coreml-cli

March 22, 2026 · View on GitHub

A command-line tool to profile CoreML models — showing per-operation compute device assignments (CPU/GPU/ANE), compilation time, and prediction latency across all MLComputeUnits configurations.

Replicates what Xcode's CoreML Performance Report does, but from the terminal and designed for programmatic use by coding agents.

Example

$ coreml-cli test_models/160ms/

Device:    Apple M4 Pro (arm64)
RAM:       48GB
OS:        macOS 26.3.1

── decoder ────────────────────────────────────────────────────────────────────
  Parakeet EOU decoder (RNNT prediction network) (Fluid Inference)
  Mixed (Float16, Float32, Int16, Int32) | torch==2.4.0 | coremltools 8.3.0
  inputs:  targets(Int32 1×1), target_length(Int32 1), h_in(Float32 1×1×640),
           c_in(Float32 1×1×640)
  outputs: decoder(Float32 1×640×1), h_out(Float32 1×1×640),
           c_out(Float32 1×1×640)
  cold compile: 128ms

  Compute Unit                 CPU    GPU    ANE   Compile   Predict
  ────────────────────────────────────────────────────────────────
  all                       100.0%   0.0%   0.0%       7ms    0.22ms
  cpu_only                  100.0%   0.0%   0.0%       6ms    0.22ms
  cpu_and_gpu               100.0%   0.0%   0.0%       6ms    0.23ms
  cpu_and_neural_engine     100.0%   0.0%   0.0%       5ms    0.26ms

── streaming_encoder ──────────────────────────────────────────────────────────
  Mixed (Float16, Float32, Int32) | torch==2.4.0 | coremltools 8.3.0
  ...
  cold compile: 3512ms

  Compute Unit                 CPU    GPU    ANE   Compile   Predict
  ────────────────────────────────────────────────────────────────
  all                         0.0% 100.0%   0.0%      46ms    6.78ms
  cpu_only                  100.0%   0.0%   0.0%      47ms    5.45ms
  cpu_and_gpu                 0.0% 100.0%   0.0%      49ms    6.67ms
  cpu_and_neural_engine       1.2%   0.0%  98.8%      51ms    2.82ms

Install

Requires macOS 14+ and uv.

git clone https://github.com/yourusername/coreml-cli
cd coreml-cli
uv sync

Usage

# Profile a single model (all compute unit configs)
uv run coreml-cli model.mlmodelc

# Profile all models in a directory
uv run coreml-cli path/to/models/

# Specific compute unit config
uv run coreml-cli model.mlmodelc --units cpu_and_neural_engine

# JSON output (for programmatic use)
uv run coreml-cli model.mlmodelc --json

# Include per-operation breakdown
uv run coreml-cli model.mlmodelc --ops

# Per-op data with private API details (backend support, estimated runtimes)
uv run coreml-cli model.mlmodelc --detailed

# ANE fallback analysis — show CPU ops grouped by rejection reason
uv run coreml-cli model.mlmodelc --fallback

# Fallback analysis as JSON (for agent consumption)
uv run coreml-cli model.mlmodelc --fallback --json

# Control benchmark iterations
uv run coreml-cli model.mlmodelc --iterations 50

# Debug logging to stderr
uv run coreml-cli model.mlmodelc --debug

What it reports

Benchmark mode (default)

For each model:

  • Cold compile time — measured once per model by bypassing the E5 compilation cache (private API: setExperimentalMLProgramEncryptedCacheUsage_(0)). Reflects what users experience the first time the model runs on their device — if this is too high, the model may not be usable. For a true first-launch measurement, restart ANECompilerService before benchmarking: sudo killall ANECompilerService.

For each compute unit configuration (all, cpu_only, cpu_and_gpu, cpu_and_neural_engine):

  • Device assignment — % of operations on CPU, GPU, and ANE (Neural Engine)
  • Compile time — cached load time (E5 bundle cache populated). This is the cost paid on every app launch.
  • Predict latency — median prediction time (5 warmup + 10 timed iterations)
  • Model metadata — precision, I/O shapes, author, description, coremltools version
  • Per-op breakdown (--ops) — each operation's name, type, assigned device, and cost weight
  • Private API data (--detailed) — selected backend, all supported backends, estimated runtime per backend, validation messages explaining why backends were rejected

Fallback analysis mode (--fallback)

Shows only ops that are not on ANE, grouped by rejection reason. Designed for the ANE optimization loop: change conversion → reconvert → --fallback → identify blockers → fix → repeat.

For each CPU-fallback op, reports:

  • Why ANE rejected it — e.g., "Unsupported tensor data type: int32", "Unsupported MIL operation"
  • How many ops — grouped by rejection reason with op type counts
  • Estimated CPU cost — how much latency the fallback adds
  • Which ops — names for tracing back to the conversion script

Common ANE rejection reasons and fixes:

  • Unsupported tensor data type: int32 — cast to float16 before these operations
  • Unsupported MIL operation "lstm" — decompose into supported ops (matmul, sigmoid, tanh)
  • Unsupported MIL operation "logical_and" — replace with float multiply workaround
  • Unable to resolve operation input — cascading from another CPU op; fix the upstream op first
  • ANE supported but scheduler chose CPU — data transfer overhead; often not worth fixing

How it works

Uses PyObjC to call macOS CoreML framework APIs directly from Python:

  1. Public API — MLComputePlan (macOS 14+) for per-operation device assignment and cost weights
  2. Private API — MLE5Engine.segmentationAnalyticsAndReturnError: for richer data including backend support matrices and estimated runtimes per backend

Heavily inspired by:

  • maderix/ANE — reverse-engineered private _ANEClient/_ANECompiler APIs for direct Neural Engine access. Their runtime introspection approach (objc_msgSend, NSClassFromString) informed how we navigate CoreML's internal object graph.
  • freedomtan/coreml_modelc_profling — per-operation profiling using both public MLComputePlan and undocumented MLE5Engine APIs. Their Objective-C implementation was the direct reference for our private profiler.

Caveats

Note that this was a weekend project, built with Claude Code.

  • Hardware-specific — compute plans and compilation are tied to the local chip. Results on an M4 Pro will differ from an M1 or A17 Pro.
  • Private APIs may break — the MLE5Engine path (--detailed) uses undocumented APIs that may change across macOS versions.
  • macOS 26 tested — CoreML enum values changed in macOS 26 (Tahoe). The tool uses framework constants to stay portable, but has only been tested on macOS 26.

License

MIT