Perf migration

.dockerignore

@@ -0,0 +1,10 @@
+# Keep the build context small: llama.cpp is fetched inside the image, models
+# are mounted at runtime, and build/output dirs are host artifacts.
+.git/
+models/
+build/
+build-*/
+outputs/
+third_party/
+eval/sim/
+*.gguf

.github/workflows/build.yml

@@ -0,0 +1,52 @@
+# Fast build + unit gate on GitHub-hosted runners. The heavier cross-platform
+# LIBERO regression sweep lives in vla-ci.yml (self-hosted).
+name: build
+
+on:
+  push:
+    branches: [main, dev]
+  pull_request:
+    branches: [main]
+
+concurrency:
+  group: build-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  cpp-unit:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - name: pixel-shuffle channel order
+        run: |
+          g++ -std=c++17 -Isrc -Wall -Wextra tests/test_vision_common.cpp -o /tmp/test_vision_common
+          /tmp/test_vision_common
+
+  py-tooling:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+      - name: converter remap
+        run: python tests/py/test_converters.py
+
+  build-gate:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - name: deps
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y -qq --no-install-recommends \
+            build-essential cmake git ca-certificates pkg-config \
+            libzmq3-dev cppzmq-dev libprotobuf-dev protobuf-compiler
+      - uses: actions/cache@v4
+        with:
+          path: build/_deps
+          key: llama-b9866-${{ runner.os }}
+      - name: build vla-server + vlm-server + vla-cli (CPU, -Wall -Wextra)
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=OFF
+          cmake --build build -j"$(nproc)" --target vla-server vlm-server vla-cli

.gitignore

@@ -53,3 +53,4 @@ eval/sim/libero/LIBERO/
 eval/sim/libero/libero_uv/
 eval/sim/simpler/SimplerEnv/
 eval/sim/simpler/simpler_uv/
+/models/

CHANGELOG.md

@@ -0,0 +1,44 @@
+# Changelog
+
+Notable changes to vla.cpp. Format loosely follows [Keep a Changelog](https://keepachangelog.com).
+
+## [0.1.1] - 2026-07-04
+
+### Added
+- `vla-cli`: one-shot inference from the command line (image + tokens to action), no server needed.
+- `scripts/quantize_gguf.py`: repack LM weights to Q8_0/Q4_0. The loader runs quantized GGUFs directly (Q8_0 roughly halves the LM, near-lossless).
+
+### Changed
+- One shared GGUF reader across the model loaders, replacing the per-arch copies.
+- Cap inbound message size (256 MiB) and image count (16) on `vla-server` and `vlm-server`.
+- Scale CPU threads to the machine core count across all loaders instead of a fixed 4.
+- Read GGUF file offsets as 64-bit and reject non-float embedding tensors in row-fetch.
+
+### Fixed
+- Reject out-of-range language tokens in OpenVLA-OFT and VLA-Adapter.
+- Zero the padded action dimensions so only real action dims carry values.
+- Reject images that do not match the model input size in VLA-Adapter and OpenVLA-OFT (out-of-bounds read on a smaller view).
+- Validate Evo-1 action dims at load so a client-supplied noise buffer cannot underrun.
+- Only enable the BitVLA CUDA path once every device buffer allocates.
+
+## [0.1.0] - 2026-07-03
+
+First tagged release. One self-contained GGUF per model (vision tower + LM + action
+expert + dataset stats), CPU or CUDA, no external mmproj and no patch to llama.cpp.
+
+### Added
+- Seven VLA policies auto-detected from the GGUF: SmolVLA, pi0, BitVLA, Evo-1, GR00T N1.5/N1.6/N1.7.
+- In-tree vision towers (SigLIP, BitSigLIP, InternViT, RADIO) on stable public ggml/llama APIs.
+- ZeroMQ + protobuf `vla-server`; a separate `vlm-server` for VLM chat.
+- BitVLA 1.58-bit custom ternary CUDA kernels.
+- Per-arch HuggingFace -> GGUF converters and mmproj-merge helpers (`scripts/`).
+- Robot eval harness for LIBERO, SimplerEnv, and ALOHA (`eval/`), with device benchmark reports.
+- Minimal CI: pixel-shuffle unit test, converter-remap test, CPU build gate.
+- `pyproject.toml` for the Python tooling and a CUDA `Dockerfile` for `vla-server`.
+
+### Changed
+- llama.cpp is fetched + pinned via CMake `FetchContent` (tag `b9866`); bumping is a
+  one-line `GIT_TAG` change. Removed the `patches/` fetch script.
+
+[0.1.1]: https://github.com/VinRobotics/vla.cpp/releases/tag/v0.1.1
+[0.1.0]: https://github.com/VinRobotics/vla.cpp/releases/tag/v0.1.0

CMakeLists.txt

@@ -16,7 +16,14 @@ set(LLAMA_BUILD_TOOLS    ON  CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_SERVER   OFF CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_TESTS    OFF CACHE BOOL "" FORCE)
-add_subdirectory(third_party/llama.cpp)
+# llama.cpp fetched + pinned at configure; bump = one-line GIT_TAG change.
+include(FetchContent)
+FetchContent_Declare(llama
+    GIT_REPOSITORY https://github.com/ggml-org/llama.cpp
+    GIT_TAG        b9866
+    GIT_SHALLOW    TRUE
+)
+FetchContent_MakeAvailable(llama)
 
 add_library(vla_core
     src/model.cpp
@@ -35,18 +42,31 @@ add_library(vla_core
 target_include_directories(vla_core
     PUBLIC
         ${CMAKE_CURRENT_SOURCE_DIR}/src
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/tools/mtmd
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/vendor
+        ${llama_SOURCE_DIR}/vendor
 )
-target_link_libraries(vla_core PUBLIC llama ggml mtmd)
+target_link_libraries(vla_core PUBLIC llama ggml)
 
 if(GGML_CUDA)
     target_compile_definitions(vla_core PUBLIC GGML_USE_CUDA)
 
     enable_language(CUDA)
     find_package(CUDAToolkit REQUIRED)
     if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES 80 86 87 89 90 100 120 CACHE STRING "" FORCE)
+        # Emit SASS per supported GPU (real) and PTX (virtual) only for the newest
+        # arch, so future cards can JIT without shipping PTX for every target.
+        # Blackwell sm_100/sm_120 require CUDA >= 12.8.
+        set(_vla_cuda_archs 80-real 86-real 87-real 89-real 90-real)
+        set(_vla_cuda_ptx 90-virtual)
+        if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 12.8)
+            list(APPEND _vla_cuda_archs 100-real 120-real)
+            set(_vla_cuda_ptx 120-virtual)
+        else()
+            message(WARNING
+                "CUDA ${CUDAToolkit_VERSION} < 12.8: omitting Blackwell sm_100/sm_120. "
+                "Pass -DCMAKE_CUDA_ARCHITECTURES=<arch> explicitly to override.")
+        endif()
+        set(CMAKE_CUDA_ARCHITECTURES ${_vla_cuda_archs} ${_vla_cuda_ptx}
+            CACHE STRING "" FORCE)
     endif()
     add_library(bitvla_cuda_kernels STATIC
         src/kernels/bitvla/bitnet_kernels.cu
@@ -79,16 +99,25 @@ target_include_directories(vlm_core
     PUBLIC
         ${CMAKE_CURRENT_SOURCE_DIR}/src
     PRIVATE
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/common
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/tools/mtmd
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/vendor
+        ${llama_SOURCE_DIR}/common
+        ${llama_SOURCE_DIR}/tools/mtmd
+        ${llama_SOURCE_DIR}/vendor
 )
 target_link_libraries(vlm_core PUBLIC llama ggml mtmd llama-common)
 
 find_package(Protobuf REQUIRED)
 find_package(PkgConfig REQUIRED)
 pkg_check_modules(ZeroMQ REQUIRED IMPORTED_TARGET libzmq)
 
+# cppzmq (zmq.hpp) is the header-only C++ binding, packaged separately from the
+# libzmq C library (Debian/Ubuntu: cppzmq-dev). The servers #include <zmq.hpp>.
+find_path(CPPZMQ_INCLUDE_DIR NAMES zmq.hpp)
+if(NOT CPPZMQ_INCLUDE_DIR)
+    message(FATAL_ERROR
+        "cppzmq (zmq.hpp) not found. Install 'cppzmq-dev' (Debian/Ubuntu) or set "
+        "CMAKE_PREFIX_PATH to your cppzmq install prefix.")
+endif()
+
 set(VLA_PROTO_SRC ${CMAKE_CURRENT_SOURCE_DIR}/src/serving/vla.proto)
 set(VLA_PROTO_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/proto-gen/serving)
 file(MAKE_DIRECTORY ${VLA_PROTO_GEN_DIR})
@@ -110,7 +139,8 @@ add_executable(vla-server
 )
 target_include_directories(vla-server PRIVATE
     ${CMAKE_CURRENT_BINARY_DIR}/proto-gen
-    ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/vendor/stb
+    ${llama_SOURCE_DIR}/vendor/stb
+    ${CPPZMQ_INCLUDE_DIR}
 )
 target_link_libraries(vla-server PRIVATE
     vla_core
@@ -139,9 +169,44 @@ add_executable(vlm-server
 )
 target_include_directories(vlm-server PRIVATE
     ${CMAKE_CURRENT_BINARY_DIR}/proto-gen
+    ${CPPZMQ_INCLUDE_DIR}
 )
 target_link_libraries(vlm-server PRIVATE
     vlm_core
     protobuf::libprotobuf
     PkgConfig::ZeroMQ
 )
+
+# One-shot inference CLI: image + tokens -> action, no server or simulator.
+add_executable(vla-cli
+    src/serving/vla-cli.cpp
+)
+target_include_directories(vla-cli PRIVATE
+    ${llama_SOURCE_DIR}/vendor/stb
+)
+target_link_libraries(vla-cli PRIVATE vla_core)
+
+# --- First-party build hygiene (never applied to the vendored llama.cpp subtree) --
+set(VLA_FIRST_PARTY_TARGETS vla_core vlm_core vla-server vlm-server vla-cli)
+
+foreach(tgt IN LISTS VLA_FIRST_PARTY_TARGETS)
+    # Warn on our own C++ only; nvcc device code keeps its own diagnostics.
+    target_compile_options(${tgt} PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-Wall -Wextra>)
+endforeach()
+
+# Link-time optimization on Release builds when the toolchain supports it.
+include(CheckIPOSupported)
+check_ipo_supported(RESULT VLA_IPO_OK OUTPUT VLA_IPO_MSG)
+if(VLA_IPO_OK AND CMAKE_BUILD_TYPE STREQUAL "Release")
+    foreach(tgt IN LISTS VLA_FIRST_PARTY_TARGETS)
+        set_property(TARGET ${tgt} PROPERTY INTERPROCEDURAL_OPTIMIZATION ON)
+    endforeach()
+else()
+    message(STATUS "vla: IPO/LTO disabled (${VLA_IPO_MSG})")
+endif()
+
+option(VLA_BUILD_TESTS "Build vla.cpp test harnesses and unit tests" OFF)
+if(VLA_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif()

Dockerfile

@@ -0,0 +1,39 @@
+# vla-server, CPU or CUDA. cmake fetches llama.cpp at build time.
+#   GPU sm_120 (default):  docker build -t vla-cpp .
+#   GPU other arch:        --build-arg CUDA_ARCH=89   (89=RTX40 90=H100 87=Orin; sm_120 needs CUDA>=12.8)
+#   older card:            --build-arg BASE_IMAGE=nvidia/cuda:12.4.1-devel-ubuntu24.04 --build-arg CUDA_ARCH=86
+#   CPU:                   --build-arg BACKEND=cpu --build-arg BASE_IMAGE=ubuntu:24.04 -t vla-cpp-cpu
+#   run:  docker run --gpus all -p5555:5555 -v $PWD/models:/models vla-cpp --bind tcp://*:5555 /models/M.gguf
+#         (CDI hosts use --device nvidia.com/gpu=all)
+
+ARG BASE_IMAGE=nvidia/cuda:12.9.1-devel-ubuntu24.04
+FROM ${BASE_IMAGE}
+
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        build-essential cmake git ca-certificates pkg-config \
+        libzmq3-dev cppzmq-dev libprotobuf-dev protobuf-compiler \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /src
+COPY . .
+
+ARG BACKEND=cuda
+ARG CUDA_ARCH=120
+# nvcc can segfault on the flash-attn kernels under high -j; lower JOBS if so.
+ARG JOBS=
+# CUDA: -devel ships only a libcuda stub (real driver injected at runtime), so
+# link ggml-cuda's driver calls against it. CPU build skips this.
+RUN set -eux; \
+    if [ "$BACKEND" = "cuda" ]; then \
+        export LIBRARY_PATH=/usr/local/cuda/lib64/stubs; \
+        cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DGGML_CUDA_GRAPHS=ON \
+              -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCH}" \
+              -DCMAKE_SHARED_LINKER_FLAGS="-lcuda" -DCMAKE_EXE_LINKER_FLAGS="-lcuda"; \
+    else \
+        cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=OFF; \
+    fi; \
+    cmake --build build -j"${JOBS:-$(nproc)}" --target vla-server vla-cli; \
+    cp build/vla-server build/vla-cli /usr/local/bin/
+
+EXPOSE 5555
+ENTRYPOINT ["vla-server"]