_graphs_enabled" ROBOT_CPP_LLAMA_GRAPH_PATCH_INDEX)
+    if(ROBOT_CPP_QWEN3VL_PARITY_PATCH_INDEX EQUAL -1 OR
+       ROBOT_CPP_LLAMA_GRAPH_PATCH_INDEX EQUAL -1)
+        message(FATAL_ERROR
+            "StarVLA requires the pinned llama.cpp overlay. Run "
+            "`./tools/apply_patches.sh` from the repository root, "
+            "then configure again.")
+    endif()
+    unset(ROBOT_CPP_QWEN3VL_MTMD_SOURCE)
+    unset(ROBOT_CPP_LLAMA_PUBLIC_HEADER)
+endif()
 set(LLAMA_BUILD_COMMON ON CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_TOOLS OFF CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
@@ -20,7 +44,16 @@ add_subdirectory(third_party/llama.cpp EXCLUDE_FROM_ALL)
 if(NOT TARGET ggml OR NOT TARGET llama)
     message(FATAL_ERROR "llama.cpp must provide ggml and llama targets")
 endif()
-
+if(ROBOT_CPP_BUILD_STARVLA)
+    # mtmd normally inherits this variable when llama.cpp builds all tools.
+    if(NOT LLAMA_INSTALL_VERSION)
+        set(LLAMA_INSTALL_VERSION ${PROJECT_VERSION})
+    endif()
+    add_subdirectory(third_party/llama.cpp/tools/mtmd EXCLUDE_FROM_ALL)
+    if(NOT TARGET mtmd)
+        message(FATAL_ERROR "llama.cpp must provide the mtmd target for Qwen-VL")
+    endif()
+endif()
 set(ROBOT_CPP_LLAMA_INCLUDE_DIRS
     ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/include
@@ -30,14 +63,29 @@ set(ROBOT_CPP_LLAMA_INCLUDE_DIRS
 )
 
 set(SMOLVLA_DIR ${CMAKE_CURRENT_SOURCE_DIR}/src/models/smolvla)
+set(STARVLA_DIR ${CMAKE_CURRENT_SOURCE_DIR}/src/models/starvla)
 set(ROBOT_SERVER_DIR ${CMAKE_CURRENT_SOURCE_DIR}/robot_server)
 set(ROBOT_CLIENT_DIR ${CMAKE_CURRENT_SOURCE_DIR}/robot_client)
 
-add_library(smolvla_runtime STATIC
+add_library(robotcpp_model_common STATIC
     src/models/ggml_backend.cpp
     src/models/ggml_backend.h
     src/models/gguf_loader.cpp
     src/models/gguf_loader.h
+    src/models/model_type.cpp
+)
+target_include_directories(robotcpp_model_common
+    PUBLIC
+        ${CMAKE_CURRENT_SOURCE_DIR}/src
+        ${ROBOT_CPP_LLAMA_INCLUDE_DIRS}
+)
+target_link_libraries(robotcpp_model_common PUBLIC ggml)
+target_compile_features(robotcpp_model_common PUBLIC cxx_std_17)
+if(NOT MSVC)
+    target_compile_options(robotcpp_model_common PRIVATE -Wno-cast-qual)
+endif()
+
+add_library(smolvla_runtime STATIC
     ${SMOLVLA_DIR}/smolvla_engine.cpp
     ${SMOLVLA_DIR}/smolvla_engine.h
     ${SMOLVLA_DIR}/state_proj.cpp
@@ -53,17 +101,13 @@ target_include_directories(smolvla_runtime
         ${SMOLVLA_DIR}
         ${ROBOT_CPP_LLAMA_INCLUDE_DIRS}
 )
-target_link_libraries(smolvla_runtime PUBLIC ggml llama)
+target_link_libraries(smolvla_runtime PUBLIC robotcpp_model_common llama)
 target_compile_features(smolvla_runtime PUBLIC cxx_std_17)
 if(NOT MSVC)
     target_compile_options(smolvla_runtime PRIVATE -Wno-cast-qual)
 endif()
 
 add_library(pi0_engine STATIC
-    src/models/ggml_backend.cpp
-    src/models/ggml_backend.h
-    src/models/gguf_loader.cpp
-    src/models/gguf_loader.h
     src/models/pi0/types.h
     src/models/pi0/action.cpp
     src/models/pi0/action.h
@@ -87,13 +131,65 @@ target_include_directories(pi0_engine
         ${CMAKE_CURRENT_SOURCE_DIR}/src
         ${ROBOT_CPP_LLAMA_INCLUDE_DIRS}
 )
-target_link_libraries(pi0_engine PUBLIC ggml llama)
+target_link_libraries(pi0_engine PUBLIC robotcpp_model_common llama)
 target_compile_features(pi0_engine PUBLIC cxx_std_17)
 if(NOT MSVC)
     target_compile_options(pi0_engine PRIVATE -Wno-cast-qual)
 endif()
 
+if(ROBOT_CPP_BUILD_STARVLA)
+add_library(starvla_runtime STATIC
+    ${STARVLA_DIR}/fast_codec.cpp
+    ${STARVLA_DIR}/fast_codec.h
+    ${STARVLA_DIR}/fast_policy.cpp
+    ${STARVLA_DIR}/fast_policy.h
+    ${STARVLA_DIR}/groot_policy.cpp
+    ${STARVLA_DIR}/groot_policy.h
+    ${STARVLA_DIR}/groot_prompt.cpp
+    ${STARVLA_DIR}/groot_prompt.h
+    ${STARVLA_DIR}/normalization.cpp
+    ${STARVLA_DIR}/normalization.h
+    ${STARVLA_DIR}/oft_image_preprocess.cpp
+    ${STARVLA_DIR}/oft_image_preprocess.h
+    ${STARVLA_DIR}/oft_prompt.cpp
+    ${STARVLA_DIR}/oft_prompt.h
+    ${STARVLA_DIR}/oft_policy.cpp
+    ${STARVLA_DIR}/oft_policy.h
+    ${STARVLA_DIR}/pi_policy.cpp
+    ${STARVLA_DIR}/pi_policy.h
+    ${STARVLA_DIR}/pi_v3_policy.cpp
+    ${STARVLA_DIR}/pi_v3_policy.h
+    ${STARVLA_DIR}/policy_gguf.h
+    ${STARVLA_DIR}/qwen3vl_bridge.cpp
+    ${STARVLA_DIR}/qwen3vl_bridge.h
+    ${STARVLA_DIR}/starvla_engine.cpp
+    ${STARVLA_DIR}/starvla_engine.h
+    third_party/llama.cpp/examples/gguf-hash/deps/sha256/sha256.c
+)
+target_include_directories(starvla_runtime
+    PUBLIC
+        ${CMAKE_CURRENT_SOURCE_DIR}/src
+        ${ROBOT_CPP_LLAMA_INCLUDE_DIRS}
+    PRIVATE
+        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/vendor
+        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/examples/gguf-hash/deps
+        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/examples/gguf-hash/deps/sha256
+)
+target_link_libraries(starvla_runtime PUBLIC robotcpp_model_common llama mtmd)
+target_compile_features(starvla_runtime PUBLIC cxx_std_17)
+if(GGML_CUDA)
+    enable_language(CUDA)
+    target_sources(starvla_runtime PRIVATE ${STARVLA_DIR}/qwen_bf16_round_cuda.cu)
+    target_compile_definitions(starvla_runtime PRIVATE ROBOTCPP_STARVLA_CUDA=1)
+    set_property(TARGET starvla_runtime PROPERTY CUDA_STANDARD 17)
+endif()
+if(NOT MSVC)
+    target_compile_options(starvla_runtime PRIVATE -Wno-cast-qual)
+endif()
+endif()
+
 add_library(robotcpp STATIC
+    src/models/argument_parse.h
     src/models/model.h
     src/models/model_factory.cpp
     src/models/pi0/pi0_model.cpp
@@ -108,6 +204,19 @@ target_include_directories(robotcpp
 )
 target_link_libraries(robotcpp PUBLIC smolvla_runtime pi0_engine)
 target_compile_features(robotcpp PUBLIC cxx_std_17)
+if(ROBOT_CPP_BUILD_STARVLA)
+    target_sources(robotcpp PRIVATE
+        ${STARVLA_DIR}/starvla_model.cpp
+        ${STARVLA_DIR}/starvla_model.h)
+    target_link_libraries(robotcpp PUBLIC starvla_runtime)
+    target_compile_definitions(robotcpp PUBLIC ROBOT_CPP_BUILD_STARVLA=1)
+endif()
+
+if(BUILD_TESTING AND ROBOT_CPP_BUILD_STARVLA)
+    add_executable(robotcpp-starvla-model-test tests/starvla/model_test.cpp)
+    target_link_libraries(robotcpp-starvla-model-test PRIVATE robotcpp)
+    add_test(NAME robotcpp-starvla-model-test COMMAND robotcpp-starvla-model-test)
+endif()
 
 if(ROBOT_CPP_BUILD_ROBOT_SERVER OR ROBOT_CPP_BUILD_ROBOT_CLIENT)
     add_library(robot_server_common STATIC
@@ -139,6 +248,7 @@ if(ROBOT_CPP_BUILD_ROBOT_CLIENT)
     set_target_properties(model-cpp-client-example PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
     target_link_libraries(model-cpp-client-example PRIVATE model_client_cpp)
     target_compile_features(model-cpp-client-example PRIVATE cxx_std_17)
+
 endif()
 
 if(ROBOT_CPP_BUILD_ROBOT_SERVER)
@@ -165,7 +275,6 @@ if(ROBOT_CPP_BUILD_ROBOT_SERVER)
     )
     target_link_libraries(model-server PRIVATE robot_server_core robotcpp)
     target_compile_features(model-server PRIVATE cxx_std_17)
-
     add_executable(smolvla-raw-predict ${ROBOT_SERVER_DIR}/test/smolvla_raw_predict.cpp)
     set_target_properties(smolvla-raw-predict PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
     target_include_directories(smolvla-raw-predict PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src ${SMOLVLA_DIR})

patches/llama.cpp/README.md

@@ -1,21 +1,20 @@
-# StarVLA llama.cpp patches
+# llama.cpp patches
 
 The project pins `third_party/llama.cpp` at commit
-`3e941b813b1acbbf06c2203a94ceb33d84748c1e`. The StarVLA runtime needs two
-small changes that are not available through that revision's public APIs:
+`3e941b813b1acbbf06c2203a94ceb33d84748c1e`. The repository applies two
+changes that are not available through that revision's public APIs:
 
-1. `0001-qwen3vl-vision-parity.patch` matches the upstream Qwen3-VL reference
-   implementation's position interpolation and exact GELU operations. These
-   changes are required for action-value parity with the original checkpoint.
+1. `0001-qwen3vl-vision-parity.patch` uses the position interpolation and exact
+   GELU operations from the Qwen3-VL implementation used by StarVLA.
 2. `0002-per-context-native-graph-control.patch` adds an optional backend API to
    disable CUDA graph capture for the text and vision contexts owned by one
-   StarVLA instance. It prevents retained CUDA graphs from violating long-loop
-   memory stability gates without globally changing other llama.cpp users.
+   StarVLA instance. This avoids retained CUDA graphs growing memory use during
+   long runs without changing the setting for other llama.cpp users.
 
-Apply both patches before configuring or building the StarVLA runtime:
+Apply the repository patch set after initializing submodules and before building:
 
 ```bash
-./tools/llama_cpp/apply_starvla_patches.sh
+./tools/apply_patches.sh
 ```
 
 The command verifies the exact llama.cpp revision and refuses a dirty or
@@ -24,8 +23,8 @@ partially patched checkout. It is safe to run again after a complete apply.
 Inspect or remove the overlay with:
 
 ```bash
-./tools/llama_cpp/apply_starvla_patches.sh --check
-./tools/llama_cpp/apply_starvla_patches.sh --revert
+./tools/apply_patches.sh --check
+./tools/apply_patches.sh --revert
 ```
 
 The parent repository commits only these patch assets. It does not advance or

robot_client/cpp/model_client.cpp

@@ -57,6 +57,7 @@ bool make_predict_request(const ModelObservation & obs, proto::predict_request &
 
     req.task  = obs.prompt;
     req.state = obs.state;
+    req.initial_noise = obs.initial_noise;
     req.images.clear();
     req.images.reserve(obs.images.size());
 

robot_client/cpp/model_client.h

@@ -21,6 +21,7 @@ struct ModelImage {
 struct ModelObservation {
     std::vector<ModelImage> images;
     std::vector<float> state;
+    std::vector<float> initial_noise;
     std::string prompt = "grab the block.";
 };
 

robot_client/python/model_client.py

@@ -9,7 +9,7 @@
 
 
 MAGIC = 0x414C5653
-VERSION = 3
+VERSION = 4
 HEADER_SIZE = 32
 
 OP_HEALTH = 1
@@ -21,8 +21,8 @@
 IMAGE_RAW_RGB_U8 = 1
 
 HEADER = struct.Struct("<IHHHHIIQI")
-PREDICT_REQ_V2_FIXED = struct.Struct("<III")
-PREDICT_REQ_V2_IMAGE = struct.Struct("<IIIIIIQ")
+PREDICT_REQ_FIXED = struct.Struct("<IIII")
+PREDICT_REQ_IMAGE = struct.Struct("<IIIIIIQ")
 PREDICT_RESP_FIXED = struct.Struct("<IIII")
 PREDICT_RESP_METRIC = struct.Struct("<Id")
 
@@ -76,6 +76,7 @@ def _recv_message(sock: socket.socket) -> tuple[int, int, int, bytes]:
 def encode_predict_observation(observation: dict[str, Any]) -> bytes:
     images = observation["images"]
     state = state_to_list(observation["state"])
+    initial_noise = state_to_list(observation.get("initial_noise"))
     prompt = str(observation["prompt"])
     if not images:
         raise ValueError("observation.images must contain at least one image")
@@ -88,13 +89,14 @@ def encode_predict_observation(observation: dict[str, Any]) -> bytes:
         encoded_images.append((name, rgb, width, height, stride))
 
     payload = bytearray()
-    payload += PREDICT_REQ_V2_FIXED.pack(
+    payload += PREDICT_REQ_FIXED.pack(
         len(encoded_images),
         len(state),
+        len(initial_noise),
         len(prompt_bytes),
     )
     for name, rgb, width, height, stride in encoded_images:
-        payload += PREDICT_REQ_V2_IMAGE.pack(
+        payload += PREDICT_REQ_IMAGE.pack(
             IMAGE_RAW_RGB_U8,
             len(name),
             width,
@@ -105,6 +107,8 @@ def encode_predict_observation(observation: dict[str, Any]) -> bytes:
         )
     for value in state:
         payload += struct.pack("<f", float(value))
+    for value in initial_noise:
+        payload += struct.pack("<f", float(value))
     payload += prompt_bytes
     for name, rgb, _width, _height, _stride in encoded_images:
         payload += name