[jakmro]

No description provided.

[Copilot]

Pull request overview

This PR removes Apple CoreML/NPU support across the Python transpiler, C++ engine, and associated docs/tests, while standardizing bundle naming/selection to CQ-bit-only variants and shifting acceleration messaging toward Metal GPU. It also introduces per-op backend pinning (CPU vs Metal) in the graph/runtime surface, including a new FFI to tag node backends.

Changes:

• Remove CoreML/NPU emission pipeline (Python) and ANE/NPU runtime paths (C++), plus related flags/docs.
• Simplify bundle variant logic by dropping --weights / platform suffixes and parsing only -cq<bits> artifacts.
• Update graph/runtime backend handling to be per-node pinnable with CPU/Metal and refresh parity tests/documentation accordingly.

Reviewed changes

Copilot reviewed 72 out of 73 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
README.md	Updates CLI/help text and benchmark wording after removing weights variants.
python/tests/test_server_live.py	Updates bundle stem comment to remove -apple variant mention.
python/tests/test_model.py	Removes platform/weights variant selection from tests.
python/tests/test_download.py	Updates archive parsing/resolution tests for CQ-bit-only artifacts.
python/tests/test_cli_run.py	Removes weights arg from CLI run test fixture.
python/tests/test_bindings.py	Removes --weights flag parsing tests.
python/tests/bundles.py	Updates bundle candidate docstring to remove -apple suffix mention.
python/README.md	Removes weights variant examples; documents per-op backend kwarg usage.
python/pyproject.toml	Drops coremltools dependency on Darwin.
python/cactus/transpile/tdt_runtime.py	Removes ANE-specific encoder code paths and env knobs.
python/cactus/transpile/runtime_compat.py	Switches to per-node backend tagging via _apply_backend + new FFI signature changes.
python/cactus/transpile/README.md	Removes mention of .mlpackage runtime dispatch.
python/cactus/transpile/npu/vision.py	Deletes CoreML vision encoder emission helper.
python/cactus/transpile/npu/source.py	Deletes CoreML source encoder emission helper.
python/cactus/transpile/npu/README.md	Deletes NPU transpiler pipeline documentation.
python/cactus/transpile/npu/pipeline.py	Deletes NPU/CoreML emission orchestrator.
python/cactus/transpile/npu/coremltools_patches.py	Deletes CoreMLTools patch layer.
python/cactus/transpile/npu/audio.py	Deletes CoreML audio encoder emission helper.
python/cactus/transpile/npu/init.py	Removes NPU package export.
python/cactus/transpile/model_adapters.py	Removes Gemma4 NPU reparam + NPU module fields in specs.
python/cactus/transpile/hf_model.py	Removes --npu* CLI flags and manifest injection of .mlpackage paths.
python/cactus/transpile/component_pipeline.py	Removes NPU-related fields from ComponentModuleSpec.
python/cactus/server.py	Updates server bundle-stem comment to remove -apple example.
python/cactus/cli/utils.py	Simplifies CQ variant parsing/selection; removes platform dimension.
python/cactus/cli/upload.py	Updates upload naming to use CQ bits only (no platform suffix).
python/cactus/cli/test.py	Only forwards --backend when explicitly set (default remains auto).
python/cactus/cli/serve.py	Only forces backend when explicitly set (default remains auto).
python/cactus/cli/runtime.py	Removes CoreML framework from Python runtime link flags.
python/cactus/cli/model.py	Removes platform/weights handling and NPU-related transpile options.
python/cactus/cli/list.py	Changes listing output format to a unified “Available models” table.
python/cactus/cli/download.py	Removes weights variant resolution and makes bundle dir CQ-bit-only.
python/cactus/cli/convert.py	Removes weights variant handling and NPU flag forwarding.
python/cactus/cli/compile.py	Removes CoreML framework from build link flags.
python/cactus/cli/common.py	Removes supported weights variants constant.
python/cactus/cli/code.py	Removes --weights handling; only forwards backend when set.
python/cactus/cli/init.py	Removes --weights/NPU CLI flags; updates backend options; adds upload command wiring.
python/cactus/init.py	Removes some exports (e.g., language detection / decode tokens).
llms.txt	Updates acceleration claim from NPU to Metal.
docs/choose-bindings.md	Updates Swift binding description from NPU to Metal.
docs/cactus_transpiler.md	Updates bundle path conventions and removes NPU emission section.
docs/cactus_graph.md	Updates backend selection guidance to CPU/Metal with per-op pinning.
docs/cactus_engine.md	Removes weights variant path and --weights apple mention.
cactus-graph/tests/test_metal_parity.cpp	Refactors parity tests for per-op backend pinning and removes “auto” resets.
cactus-graph/src/param_io.cpp	Treats backend as runtime-only (no longer serialized).
cactus-graph/src/ops_tensor.cpp	Removes NPU-not-implemented guard now that NPU backend is removed.
cactus-graph/src/ops_nn.cpp	Gates Metal kernels by node backend pin + availability, removes NPU stubs.
cactus-graph/src/ops_conv.cpp	Removes NPU-not-implemented guards now that NPU backend is removed.
cactus-graph/src/ops_cache.cpp	Updates residency logic to use cactus_default_backend()==METAL.
cactus-graph/src/metal_plan.cpp	Prevents Metal fusion planning over CPU-pinned nodes.
cactus-graph/src/io.cpp	Adds backend tagging to embedding node creation.
cactus-graph/src/graph_ffi.cpp	Removes per-op backend args from many FFI calls; adds cactus_graph_set_node_backend.
cactus-graph/src/execute.cpp	Reworks backend selection and adds cactus_default_backend() used as the default for ops.
cactus-graph/cactus_graph.h	Removes NPU backend, adds cactus_default_backend(), expands backend params across graph ops.
cactus-engine/tests/test_utils.cpp	Adjusts backend env handling logic for tests.
cactus-engine/tests/run.cpp	Updates backend CLI validation/error message (no “auto”).
cactus-engine/tests/ios/configure_xcode.rb	Removes CoreML framework from iOS test app link flags.
cactus-engine/test.sh	Changes default backend from auto to empty/unset.
cactus-engine/src/npu.cpp	Deletes NPU stub implementation.
cactus-engine/src/npu_ane.mm	Deletes ANE/CoreML implementation.
cactus-engine/src/npu_ane.h	Deletes ANE/CoreML header.
cactus-engine/src/model.cpp	Removes NPU encoder load/execute paths and updates Metal/backend checks.
cactus-engine/src/model_npu.cpp	Deletes NPU model integration code.
cactus-engine/src/engine.h	Removes NPU encoder interfaces and related model fields.
cactus-engine/CMakeLists.txt	Removes NPU sources and CoreML framework linking.
cactus-engine/cactus_engine.h	Adds new audio preprocess FFI declaration.
cactus-code/package-lock.json	Removes some optional deps entries (e.g., ws, zod) from lockfile.
blog/lfm2.5_350m.md	Updates narrative away from “NPU” to “dedicated accelerators” and generalizes wording.
blog/hybrid_transcription.md	Removes “NPUs” claim in engine description.
blog/gemma4.md	Updates performance narrative and Swift binding description from NPU to Metal.
bindings/rust/cactus.rs	Swaps CoreML framework link for Metal + MPS on macOS.
apple/CMakeLists.txt	Removes CoreML framework from Apple framework link flags.

Files not reviewed (1)

• cactus-code/package-lock.json: Generated file

Comments suppressed due to low confidence (1)

cactus-engine/tests/test_utils.cpp:17

• TestUtils::apply_backend() will now treat CACTUS_TEST_BACKEND=auto as an invalid backend and print "unavailable; using default". Several test runners default this env var to auto (e.g. cactus-engine/tests/android/run.sh and cactus-engine/tests/ios/run.sh), so this change will add noisy output and makes it harder to tell when a backend is truly misconfigured.

Either ignore "auto" here (treat it the same as unset), or teach the backend selector to accept "auto" explicitly and print a clearer message.

    const char* b = std::getenv("CACTUS_TEST_BACKEND");
    if (!b || !*b) return;
    if (cactus_set_backend(b) == 0)
        std::cout << "Backend: " << (std::strcmp(b, "metal") == 0 ? "Metal GPU" : "CPU") << "\n";
    else
        std::cout << "Backend '" << b << "' unavailable; using default\n";

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.