NPU: move the Qwen3.6-35B-A3B route to a private add-on, keep a generic hook

[bong-water-water-bong]

The Qwen3.6-35B-A3B NPU route is now closed source. Its code has moved to a private add-on, and the engine keeps only a small hook for it.

Removed

• The third_party/OpenFlowLM-Next submodule and its NOTICE entry
• scripts/build-lax.sh
• npu/lax*
• app/npu_lax* and the 1bit npu-lax command
• tests/npu_lax* and tests/golden/npu_lax/
• docs/npu-lax.md
• The ONEBIT_NPU_LAX* CMake options
• The --npu-kernels, --npu-transport and --npu-snapshots flags

The docs that mentioned the route now have a short note instead.

Added

• npu/private_route.{h,cpp}: a registry of private NPU routes keyed by model_type, plus optional 1bit subcommands.
• CMake option -DONEBIT_NPU_PRIVATE=<npu-kernels checkout> (needs -DONEBIT_NPU=ON). It runs add_subdirectory on the checkout's addons/, links onebit_npu_private and defines ONEBIT_NPU_PRIVATE. At startup, 1bit then calls register_private_addon().
• 1bit serve --npu-opt KEY=VALUE (in unified: --opt), which passes options through to a route.
• Without the add-on, 1bit serve -m <Qwen3.6-35B-A3B dir> --device npu exits with: the Qwen3.6-35B-A3B NPU route is not part of this build; build with -DONEBIT_NPU_PRIVATE=<npu-kernels checkout>.
• tests/npu_private_route_test.cpp (ctest npu_private_route), added to the CI target list.
• docs/npu.md, "Private routes".

The fast lane is unchanged.

Tested on Strix Halo (2026-09-25)

• Without the option (-DONEBIT_NPU=ON):
  • The CI target list builds and ctest passes 5/5.
  • serve on the 35B directory prints the message above and exits 1.
  • tests/serve_e2e.sh with Qwen3-0.6B on npu passes.
• With -DONEBIT_NPU_PRIVATE:
  • The add-on's tests pass: parity against the fp64 reference (argmax 846 / 198 / 3710, corr >= 0.99999), and a serve round trip that answers Paris with no xclbin opened.
  • serve_e2e.sh with Qwen3-0.6B on npu still passes.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit a544dae

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit a544dae)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Private Route Availability Check The npu_model_problem function checks for the existence of model files and lane kernels, but does not validate that the private route's unavailable function returns an empty string when the route is available. This could lead to incorrect error messages if a private route is registered but its unavailable function has a bug or returns a non-empty string even when the route should be available. std::string npu_model_problem(const std::string& dir, const npu::PrivateOptions& opts) { if (!has_model_files(dir)) return dir + " is not an NPU model directory (model.q4nx, config.json, tokenizer.json)"; const std::string type = npu::model_type_of(dir); if (const auto* r = npu::find_private_route(type)) { const std::string why = r->unavailable ? r->unavailable(dir, opts) : ""; return why.empty() ? "" : dir + ": " + why; } if (const std::string why = npu::private_route_missing(type); !why.empty()) return dir + ": " + why; if (!has_lane_kernels(dir)) return dir + " is not an NPU model directory (the fast lane needs npu/)"; return ""; } Option Parsing for Private Routes The --npu-opt option parsing in run_serve does not validate that the parsed options are used correctly by the private route. If a private route requires specific options that are not provided, it may fail silently or behave unexpectedly. The error handling for missing or invalid options should be more robust. npu::PrivateOptions opts; for (const auto& kv : o.npu_opts) npu::parse_private_option(kv, opts); if (const std::string why = npu_model_problem(o.model, opts); !why.empty()) throw std::runtime_error(why); // The NPU serves in process (unified.cpp). std::vector<std::string> args = {"-m", o.model, "-p", std::to_string(o.port), "--host", o.host}; if (!o.alias.empty()) { args.push_back("--alias"); args.push_back(o.alias); } for (const auto& kv : o.npu_opts) { args.push_back("--opt"); args.push_back(kv); } Private Route Registration The register_private_route function replaces an existing route with the same model_type but does not provide a mechanism to verify that the new route is compatible with the existing one. This could lead to unexpected behavior if a private route is registered with a different interface than the one previously registered. void register_private_route(PrivateRoute route) { for (auto& r : routes()) if (r.model_type == route.model_type) { r = std::move(route); return; } routes().push_back(std::move(route)); } Private Addon Validation The CMake logic for ONEBIT_NPU_PRIVATE checks for the existence of addons/CMakeLists.txt but does not validate that the onebit_npu_private target is correctly defined and linked. If the add-on's CMakeLists.txt is malformed or does not define the required target, the build may succeed but the private route functionality will not work as expected. set(ONEBIT_NPU_PRIVATE "" CACHE PATH "An npu-kernels checkout whose addons/ are built into 1bit (needs ONEBIT_NPU)") if(ONEBIT_NPU_PRIVATE) if(NOT ONEBIT_NPU) message(FATAL_ERROR "ONEBIT_NPU_PRIVATE needs ONEBIT_NPU=ON") endif() if(NOT EXISTS "${ONEBIT_NPU_PRIVATE}/addons/CMakeLists.txt") message(FATAL_ERROR "ONEBIT_NPU_PRIVATE=${ONEBIT_NPU_PRIVATE} has no addons/CMakeLists.txt") endif() add_subdirectory(${ONEBIT_NPU_PRIVATE}/addons ${CMAKE_BINARY_DIR}/npu_private) if(NOT TARGET onebit_npu_private) message(FATAL_ERROR "${ONEBIT_NPU_PRIVATE}/addons defines no onebit_npu_private target") endif() target_link_libraries(onebit PRIVATE onebit_npu_private) target_compile_definitions(onebit PRIVATE ONEBIT_NPU_PRIVATE) endif()

[github-actions]

Persistent review updated to latest commit a544dae