Refactor start.sh: streamline model download and argument building

Co-authored-by: Copilot <copilot@github.com>

docs/pipeline/README.md

@@ -18,6 +18,7 @@ descriptions via a local GGUF model or a deterministic mock.
   - [Build](#build)
   - [Model](#model)
   - [Run](#run)
+- [Docker / RunPod](#docker--runpod)
 - [Architecture](#architecture)
   - [Pipeline Stages](#pipeline-stages)
   - [Key Components](#key-components)
@@ -51,7 +52,7 @@ step.
 
 ### Build
 
-Requirements: C++20 compiler, CMake 3.24+, libcurl, Boost (JSON and
+Requirements: C++20 compiler, CMake 3.31+, OpenSSL, Boost (JSON and
 ProgramOptions). SQLite is fetched from the upstream amalgamation, so no system
 SQLite package is required.
 
@@ -60,6 +61,16 @@ cmake -S . -B build
 cmake --build build
 ```
 
+CMake automatically detects whether a compatible llama.cpp installation is
+present on the system (`libllama`, `libggml`, `libggml-base`, and `llama.h`
+visible on the default search paths). If found, it links against those
+libraries and skips the FetchContent build. If not found, it fetches and builds
+llama.cpp from source at tag `b9012`. No additional flags are required in
+either case.
+
+Metal is enabled automatically on Apple Silicon. CUDA or HIP/ROCm is detected
+automatically on Linux when the relevant toolkit is present.
+
 ### Model
 
 > Skip this step if you only need `--mocked`.
@@ -74,33 +85,124 @@ curl -L \
 ### Run
 
 Run from `build/` so the copied `locations.json` and `prompts/` are available.
-Each run also writes a fresh dated SQLite file such as
+Each run writes a fresh dated SQLite file such as
 `biergarten_seed_2026-04-19T15-30-45.123456Z.sqlite` into the working directory.
 
 ```bash
 ./biergarten-pipeline --mocked
-./biergarten-pipeline --model models/google_gemma-4-E4B-it-Q6_K.gguf --temperature 1.0 --top-p 0.95 --top-k 64 --n-ctx 8192 --seed -1
+
+./biergarten-pipeline \
+  --model ../models/google_gemma-4-E4B-it-Q6_K.gguf \
+  --prompt-dir prompts \
+  --temperature 1.0 --top-p 0.95 --top-k 64 --n-ctx 8192 --seed -1
 ```
 
 #### CLI Flags
 
-| Flag            | Purpose                                                 |
+| Flag            | Purpose                                                                                              |
-| --------------- | ------------------------------------------------------- |
+| --------------- | ---------------------------------------------------------------------------------------------------- |
-| `--mocked`      | Deterministic mock generator, no model required.        |
+| `--mocked`      | Deterministic mock generator, no model required.                                                     |
-| `--model, -m`   | Path to a GGUF file. Required unless `--mocked` is set. |
+| `--model, -m`   | Path to a GGUF file. Required unless `--mocked` is set.                                              |
-| `--temperature` | Sampling temperature. Default: `1.0`.                   |
+| `--prompt-dir`  | Directory containing prompt files (e.g. `BREWERY_GENERATION.md`). Required unless `--mocked` is set. |
-| `--top-p`       | Nucleus sampling. Default: `0.95`.                      |
+| `--output, -o`  | Directory for generated SQLite artifacts. Default: `output`.                                         |
-| `--top-k`       | Top-k sampling. Default: `64`.                          |
+| `--log-path`    | Path for application logs. Default: `pipeline.log`.                                                  |
-| `--n-ctx`       | Context window size. Default: `8192`.                   |
+| `--temperature` | Sampling temperature. Default: `1.0`.                                                                |
-| `--seed`        | Random seed. Default: `-1` (random at runtime).         |
+| `--top-p`       | Nucleus sampling. Default: `0.95`.                                                                   |
-| `--help, -h`    | Print usage and exit.                                   |
+| `--top-k`       | Top-k sampling. Default: `64`.                                                                       |
+| `--n-ctx`       | Context window size. Default: `8192`.                                                                |
+| `--seed`        | Random seed. Default: `-1` (random at runtime).                                                      |
+| `--help, -h`    | Print usage and exit.                                                                                |
 
 `--mocked` and `--model` are mutually exclusive. Omitting both exits with an
 error before the pipeline starts. Sampling flags are ignored when `--mocked` is
 set.
 
 The post-build step copies `prompts/` into `build/prompts/`. Rebuild after
-editing `prompts/system.md`.
+editing any prompt file.
+
+---
+
+## Docker / RunPod
+
+The `tooling/pipeline/runpod/` directory contains a GPU-ready container
+configuration for running the pipeline on RunPod or any Docker host with an
+NVIDIA GPU.
+
+### How it works
+
+The container uses a two-stage build. The first stage pulls prebuilt
+`libllama`, `libggml`, and backend plugin libraries (including `libggml-cuda.so`
+and the CPU variant plugins) from `ghcr.io/ggml-org/llama.cpp:full-cuda`. The
+second stage copies those libraries into `/usr/local/lib` and runs `ldconfig` so
+the dynamic linker and `dlopen` calls from `ggml_backend_load_all()` can resolve
+the CUDA backend plugin at runtime. llama.cpp headers are cloned at the matching
+tag and installed into `/usr/local/include`. CMake auto-detects both and skips
+the FetchContent source build entirely, keeping image build times short.
+
+`GGML_BACKEND_PATH` is set to `/usr/local/lib` so llama.cpp knows where to scan
+for backend plugins.
+
+### Build the image
+
+Run from the `tooling/pipeline/` directory (the CMake project root), not from
+inside `runpod/`, so the `COPY . .` step picks up the full project context.
+
+```bash
+docker build -t biergarten-pipeline:latest -f runpod/Dockerfile .
+```
+
+To monitor the full build output and confirm CMake selects the system llama.cpp:
+
+```bash
+docker build \
+  --progress=plain \
+  --no-cache \
+  -t biergarten-pipeline:latest \
+  -f runpod/Dockerfile \
+  . 2>&1 | tee build.log
+```
+
+Look for `[biergarten] Found system llama.cpp — skipping FetchContent` in the
+output to confirm the fast path was taken.
+
+### Run in mocked mode
+
+No model or GPU required. Useful for validating the pipeline logic and SQLite
+export path.
+
+```bash
+docker run --rm \
+  -e BIERGARTEN_MODE=mocked \
+  -v "$PWD/output:/workspace/output" \
+  -v "$PWD/logs:/workspace/logs" \
+  biergarten-pipeline:latest
+```
+
+### Run in live mode
+
+Mount your GGUF model before starting. The container validates the model path
+before launching the binary.
+
+```bash
+docker run --rm \
+  --runtime=nvidia \
+  -e BIERGARTEN_MODE=live \
+  -e GGML_BACKEND_PATH="/usr/local/lib/libggml-cuda.so" \
+  -v "$PWD/models:/workspace/models" \
+  -v "$PWD/output:/workspace/output" \
+  -v "$PWD/logs:/workspace/logs" \
+  biergarten-pipeline:latest
+```
+
+The model must be present at `./models/google_gemma-4-E4B-it-Q6_K.gguf` on the
+host. See [Model](#model) above for the download command.
+
+### RunPod deployment
+
+Use a GPU pod template. Mount persistent storage for `/workspace/models`,
+`/workspace/output`, and `/workspace/logs`. Set `BIERGARTEN_MODE=live` in the
+template environment. See `tooling/pipeline/runpod/pod-template.yaml` for a
+starter template.
 
 ---
 
@@ -197,16 +299,18 @@ code, latitude, and longitude for each entry.
 ## Tech Stack
 
 - C++20
-- CMake 3.24+
+- CMake 3.31+
 - Boost.JSON, Boost.ProgramOptions, Boost.DI
 - spdlog
-- libcurl
+- cpp-httplib (with OpenSSL)
 - SQLite amalgamation fetched and compiled via CMake FetchContent
-- llama.cpp
+- llama.cpp (auto-detected from system install or fetched via FetchContent)
+- Docker with NVIDIA CUDA 12.6 base image for GPU container builds
+- RunPod for cloud GPU inference
 
-The build fetches Boost.DI, spdlog, llama.cpp, and SQLite via CMake. Metal is
+The build fetches Boost.DI, spdlog, and SQLite via CMake. llama.cpp is fetched
-enabled on Apple Silicon; CUDA or HIP/ROCm is detected on Linux when the toolkit
+only when a system installation is not detected. Metal is enabled on Apple
-is present.
+Silicon; CUDA or HIP/ROCm is detected on Linux when the toolkit is present.
 
 > **Code Style:** Modern C++20 throughout — RAII for ownership,
 > `std::unique_ptr` for injected dependencies, `std::optional` for parse
@@ -218,7 +322,7 @@ is present.
 
 ## Tested Hardware
 
-### ARM macOS - M1 Pro
+### ARM macOS — M1 Pro
 
 |           |                                   |
 | --------- | --------------------------------- |
@@ -229,7 +333,7 @@ is present.
 | Model     | Gemma 4 E4B                       |
 | Inference | llama.cpp with Metal              |
 
-### x86_64 Linux - NVIDIA RTX 2000
+### x86_64 Linux — NVIDIA RTX 2000
 
 |           |                                |
 | --------- | ------------------------------ |
@@ -240,6 +344,15 @@ is present.
 | Model     | Gemma 4 E4B                    |
 | Inference | llama.cpp with CUDA 12.x       |
 
+### x86_64 Linux — Docker / RunPod (NVIDIA CUDA)
+
+|           |                                             |
+| --------- | ------------------------------------------- |
+| Host      | RunPod GPU pod                              |
+| Base      | nvidia/cuda:12.6.3-devel-ubuntu24.04        |
+| Model     | Gemma 4 E4B Q6_K                            |
+| Inference | llama.cpp prebuilt CUDA backends via dlopen |
+
 ---
 
 ## Fixture Strategy
@@ -260,8 +373,9 @@ is present.
 | `includes/`                  | Public headers and shared models.                  |
 | `src/`                       | Implementation files.                              |
 | `locations.json`             | Curated city input copied into the build tree.     |
-| `prompts/`                   | System prompt used by the model-backed path.       |
+| `prompts/`                   | System prompts used by the model-backed path.      |
 | `diagrams/`                  | Architecture and pipeline diagrams.                |
+| `tooling/pipeline/runpod/`   | Dockerfile, launcher, and RunPod pod template.     |
 | `ETHICS-AND-KNOWN-ISSUES.md` | Ethics, bias, hallucination analysis, mitigations. |
 
 ---
@@ -276,6 +390,7 @@ is present.
 - `src/data_generation/llama/` — local inference, prompt loading, output
   validation.
 - `src/data_generation/mock/` — deterministic fallback.
+- `tooling/pipeline/runpod/` — container build and runtime launcher.
 
 ---
 

docs/pipeline/diagrams/current/activity.puml

@@ -29,7 +29,7 @@ if (Are arguments valid?) then (no)
 else (yes)
 endif
 
-:Init CurlGlobalState & LlamaBackendState;
+:Init OpenSSL global state & LlamaBackendState;
 :di::make_injector(...);
 :injector.create<std::unique_ptr<BiergartenDataGenerator>>();
 :BiergartenDataGenerator::Run();

docs/pipeline/diagrams/current/class.puml

@@ -52,7 +52,7 @@ interface WebClient <<interface>> {
   + UrlEncode(value : const std::string&) : std::string
 }
 
-class CURLWebClient {
+class HttpWebClient {
   + Get(url : const std::string&) : std::string
   + UrlEncode(value : const std::string&) : std::string
 }
@@ -130,7 +130,7 @@ BiergartenDataGenerator *-- IExportService : owns
 IEnrichmentService <|.. WikipediaService : implements
 WikipediaService *-- WebClient : owns
 
-WebClient <|.. CURLWebClient : implements
+WebClient <|.. HttpWebClient : implements
 
 DataGenerator <|.. MockGenerator : implements
 DataGenerator <|.. LlamaGenerator : implements

docs/pipeline/diagrams/planned/activity.puml

@@ -13,7 +13,7 @@ if (Invalid args?) then (yes)
   stop
 else (no)
 endif
-:Init CurlGlobalState & LlamaBackendState;
+:Init OpenSSL global state & LlamaBackendState;
 :Build DI injector;
 
 :Initialize SqliteExportService;

docs/pipeline/diagrams/planned/class.puml

@@ -356,7 +356,7 @@ package "Infrastructure: Enrichment" {
     + UrlEncode(value : const std::string&) : std::string
   }
 
-  class CURLWebClient {
+  class HttpWebClient {
     + Get(url : const std::string&) : std::string
     + UrlEncode(value : const std::string&) : std::string
   }
@@ -520,7 +520,7 @@ CheckinDistributionStrategy    <|.. RandomCheckinStrategy
 FollowGenerationStrategy       <|.. RandomFollowStrategy
 FollowGenerationStrategy       <|.. ActivityWeightedFollowStrategy
 EnrichmentService              <|.. WikipediaService
-WebClient                      <|.. CURLWebClient
+WebClient                      <|.. HttpWebClient
 DataGenerator                  <|.. MockGenerator
 DataGenerator                  <|.. LlamaGenerator
 PromptFormatter                <|.. Gemma4JinjaPromptFormatter

tooling/pipeline/.dockerignore

@@ -0,0 +1,9 @@
+build/
+cmake-build-debug/
+.git/
+.idea/
+**/*.sqlite
+**/*.log
+**/*.sqlite3
+**/*.db
+

tooling/pipeline/CMakeLists.txt

@@ -1,41 +1,45 @@
 cmake_minimum_required(VERSION 3.31)
 project(biergarten-pipeline)
 
+# Set policy to allow FetchContent_Populate for header-only libraries
+# that have outdated CMakeLists.txt files
+cmake_policy(SET CMP0169 OLD)
+
 # 1. Build Options
 
 option(BIERGARTEN_MOCK_ONLY "Build with mock data generators only — skips llama.cpp" OFF)
-if (BIERGARTEN_MOCK_ONLY)
+if(BIERGARTEN_MOCK_ONLY)
-    message(STATUS "[biergarten] MOCK_ONLY build — llama.cpp will not be compiled.")
+        message(STATUS "[biergarten] MOCK_ONLY build — llama.cpp will not be compiled.")
-endif ()
+endif()
 
 # 2. Platform & GPU Detection
-if (NOT UNIX)
+if(NOT UNIX)
-    message(FATAL_ERROR "[biergarten] Windows is not supported. Please use Linux (Fedora 43) or macOS (M1 Pro).")
+        message(FATAL_ERROR "[biergarten] Windows is not supported. Please use Linux (Fedora 43) or macOS (M1 Pro).")
-endif ()
+endif()
 
-if (APPLE)
+if(APPLE)
-    if (CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
+        if(CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
-        message(STATUS "[biergarten] Apple Silicon detected — enabling Metal acceleration.")
+                message(STATUS "[biergarten] Apple Silicon detected — enabling Metal acceleration.")
-        set(GGML_METAL ON CACHE BOOL "Enable Metal for Apple Silicon" FORCE)
+                set(GGML_METAL ON CACHE BOOL "Enable Metal for Apple Silicon" FORCE)
-    else ()
+        else()
-        message(STATUS "[biergarten] Intel Mac detected — using CPU / Accelerate framework.")
+                message(STATUS "[biergarten] Intel Mac detected — using CPU / Accelerate framework.")
-        set(GGML_METAL OFF CACHE BOOL "Disable Metal for Intel Macs" FORCE)
+                set(GGML_METAL OFF CACHE BOOL "Disable Metal for Intel Macs" FORCE)
-    endif ()
+        endif()
-else ()
+else()
-    find_package(CUDAToolkit QUIET)
+        find_package(CUDAToolkit QUIET)
-    find_package(hip CONFIG QUIET)
+        find_package(hip CONFIG QUIET)
 
-    if (CUDAToolkit_FOUND)
+        if(CUDAToolkit_FOUND)
-        message(STATUS "[biergarten] NVIDIA GPU detected — enabling CUDA acceleration.")
+                message(STATUS "[biergarten] NVIDIA GPU detected — enabling CUDA acceleration.")
-        set(GGML_CUDA ON CACHE BOOL "Enable CUDA for NVIDIA GPUs" FORCE)
+                set(GGML_CUDA ON CACHE BOOL "Enable CUDA for NVIDIA GPUs" FORCE)
-        set(CMAKE_CUDA_ARCHITECTURES native)
+                set(CMAKE_CUDA_ARCHITECTURES native)
-    elseif (hip_FOUND OR DEFINED ENV{ROCM_PATH} OR EXISTS "/opt/rocm")
+        elseif(hip_FOUND OR DEFINED ENV{ROCM_PATH} OR EXISTS "/opt/rocm")
-        message(STATUS "[biergarten] AMD GPU detected — enabling HIP/ROCm acceleration.")
+                message(STATUS "[biergarten] AMD GPU detected — enabling HIP/ROCm acceleration.")
-        set(GGML_HIPBLAS ON CACHE BOOL "Enable HIP for AMD GPUs" FORCE)
+                set(GGML_HIPBLAS ON CACHE BOOL "Enable HIP for AMD GPUs" FORCE)
-    else ()
+        else()
-        message(STATUS "[biergarten] No NVIDIA or AMD GPU found — falling back to CPU.")
+                message(STATUS "[biergarten] No NVIDIA or AMD GPU found — falling back to CPU.")
-    endif ()
+        endif()
-endif ()
+endif()
 
 # 3. Project-wide Settings
 set(CMAKE_CXX_STANDARD 20)
@@ -51,16 +55,23 @@ include(FetchContent)
 find_package(Boost REQUIRED COMPONENTS json program_options)
 
 # Boost.DI (unofficial Boost extension, must declare separately from main Boost dependency)
+# Header-only library, so we only fetch without invoking its CMakeLists.txt
 FetchContent_Declare(
         boost-di
         GIT_REPOSITORY https://github.com/boost-ext/di.git
         GIT_TAG v1.3.0
+        GIT_SHALLOW TRUE
 )
-FetchContent_MakeAvailable(boost-di)
+FetchContent_GetProperties(boost-di)
-if (TARGET Boost.DI AND NOT TARGET boost::di)
+if(NOT boost-di_POPULATED)
-    add_library(boost::di ALIAS Boost.DI)
+        FetchContent_Populate(boost-di)
-endif ()
+endif()
 
+add_library(boost_di INTERFACE)
+add_library(boost::di ALIAS boost_di)
+target_include_directories(boost_di INTERFACE
+        $<BUILD_INTERFACE:${boost-di_SOURCE_DIR}/include>
+)
 # SQLite amalgamation
 FetchContent_Declare(
         sqlite_amalgamation
@@ -69,21 +80,38 @@ FetchContent_Declare(
         EXCLUDE_FROM_ALL
 )
 FetchContent_MakeAvailable(sqlite_amalgamation)
-if (NOT TARGET sqlite3)
+if(NOT TARGET sqlite3)
-    add_library(sqlite3 STATIC ${sqlite_amalgamation_SOURCE_DIR}/sqlite3.c)
+        add_library(sqlite3 STATIC ${sqlite_amalgamation_SOURCE_DIR}/sqlite3.c)
-    target_include_directories(sqlite3 PUBLIC ${sqlite_amalgamation_SOURCE_DIR})
+        target_include_directories(sqlite3 PUBLIC ${sqlite_amalgamation_SOURCE_DIR})
-    target_compile_definitions(sqlite3 PUBLIC SQLITE_THREADSAFE=1)
+        target_compile_definitions(sqlite3 PUBLIC SQLITE_THREADSAFE=1)
-endif ()
+endif()
 
 # llama.cpp — skipped for mock-only builds
-if (NOT BIERGARTEN_MOCK_ONLY)
+if(NOT BIERGARTEN_MOCK_ONLY)
-    FetchContent_Declare(
+        find_library(LLAMA_LIB NAMES llama)
-            llama-cpp
+        find_library(GGML_LIB NAMES ggml)
-            GIT_REPOSITORY https://github.com/ggml-org/llama.cpp.git
+        find_library(GGML_BASE_LIB NAMES ggml-base)
-            GIT_TAG b8742
+        find_path(LLAMA_INC_DIR NAMES llama.h PATH_SUFFIXES include)
-    )
+
-    FetchContent_MakeAvailable(llama-cpp)
+        if(LLAMA_LIB AND GGML_LIB AND GGML_BASE_LIB AND LLAMA_INC_DIR)
-endif ()
+                message(STATUS "[biergarten] Found system llama.cpp — skipping FetchContent")
+
+                add_library(llama SHARED IMPORTED)
+                set_target_properties(llama PROPERTIES
+                        IMPORTED_LOCATION "${LLAMA_LIB}"
+                        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INC_DIR}"
+                        INTERFACE_LINK_LIBRARIES "${GGML_LIB};${GGML_BASE_LIB}"
+                )
+        else()
+                message(STATUS "[biergarten] System llama.cpp not found — fetching via FetchContent")
+                FetchContent_Declare(
+                        llama-cpp
+                        GIT_REPOSITORY https://github.com/ggml-org/llama.cpp.git
+                        GIT_TAG b9012
+                )
+                FetchContent_MakeAvailable(llama-cpp)
+        endif()
+endif()
 
 # spdlog
 FetchContent_Declare(
@@ -153,16 +181,16 @@ target_sources(${PROJECT_NAME} PRIVATE
 )
 
 # --- data_generation: llama (skipped for mock-only builds) ---
-if (NOT BIERGARTEN_MOCK_ONLY)
+if(NOT BIERGARTEN_MOCK_ONLY)
-    target_sources(${PROJECT_NAME} PRIVATE
+        target_sources(${PROJECT_NAME} PRIVATE
-            src/data_generation/llama/load.cc
+                src/data_generation/llama/load.cc
-            src/data_generation/llama/helpers.cc
+                src/data_generation/llama/helpers.cc
-            src/data_generation/llama/generate_brewery.cc
+                src/data_generation/llama/generate_brewery.cc
-            src/data_generation/llama/infer.cc
+                src/data_generation/llama/infer.cc
-            src/data_generation/llama/llama_generator.cc
+                src/data_generation/llama/llama_generator.cc
-            src/data_generation/llama/generate_user.cc
+                src/data_generation/llama/generate_user.cc
-    )
+        )
-endif ()
+endif()
 
 # --- services: wikipedia ---
 target_sources(${PROJECT_NAME} PRIVATE
@@ -189,8 +217,6 @@ target_sources(${PROJECT_NAME} PRIVATE
 # 6. Include Directories, Link Libraries & Compile Definitions
 target_include_directories(${PROJECT_NAME} PRIVATE
         includes
-        $<$<NOT:$<BOOL:${BIERGARTEN_MOCK_ONLY}>>:${llama-cpp_SOURCE_DIR}/include>
-        $<$<NOT:$<BOOL:${BIERGARTEN_MOCK_ONLY}>>:${llama-cpp_SOURCE_DIR}/common>
 )
 
 target_link_libraries(${PROJECT_NAME} PRIVATE
@@ -225,4 +251,4 @@ add_custom_command(TARGET ${PROJECT_NAME} POST_BUILD
         COMMAND ${CMAKE_COMMAND} -E copy_directory
         ${CMAKE_SOURCE_DIR}/prompts
         ${CMAKE_BINARY_DIR}/prompts
-)
+)

tooling/pipeline/includes/data_generation/llama_generator.h

@@ -14,10 +14,10 @@
 #include <string>
 #include <string_view>
 
+#include "../services/prompting/prompt_directory.h"
 #include "data_generation/data_generator.h"
 #include "data_generation/prompt_formatting/prompt_formatter.h"
 #include "data_model/models.h"
-#include "../services/prompting/prompt_directory.h"
 
 struct llama_model;
 struct llama_context;
@@ -129,6 +129,7 @@ class LlamaGenerator final : public DataGenerator {
   uint32_t sampling_top_k_ = kDefaultSamplingTopK;
   std::mt19937 rng_;
   uint32_t n_ctx_ = kDefaultContextSize;
+  int n_gpu_layers_ = 0;
   std::unique_ptr<IPromptFormatter> prompt_formatter_;
   std::unique_ptr<IPromptDirectory> prompt_directory_;
 };

tooling/pipeline/includes/data_model/models.h

@@ -3,7 +3,8 @@
 
 /**
  * @file data_model/models.h
- * @brief Core data models: locations, application configuration, and generation inputs.
+ * @brief Core data models: locations, application configuration, and generation
+ * inputs.
  */
 
 #include <boost/program_options.hpp>
@@ -94,6 +95,9 @@ struct GeneratorOptions {
   /// @brief Use mocked generator instead of actual LLM inference.
   bool use_mocked = false;
 
+  /// @brief Number of layers to offload to GPU.
+  int n_gpu_layers = 0;
+
   /// @brief Specific sampling parameters for this generator.
   /// If nullopt, the application should use global defaults.
   std::optional<SamplingOptions> sampling;

tooling/pipeline/runpod/Dockerfile

@@ -0,0 +1,72 @@
+# --- Stage 1: Build Environment (The "Heavy" Stage) ---
+FROM nvidia/cuda:12.6.3-devel-ubuntu24.04 AS builder
+
+ENV DEBIAN_FRONTEND=noninteractive \
+  CMAKE_GENERATOR=Ninja
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+  build-essential ca-certificates curl git libboost-json-dev \
+  libboost-program-options-dev libssl-dev ninja-build pkg-config zlib1g-dev \
+  && rm -rf /var/lib/apt/lists/*
+
+# Install modern CMake
+RUN curl -L https://github.com/Kitware/CMake/releases/download/v3.31.0/cmake-3.31.0-linux-x86_64.sh -o cmake.sh && \
+  sh cmake.sh --skip-license --prefix=/usr/local && rm cmake.sh
+
+# Get headers for C++ build
+RUN curl -L https://github.com/ggml-org/llama.cpp/archive/refs/tags/b9012.tar.gz -o /tmp/llama-src.tar.gz && \
+  tar -xzf /tmp/llama-src.tar.gz -C /tmp && \
+  cp -r /tmp/llama.cpp-b9012/include/* /usr/local/include/ && \
+  cp -r /tmp/llama.cpp-b9012/ggml/include/* /usr/local/include/
+
+# Pull llama.cpp binaries to use during build if needed
+COPY --from=ghcr.io/ggml-org/llama.cpp:full-cuda /app/lib*.so* /usr/local/lib/
+
+WORKDIR /app
+COPY . .
+
+# Build the C++ pipeline
+RUN cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release && \
+  cmake --build build -j$(nproc)
+
+# --- Stage 2: Runtime Environment (The "Slim" Stage) ---
+FROM nvidia/cuda:12.6.3-runtime-ubuntu24.04 AS runtime
+
+# Install only necessary runtime shared libraries
+RUN apt-get update && apt-get install -y --no-install-recommends \
+  curl \
+  ca-certificates \
+  libboost-json1.83.0 \
+  libboost-program-options1.83.0 \
+  libgomp1 \
+  libssl3 \
+  zlib1g \
+  && rm -rf /var/lib/apt/lists/*
+
+ENV APP_ROOT=/app \
+  LD_LIBRARY_PATH="/usr/local/lib:${LD_LIBRARY_PATH}"
+
+WORKDIR /app/build
+
+# Copy only the compiled binaries from the builder
+COPY --from=builder /app/build/biergarten-pipeline ./
+
+# Copy required config files
+COPY locations.json /app/build/
+COPY beer-styles.json /app/build/
+
+# Copy prompt templates
+COPY prompts /app/prompts
+
+# Copy only the necessary shared libraries from builder/llama-bin
+COPY --from=ghcr.io/ggml-org/llama.cpp:full-cuda /app/lib*.so* /usr/local/lib/
+
+# Co-locate plugins
+RUN cp /usr/local/lib/libggml-cuda.so . 2>/dev/null || true && \
+  cp /usr/local/lib/libggml-cpu*.so . 2>/dev/null || true
+
+# Setup Start Script
+COPY ./runpod/start.sh /usr/local/bin/biergarten-start
+RUN chmod +x /usr/local/bin/biergarten-start
+
+ENTRYPOINT ["/usr/local/bin/biergarten-start"]

tooling/pipeline/runpod/README.md

@@ -0,0 +1,8 @@
+```bash
+touch runpod/start.sh
+docker build \
+ --progress=plain \
+ -t biergarten-pipeline:latest \
+ -f runpod/Dockerfile \
+ . 2>&1 | tee build.log
+```

tooling/pipeline/runpod/pod-template.yaml

@@ -0,0 +1,22 @@
+name: biergarten-pipeline-live
+imageName: biergarten-pipeline:latest
+category: NVIDIA
+containerDiskInGb: 50
+volumeInGb: 50
+volumeMountPath: /workspace
+dockerEntrypoint:
+  - /usr/local/bin/biergarten-start
+dockerStartCmd: []
+isPublic: false
+isServerless: false
+env:
+  BIERGARTEN_MODE: live
+  BIERGARTEN_MODEL_PATH: /workspace/models/google_gemma-4-E4B-it-Q6_K.gguf
+  BIERGARTEN_PROMPT_DIR: /workspace/app/build/prompts
+  BIERGARTEN_OUTPUT_DIR: /workspace/output
+  BIERGARTEN_LOG_PATH: /workspace/logs/pipeline.log
+  BIERGARTEN_TEMPERATURE: "1.0"
+  BIERGARTEN_TOP_P: "0.95"
+  BIERGARTEN_TOP_K: "64"
+  BIERGARTEN_N_CTX: "8192"
+  BIERGARTEN_SEED: "-1"

tooling/pipeline/runpod/start.sh

@@ -0,0 +1,58 @@
+#!/bin/bash
+set -e
+
+MODEL_PATH="${BIERGARTEN_MODEL_PATH:-/workspace/models/google_gemma-4-E4B-it-Q6_K.gguf}"
+OUTPUT_DIR="${BIERGARTEN_OUTPUT_DIR:-/workspace/output}"
+LOG_PATH="${BIERGARTEN_LOG_PATH:-/workspace/logs/pipeline.log}"
+EXECUTABLE="/app/build/biergarten-pipeline"
+PROMPT_DIR="/app/prompts"
+
+echo "--- Starting Biergarten Pipeline Environment Check ---"
+
+# Ensure directories exist
+mkdir -p "$OUTPUT_DIR"
+mkdir -p "$(dirname "$LOG_PATH")"
+mkdir -p "$(dirname "$MODEL_PATH")"
+
+# Download model if missing
+if [ ! -f "$MODEL_PATH" ]; then
+    echo "Model not found. Downloading (this may take a while)..."
+
+    curl -L -C - \
+      -o "$MODEL_PATH" \
+      "https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF/resolve/main/google_gemma-4-E4B-it-Q6_K.gguf?download=true"
+
+    echo "Download complete."
+fi
+
+# Verify model exists
+if [ ! -f "$MODEL_PATH" ]; then
+    echo "ERROR: Model still not found after download attempt."
+    exit 1
+fi
+
+# Default GPU layers
+GL_LAYERS="${BIERGARTEN_GL_LAYERS:-40}"
+
+# Build args
+ARGS=(
+    "--model"      "$MODEL_PATH"
+    "--prompt-dir" "$PROMPT_DIR"
+    "--output"     "$OUTPUT_DIR"
+    "--log-path"   "$LOG_PATH"
+    "--n-gpu-layers" "$GL_LAYERS"
+)
+
+# Optional params
+[[ -n "$BIERGARTEN_TEMPERATURE" ]] && ARGS+=("--temperature"  "$BIERGARTEN_TEMPERATURE")
+[[ -n "$BIERGARTEN_TOP_P" ]]       && ARGS+=("--top-p"        "$BIERGARTEN_TOP_P")
+[[ -n "$BIERGARTEN_TOP_K" ]]       && ARGS+=("--top-k"        "$BIERGARTEN_TOP_K")
+[[ -n "$BIERGARTEN_N_CTX" ]]       && ARGS+=("--n-ctx"        "$BIERGARTEN_N_CTX")
+[[ -n "$BIERGARTEN_SEED" ]]        && ARGS+=("--seed"         "$BIERGARTEN_SEED")
+
+# Extra args
+[[ -n "$BIERGARTEN_EXTRA_ARGS" ]] && ARGS+=($BIERGARTEN_EXTRA_ARGS)
+
+echo "--- Executing: $EXECUTABLE ${ARGS[*]} ---"
+
+exec "$EXECUTABLE" "${ARGS[@]}"

tooling/pipeline/src/application_options/parse_arguments.cc

@@ -50,6 +50,8 @@ std::optional<ApplicationOptions> ParseArguments(const int argc, char** argv) {
     opt("prompt-dir", prog_opts::value<std::string>()->default_value(""),
         "Directory containing named prompt files (e.g. BREWERY_GENERATION.md)."
         " Required when not using --mocked.");
+    opt("n-gpu-layers", prog_opts::value<int>()->default_value(0),
+        "Number of layers to offload to GPU");
   };
 
   add_sampling_options();
@@ -85,6 +87,7 @@ std::optional<ApplicationOptions> ParseArguments(const int argc, char** argv) {
 
     const bool use_mocked = var_map["mocked"].as<bool>();
     const std::string model_path = var_map["model"].as<std::string>();
+    const int n_gpu_layers = var_map["n-gpu-layers"].as<int>();
 
     // Enforce mutual exclusivity before any further configuration is applied.
     if (use_mocked && !model_path.empty()) {
@@ -110,6 +113,7 @@ std::optional<ApplicationOptions> ParseArguments(const int argc, char** argv) {
 
     options.generator.use_mocked = use_mocked;
     options.generator.model_path = model_path;
+    options.generator.n_gpu_layers = n_gpu_layers;
 
     // Only populate sampling config when the user explicitly overrides at
     // least one value. Leaving it as std::nullopt lets LlamaGenerator fall

tooling/pipeline/src/data_generation/llama/llama_generator.cc

@@ -89,6 +89,7 @@ LlamaGenerator::LlamaGenerator(
   }
 
   n_ctx_ = sampling.n_ctx;
+  n_gpu_layers_ = options.generator.n_gpu_layers;
 
   this->Load(model_path);
 }

tooling/pipeline/src/data_generation/llama/load.cc

@@ -12,6 +12,7 @@
 #include <utility>
 
 #include "data_generation/llama_generator.h"
+#include "ggml-backend.h"
 #include "llama.h"
 
 // Maximum batch size for decode operations. Capping the batch prevents
@@ -22,7 +23,12 @@ void LlamaGenerator::Load(const std::string& model_path) {
   context_.reset();
   model_.reset();
 
-  const llama_model_params model_params = llama_model_default_params();
+  // Specifically load dynamic ggml backends (like CUDA) that are provided
+  // externally before attempting to load a model.
+  ggml_backend_load_all();
+
+  llama_model_params model_params = llama_model_default_params();
+  model_params.n_gpu_layers = n_gpu_layers_;
   LlamaGenerator::ModelHandle loaded_model(
       llama_model_load_from_file(model_path.c_str(), model_params));
   if (!loaded_model) {