Embed the engine into Lemonade, not Lemonade into the engine: drop the vendored Lemonade

.github/workflows/bump-hrx.yml

@@ -18,7 +18,7 @@
 # (docs/hrx.md, "Staying current"). When AMD moves its pins, this workflow:
 #   1. syncs 1bit-MONSTER/llama.cpp: master to ggml-org, 1bit/hrx-vulkan to AMD's pin;
 #   2. opens a PR here moving both submodules.
-# GitHub-hosted CI builds that PR without HRX; run tests/hrx_lemonade_e2e.sh on
+# GitHub-hosted CI builds that PR without HRX; run tests/serve_e2e.sh on
 # Strix Halo before merging it.
 #
 # Needs the secret HRX_BUMP_TOKEN: a fine-grained token with Contents
@@ -127,4 +127,4 @@ jobs:
           | third_party/hrx-system (ROCm/hrx-system) | \`${OURS_HRX:0:12}\` | \`${AMD_HRX:0:12}\` |
 
           CI here builds without HRX. Before merging, on Strix Halo:
-          \`cmake -B build -G Ninja -DONEBIT_HRX=ON && cmake --build build --target onebit && tests/hrx_lemonade_e2e.sh build/1bit build/hrx/llama/bin/llama-server\`"
+          \`cmake -B build -G Ninja -DONEBIT_HRX=ON && cmake --build build --target onebit && for d in vulkan hrx; do tests/serve_e2e.sh build/1bit <Qwen3-0.6B Q4_K_M .gguf> \$d; done\`"

CMakeLists.txt

@@ -28,24 +28,40 @@ enable_testing()
 
 # Components are ported in the order listed in docs/PORTING.md.
 
-# ── Step 1: Lemonade's server core, embedded ───────────────────────────────
-# Vendored at third_party/lemonade (v11.9.0 plus the local deltas listed in
-# its UPSTREAM.md, including the `onebit` backend).
-set(BUILD_WEB_APP OFF CACHE BOOL "" FORCE)
-set(BUILD_TESTING OFF CACHE BOOL "" FORCE)
-add_subdirectory(third_party/lemonade EXCLUDE_FROM_ALL)
+# ── The engine binary: `1bit serve`, an OpenAI-compatible API (docs/serve.md)
+# The engine is a backend a host such as Lemonade launches; it no longer embeds
+# one. JSON and the HTTP server/client are pinned here, at the commits the
+# engine used before (nlohmann/json v3.11.3, cpp-httplib v0.47.0).
+include(FetchContent)
+find_package(Threads REQUIRED)
+set(JSON_BuildTests OFF CACHE INTERNAL "")
+FetchContent_Declare(json
+    GIT_REPOSITORY https://github.com/nlohmann/json.git
+    GIT_TAG 9cca280a4d0ccf0c08f47a99aa71d1b0e52f8d03)  # v3.11.3
+set(HTTPLIB_USE_OPENSSL_IF_AVAILABLE OFF CACHE BOOL "" FORCE)
+set(HTTPLIB_USE_ZLIB_IF_AVAILABLE OFF CACHE BOOL "" FORCE)
+set(HTTPLIB_USE_BROTLI_IF_AVAILABLE OFF CACHE BOOL "" FORCE)
+set(HTTPLIB_USE_ZSTD_IF_AVAILABLE OFF CACHE BOOL "" FORCE)
+set(HTTPLIB_INSTALL OFF CACHE BOOL "" FORCE)
+FetchContent_Declare(httplib
+    GIT_REPOSITORY https://github.com/yhirose/cpp-httplib.git
+    GIT_TAG fe332fa06bac76a1c6d402c08f414052999347da)  # v0.47.0
+FetchContent_MakeAvailable(json httplib)
 
 add_executable(onebit app/main.cpp app/serve.cpp)
 set_target_properties(onebit PROPERTIES OUTPUT_NAME 1bit)
-target_link_libraries(onebit PRIVATE lemonade-server-core)
+target_link_libraries(onebit PRIVATE nlohmann_json::nlohmann_json httplib::httplib Threads::Threads)
+
+# Runs everywhere, CI included: serve's proxy against a fake backend.
+add_test(NAME smoke_serve COMMAND ${CMAKE_SOURCE_DIR}/tests/smoke_serve.sh $<TARGET_FILE:onebit>)
 
 # ── `1bit serve` end to end (docs/serve.md) ───────────────────────────────
 # Needs a GGUF (e.g. Qwen3-0.6B Q4_K_M); tests each device this build has.
 set(ONEBIT_SERVE_TEST_GGUF "" CACHE FILEPATH "GGUF the serve_e2e tests load")
 
 # ── Step 2: HRX + Vulkan in one llama.cpp build (docs/hrx.md) ─────────────
 # Off by default: it needs TheRock/ROCm and the three HRX submodules, and it
-# only runs on AMD GPUs. `1bit lemonade` then uses this llama-server for both
+# only runs on AMD GPUs. `1bit serve` then uses this llama-server for both
 # the llamacpp-hrx recipe (HRX0) and the llamacpp recipe's Vulkan backend.
 option(ONEBIT_HRX "Build HRX + Vulkan llama-server and wire it into Lemonade" OFF)
 if(ONEBIT_HRX)
@@ -58,13 +74,11 @@ if(ONEBIT_HRX)
                      COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} ${_dev})
         endforeach()
     endif()
-    add_test(NAME hrx_lemonade_e2e
-             COMMAND ${CMAKE_SOURCE_DIR}/tests/hrx_lemonade_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_HRX_SERVER})
 endif()
 
 # ── ZINC (docs/zinc.md) ────────────────────────────────────────────────────
 # Off by default: it builds third_party/zinc with scripts/build-zinc.sh (which
-# fetches its own Zig) for one GPU backend. `1bit lemonade` then serves the zinc
+# fetches its own Zig) for one GPU backend. `1bit serve --device zinc` then runs the zinc
 # recipe with this binary. ONEBIT_ZINC_BACKEND=cuda is the NVIDIA build.
 option(ONEBIT_ZINC "Build ZINC (third_party/zinc) and wire it into Lemonade" OFF)
 if(ONEBIT_ZINC)
@@ -84,8 +98,6 @@ if(ONEBIT_ZINC)
         add_test(NAME serve_e2e_zinc
                  COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} zinc)
     endif()
-    add_test(NAME zinc_lemonade_e2e
-             COMMAND ${CMAKE_SOURCE_DIR}/tests/zinc_lemonade_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_ZINC_SERVER})
 endif()
 
 # ── Hugging Face tokenizers (docs/tokenizers.md) ───────────────────────────
@@ -183,11 +195,11 @@ if(ONEBIT_NPU)
     endif()
 endif()
 
-add_test(NAME smoke_lemonade COMMAND ${CMAKE_SOURCE_DIR}/tests/smoke_lemonade.sh $<TARGET_FILE:onebit>)
 
-# MLX on Apple Silicon through Lemonade's mlx backend (docs/apple.md).
-set(ONEBIT_MLX_SERVER "" CACHE FILEPATH "lemon-mlx-engine's server binary, for tests/mlx_lemonade_e2e.sh")
+# MLX on Apple Silicon: `1bit serve --device mlx` (docs/apple.md).
+set(ONEBIT_MLX_SERVER "" CACHE FILEPATH "lemon-mlx-engine's server binary, for serve_e2e_mlx")
 if(APPLE AND ONEBIT_MLX_SERVER)
-    add_test(NAME mlx_lemonade_e2e
-             COMMAND ${CMAKE_SOURCE_DIR}/tests/mlx_lemonade_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_MLX_SERVER})
+    add_test(NAME serve_e2e_mlx
+             COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> mlx-community/Qwen3-0.6B-4bit mlx
+                     --mlx-server ${ONEBIT_MLX_SERVER})
 endif()

NOTICE

@@ -13,19 +13,6 @@ third_party/ (or in its upstream repository for components that are fetched at
 build time rather than stored here). The copyright lines below are taken from
 each project's own LICENSE file or source headers.
 
--------------------------------------------------------------------------------
-Included in this repository, with modifications
--------------------------------------------------------------------------------
-
-Lemonade (third_party/lemonade)
-  https://github.com/lemonade-sdk/lemonade
-  Copyright (C) 2022-2024, Advanced Micro Devices, Inc.
-  License: Apache-2.0
-  Modified: the local changes (the onebit, mlx and zinc backends, the
-  hrx_device option, the embeddability patch and model-registry entries) are
-  listed in third_party/lemonade/UPSTREAM.md, as Apache-2.0 section 4(b)
-  requires. Lemonade bundles AixLog, Copyright (C) 2017-2021 Johannes Pohl (MIT).
-
 -------------------------------------------------------------------------------
 Pinned at exact commits (git submodules) and built from source
 -------------------------------------------------------------------------------
@@ -86,8 +73,7 @@ convaiinnovations/laya, pinned in config/laya.json)
   License: Apache-2.0
 
 -------------------------------------------------------------------------------
-Fetched by the build at pinned commits (third_party/lemonade/CMakeLists.txt),
-unless a suitable system copy is found
+Fetched by the build at pinned commits (CMakeLists.txt)
 -------------------------------------------------------------------------------
 
 JSON for Modern C++ (nlohmann/json v3.11.3)
@@ -99,34 +85,3 @@ cpp-httplib (v0.47.0)
   https://github.com/yhirose/cpp-httplib
   Copyright (c) 2017 yhirose
   License: MIT
-
-CLI11 (v2.4.2)
-  https://github.com/CLIUtils/CLI11
-  Copyright (c) 2017-2026 University of Cincinnati, developed by Henry Schreiner
-  under NSF AWARD 1414736
-  License: BSD-3-Clause
-
-curl (8.5.0)
-  https://github.com/curl/curl
-  Copyright (c) 1996-2026, Daniel Stenberg, <daniel@haxx.se>, and many contributors
-  License: curl
-
-Zstandard (zstd v1.5.7)
-  https://github.com/facebook/zstd
-  Copyright (c) Meta Platforms, Inc. and affiliates.
-  License: BSD-3-Clause OR GPL-2.0-only (used here under BSD-3-Clause)
-
-Mbed TLS (v3.6.2)
-  https://github.com/Mbed-TLS/mbedtls
-  Copyright The Mbed TLS Contributors
-  License: Apache-2.0 OR GPL-2.0-or-later (used here under Apache-2.0)
-
-libwebsockets (v4.3.3)
-  https://github.com/warmcat/libwebsockets
-  Copyright (C) 2010-2020 Andy Green <andy@warmcat.com> and contributors
-  License: MIT (some sources under similar permissive licenses; see its LICENSE)
-
-Brotli (v1.1.0, macOS builds only)
-  https://github.com/google/brotli
-  Copyright (c) 2009, 2010, 2013-2016 by the Brotli Authors
-  License: MIT

README.md

@@ -16,22 +16,27 @@ limitations under the License.
 -->
 # 1bit engine
 
-One binary that runs [Lemonade](https://github.com/lemonade-sdk/lemonade) completely,
-with AMD Ryzen AI hardware behind it:
+An inference engine for AMD Ryzen AI, built to run inside
+[Lemonade](https://github.com/lemonade-sdk/lemonade): the `1bit` binary serves one model behind an
+OpenAI-compatible API, whatever device runs it, and Lemonade (or any OpenAI client) launches it
+the way it launches `llama-server`. Behind that one API:
 
 - the XDNA 2 NPU engine
 - HRX and Vulkan on the Radeon iGPU, compiled together in one llama.cpp build
 - ZINC, which also reaches NVIDIA GPUs (CUDA) and Apple GPUs (Metal)
+- MLX on Apple Silicon, through lemon-mlx-engine
 - Laya, which decides where each request runs
 - every Hugging Face model architecture, kept current by a daily census
 
-> **Status:** steps 1–3 of 5 have landed: embedded Lemonade ([docs/lemonade.md](docs/lemonade.md)),
-> HRX + Vulkan in one build on AMD's live ggml-hrx ([docs/hrx.md](docs/hrx.md)), and the NPU engine on
-> full ELFs with the upstream XDNA stack pinned ([docs/npu.md](docs/npu.md); its layer kernel is
-> not yet built from source). Also: MLX on Apple Silicon through Lemonade's `mlx` recipe
-> ([docs/apple.md](docs/apple.md)), and ZINC, which reaches NVIDIA GPUs through its CUDA backend
-> ([docs/zinc.md](docs/zinc.md)). Next is step 4, the Laya router. The working engine is being ported from
-> [1bit-MONSTER](https://github.com/1bit-MONSTER/1bit-MONSTER) in five steps; see
+> **Status:** `1bit serve` is the engine's front door ([docs/serve.md](docs/serve.md)): the NPU,
+> Vulkan, HRX and ZINC each pass its end-to-end test on Strix Halo. Following geramyL's review, the
+> engine is embedded into Lemonade rather than embedding it: the vendored Lemonade is gone, and a
+> Lemonade recipe that launches `1bit serve` is being prepared for upstream ([docs/lemonade.md](docs/lemonade.md)).
+> Ported so far: HRX + Vulkan in one build on AMD's live ggml-hrx ([docs/hrx.md](docs/hrx.md)), the NPU
+> engine on full ELFs with the upstream XDNA stack pinned ([docs/npu.md](docs/npu.md); its layer kernel
+> is not yet built from source), ZINC ([docs/zinc.md](docs/zinc.md)) and MLX
+> ([docs/apple.md](docs/apple.md)). Next is step 4, the Laya router. The working engine is being ported
+> from [1bit-MONSTER](https://github.com/1bit-MONSTER/1bit-MONSTER); see
 > [docs/PORTING.md](docs/PORTING.md). Measured results are on the
 > [wiki](https://github.com/1bit-MONSTER/engine/wiki). This repository holds the verified code
 > without the development history.
@@ -46,7 +51,7 @@ Apache-2.0. See [LICENSE](LICENSE).
 vibecoding, and that is where all of this started. Without it, this engine would not exist.
 
 **The Lemonade team and AMD's developers.** [Lemonade](https://github.com/lemonade-sdk/lemonade)
-runs inside this binary. AMD's developers left breadcrumbs all over the place: the XDNA driver
+is the home this engine is built to run in. AMD's developers left breadcrumbs all over the place: the XDNA driver
 and XRT, IRON and Peano, HRX, their tested llama.cpp integration, their issues, their examples.
 This engine is what following those breadcrumbs built.
 
@@ -65,9 +70,8 @@ The repositories this engine is built on, in order of importance:
 | 9 | [huggingface/tokenizers](https://github.com/huggingface/tokenizers) | Every model's `tokenizer.json`, byte-exact, behind our C ABI | Apache-2.0 |
 | 10 | [NandhaKishorM/laya](https://github.com/NandhaKishorM/laya) | The router that decides where each request runs | Apache-2.0 |
 
-Also built on [nlohmann/json](https://github.com/nlohmann/json) (MIT),
-[yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) (MIT), and the libraries Lemonade
-builds with: curl, zstd, Mbed TLS, libwebsockets, CLI11 and Brotli.
+Also built on [nlohmann/json](https://github.com/nlohmann/json) (MIT)
+and [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) (MIT).
 
 Every third-party copyright and license is listed in [NOTICE](NOTICE). Each project keeps its
 own license; nothing here relicenses anyone's work.