serve: drop the ONNX route and its packaging

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -42,7 +42,7 @@ body:
     id: device
     attributes:
       label: Device
-      options: [auto, npu, hrx, cpu, ds4, onnx, not sure]
+      options: [auto, npu, hrx, cpu, ds4, not sure]
     validations:
       required: true
   - type: input

.github/ISSUE_TEMPLATE/model_request.yml

@@ -41,7 +41,7 @@ body:
     attributes:
       label: Device you want it on
       multiple: true
-      options: [npu, hrx, cpu, ds4, onnx, any]
+      options: [npu, hrx, cpu, ds4, any]
     validations:
       required: true
   - type: textarea

.gitmodules

@@ -57,12 +57,6 @@
 	url = https://github.com/1bit-MONSTER/lemonade.git
 	branch = main
 	shallow = true
-[submodule "third_party/ryzenai-server"]
-	path = third_party/ryzenai-server
-	url = https://github.com/lemonade-sdk/ryzenai-server
-# DwarfStar (docs/dwarfstar.md): our fork 1bit-MONSTER/ds4, branch 1bit/main = upstream
-# antirez/ds4 main plus our commits (1BP packages, docs/1BP.md there); built by
-# scripts/build-ds4.sh; .github/workflows/bump-ds4.yml merges upstream into 1bit/main.
 [submodule "third_party/ds4"]
 	path = third_party/ds4
 	url = https://github.com/1bit-MONSTER/ds4.git

.pr_agent.toml

@@ -40,7 +40,7 @@ persistent_comment = true
 extra_instructions = """
 This is the 1bit engine: a C++26 inference engine behind an OpenAI-compatible API (`1bit serve`), running inside
 Lemonade, on AMD Strix Halo's XDNA 2 NPU (fast lane, the from-scratch dx kernels), the Radeon iGPU (HRX
-through a pinned llama.cpp build with HIP kernels; no Vulkan, ROCm backend, Metal, ZINC or CUDA), DwarfStar and ONNX. Review like an engineer who runs this on the hardware:
+through a pinned llama.cpp build with HIP kernels; no Vulkan, ROCm backend, Metal, ZINC or CUDA) and DwarfStar. Review like an engineer who runs this on the hardware:
 - correctness first: tensor shapes and strides, KV-cache positions and rollback, bf16/f32 rounding (dx kernels
   round to nearest even), NPU hardware-context and runlist lifetimes, lifetimes of child processes and ports;
 - C++: ownership and RAII, no allocations in per-token loops, exceptions reaching main, thread safety of shared state;

CMakeLists.txt

@@ -202,7 +202,7 @@ if(ONEBIT_HRX)
         foreach(_dev hrx cpu)
             add_test(NAME serve_e2e_${_dev}
                      COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} ${_dev})
-            # hrx skips without a gfx1151 (ONEBIT_HRX_E2E_TARGETS), as onnx_npu_smoke does without an NPU
+            # hrx skips without a gfx1151 (ONEBIT_HRX_E2E_TARGETS)
             set_tests_properties(serve_e2e_${_dev} PROPERTIES SKIP_RETURN_CODE 77)
         endforeach()
     endif()
@@ -227,25 +227,6 @@ if(ONEBIT_DS4)
     target_compile_definitions(onebit PRIVATE ONEBIT_DS4_SERVER="${_onebit_ds4_server}")
 endif()
 
-# ── ONNX Runtime GenAI (docs/onnx.md) ──────────────────────────────────────
-# Off by default: builds ryzenai-server (MIT) with scripts/build-onnx.sh as its own program;
-# `1bit serve --device onnx` runs it as a child process.
-option(ONEBIT_ONNX "Build ryzenai-server (third_party/ryzenai-server) for --device onnx" OFF)
-if(ONEBIT_ONNX)
-    set(ONEBIT_ONNX_SERVER "${CMAKE_BINARY_DIR}/onnx/ryzenai-server")
-    add_custom_target(onnx_server
-        COMMAND ${CMAKE_SOURCE_DIR}/scripts/build-onnx.sh ${CMAKE_BINARY_DIR}/onnx
-        BYPRODUCTS ${ONEBIT_ONNX_SERVER}
-        USES_TERMINAL)
-    add_dependencies(onebit onnx_server)
-    onebit_build_relative(ONEBIT_ONNX_SERVER _onebit_onnx_server)
-    target_compile_definitions(onebit PRIVATE ONEBIT_ONNX_SERVER="${_onebit_onnx_server}")
-    # Linux NPU through Ryzen AI Software (docs/onnx.md): skips without ONEBIT_RYZENAI_ROOT,
-    # an NPU and a model (ONEBIT_ONNX_NPU_MODEL)
-    add_test(NAME onnx_npu_smoke COMMAND ${CMAKE_SOURCE_DIR}/tests/onnx_npu_smoke.sh $<TARGET_FILE:onebit>)
-    set_tests_properties(onnx_npu_smoke PROPERTIES SKIP_RETURN_CODE 77)
-endif()
-
 # ── Hugging Face tokenizers (docs/tokenizers.md) ───────────────────────────
 # Off by default: it needs cargo. Any model's tokenizer.json, through our C ABI
 # over third_party/tokenizers (the same release as the tokenizers npm/PyPI

NOTICE

@@ -119,29 +119,6 @@ Lemonade (third_party/lemonade)
   built as its own program (scripts/build-lemonade.sh) that runs `1bit serve`
   as a backend. None of its code is linked into this product.
 
-ryzenai-server (third_party/ryzenai-server)
-  https://github.com/lemonade-sdk/ryzenai-server
-  Copyright (c) 2026 lemonade-sdk. License: MIT.
-  Built as a separate program (scripts/build-onnx.sh) that `1bit serve --device
-  onnx` runs as a child process; none of its code is linked into this product.
-  Its build fetches cpp-httplib (MIT), nlohmann/json (MIT) and the ONNX Runtime
-  GenAI headers (MIT). The release package carries the built program in onnx/,
-  with its licence as onnx/ryzenai-server-LICENSE.txt. The AMD_LICENSE file in
-  its repository covers AMD's Ryzen AI Software, which this product does not ship
-  or copy: the NPU and hybrid modes use the user's own install (RYZENAI_INSTALL_PATH
-  at build time, ONEBIT_RYZENAI_ROOT at run time; docs/onnx.md, "Linux NPU"), and
-  the release build never builds against it (scripts/weekly.sh checks the package).
-
-ONNX Runtime GenAI and ONNX Runtime
-  https://github.com/microsoft/onnxruntime-genai
-  https://github.com/microsoft/onnxruntime
-  Copyright (c) Microsoft Corporation. License: MIT.
-  scripts/build-onnx.sh fetches the release libraries (ONNX Runtime GenAI 0.11.2,
-  ONNX Runtime 1.23.2, pinned by sha256) and places them beside ryzenai-server.
-  The release package carries libonnxruntime-genai.so and libonnxruntime.so in
-  onnx/, with each project's LICENSE and ThirdPartyNotices.txt beside them
-  (onnx/onnxruntime-genai-*.txt, onnx/onnxruntime-*.txt).
-
 Hugging Face tokenizers (third_party/tokenizers)
   https://github.com/huggingface/tokenizers
   Copyright Hugging Face, Inc. and the tokenizers authors

README.md

@@ -59,9 +59,6 @@ hands to 1bit, it runs the engine as one of its backends, behind an OpenAI-compa
   same build on the CPU.
 - **The NPU engine** on the XDNA 2 NPU, in progress: correctness first, speed next
   ([docs/npu.md](docs/npu.md)).
-- **ONNX Runtime GenAI** (Lemonade's ONNX format) on the CPU, and on the NPU through the user's
-  own install of AMD's Ryzen AI Software, on Linux
-  ([#386](https://github.com/1bit-MONSTER/engine/pull/386), [docs/onnx.md](docs/onnx.md)).
 - **DwarfStar**, for DeepSeek V4 Flash, GLM 5.x and Qwen3.8-Flash-Next in its own GGUFs, with SSD
   expert streaming ([docs/dwarfstar.md](docs/dwarfstar.md)).
 - **Laya**, opt-in, which picks where each request runs ([docs/laya.md](docs/laya.md)), and

app/main.cpp

@@ -15,7 +15,7 @@
 // 1bit: the engine binary. Subcommands:
 //
 //   1bit serve -m <model>            one model on any device (NPU, HRX, CPU,
-//                                    DwarfStar, ONNX) behind an OpenAI-compatible API;
+//                                    DwarfStar) behind an OpenAI-compatible API;
 //                                    how the engine runs inside Lemonade (docs/serve.md).
 //   1bit unified -m <model dir>      one native model on the NPU fast lane behind
 //                                    OpenAI endpoints; serve's NPU route.