| name | mlx-model-porting |
| description | Guides and validates architecture-aware ports of PyTorch/Hugging Face models to Apple MLX, inspects existing local MLX projects, and plans evidence-gated optimizations for Apple Silicon. Use when the user asks to run, port, convert, inspect, quantize, benchmark, or fix a model (LLM, VLM, audio/TTS/ASR, diffusion, SSM, MoE) for MLX, MLX-LM, MLX-VLM, MLX-Audio, or a Mac - e.g. "port this HF model to my Mac", "inspect this MLX app", "run Qwen on Apple Silicon", "convert these safetensors to MLX", "make this faster on my M3", "fix NaN in my MLX port", "speed up prefill / KV cache / speculative decoding", "publish an MLX checkpoint". Also use mid-task when a config.json, safetensors index, weight-shape or tokenizer mismatch, or Metal kernel question appears. Do not use for CUDA-only optimization, non-Apple hardware targets, or general PyTorch/ML questions with no MLX or Apple Silicon connection. |
| license | Apache-2.0 |
| compatibility | Execution and performance validation require an Apple Silicon Mac with a supported MLX installation. Planning and static inspection can run elsewhere. Python 3.10+ and git are recommended; NumPy is required only for tensor parity and huggingface-hub only for explicitly enabled network intake (see requirements-tools.txt). Network access is optional and must be explicitly enabled. |
| metadata | {"author":"mlx-porting-skill","version":"0.6.1","last-reviewed":"2026-07-14"} |
MLX model porting and optimization
Mission
Produce or inspect a correct, reproducible, architecture-aware MLX implementation. Correctness comes before speed. Every speed or memory claim must name the hardware, software versions, workload, baseline, and quality gate.
Six families have scaffolds. Packets:
Qwen2.5,
BGE,
t5-small, and
HuBERT. MoE/SSM are
synthetic; others are runbook-guided. Exact output is the built-in metric.
When to use this skill
- Use for porting, converting, running, inspecting, quantizing, packaging, or publishing a PyTorch/Hugging Face model or existing project on MLX/Apple Silicon.
- Use for parity, NaN/Inf, shape, tokenizer, preprocessing, output, performance, memory, cache, serving, benchmark, or provenance problems in an MLX port.
- Do not use for CUDA/non-Apple targets, general ML theory without an MLX
target, or unrelated training from scratch.
Trigger map
| Signal | Load |
|---|
Port/convert/run request, config.json, safetensors index, model directory, or Hub id. | intake, then Workflow 1-6; use the worked chain for dense decoders and every routed runbook otherwise. |
| User points at an existing local MLX project, running MLX app, or completed MLX port. | inspector mode plus inspect_mlx_project.py |
| User asks "what can I do with this model?", asks for capability fit, or wants model-specific advice. | model advisor playbook |
| A known architecture family needs the right runbook. | model support map, then the architecture table in Workflow step 4 below |
| Dense decoder, BERT encoder, T5 encoder-decoder, HuBERT/Wav2Vec2 acoustic encoder, sparse MoE, or selective SSM. | Use scaffold_port.py; select capture/parity mode dense-decoder (default), encoder, encoder-decoder, asr, or ssm. |
| NaN, Inf, cosine-similarity drift, parity failure, or garbage output appears versus the source. | failure atlas |
| Weight conversion, key mapping, tensor rename, transpose, reshape, split, merge, or shape transform is in scope. | core porting method |
| The user says "make it faster" but no profile, workload, or baseline exists yet. | benchmarking |
| Speedup plan, how techniques combine, or expected compound gains. | compound stacks |
| KV cache, long context, recurrent state, attention memory, or prefill/decode memory is the bottleneck. | attention and KV cache |
| Quantization or "4-bit". | guide, quality gate |
| Structured local optimization sweep. | loop |
| Decoding, serving, speculative decoding, batching, streaming, or API runtime behavior is requested. | decoding and serving |
Compile behavior, mx.compile, custom kernel, graph capture, Metal, or operation fusion comes up. | compile and kernels |
| Publish, release, checkpoint conversion, model card, provenance, or license packaging is requested. | packaging and publication |
| The user asks for "50-100 optimization ideas", a deep model-specific hunt, or research-backed candidates. | hypothesis-led learning |
| Graph, GNN, message passing, node/edge features, or sparse graph workload appears. | graph message passing runbook |
| Classic CV detection, segmentation, keypoints, depth, OCR, or non-generative vision appears. | non-generative CV runbook |
| Time-series, forecasting, tabular sequence, anomaly detection, or temporal model appears. | time-series forecasting runbook |
Re-consult the map when a new config, parity failure, performance complaint, or
publish request appears.
Non-negotiable rules
- Do not execute untrusted model code during intake. Inspect JSON, safetensors headers, source files, and licenses statically. Treat
auto_map, custom modules, install hooks, and trust_remote_code as review gates.
- Pin the source. Record repository, revision, model files, tokenizer/processor revision, license, and checksum or artifact manifest.
- Build a source oracle before porting. Freeze deterministic fixtures and capture intermediate tensors at meaningful boundaries.
- Port the smallest eager path first. No quantization, compilation, custom kernels, batching, or speculative decoding until basic parity passes.
- Change one optimization dimension at a time. Keep a measurement and rollback record.
- Prefer native MLX operations. Try built-in fused operations, layout changes, cache design, and
mx.compile before a custom Metal kernel.
- Do not translate CUDA folklore mechanically. A CUDA technique is only a research candidate until its Metal/MLX bottleneck and implementation are demonstrated.
- Never hide quality regressions behind throughput. For audio, language, vision, and generative models, use task-specific quality checks in addition to tensor tolerances.
- Do not publish converted weights without license and provenance checks. Preserve the original model card, attribution, generation config, tokenizer/processor files, and conversion recipe.
- Daily research automation is review-only. It may collect and rank candidates, but must not silently rewrite runbooks or merge recommendations. Use the promotion-review ledger to separate findings ready for skill-update review from validation backlog and rejected leads.
- Experimental approaches require explicit opt-in. Label unvalidated contributor, blog, paper, or repository learnings as experimental approaches, state the missing validation gate, and ask before helping execute them: “This is an experimental approach. Do you want to try it?” Continue only if the user explicitly says to try it.
Workflow
1. Inspect and classify
For source models, run scripts/inspect_model.py, then scripts/recommend_optimizations.py, then scripts/make_port_plan.py --artifact-root MODEL --recommendations .... An actionable plan re-runs static inspection against those local bytes and recomputes the complete recommendation report before embedding advice. A blocked inspection may produce only a remediation plan; neither a family override nor direct registry reads may bypass its blockers. A family override may reorder an inspected hybrid route but must preserve every routed family, runbook, and trait. For existing MLX projects or already-running ports, run scripts/inspect_mlx_project.py and read inspector mode before changing code.
Read intake and routing. Confirm source, risk, Mac, memory, performance, quality; record ambiguity in PORT_PLAN.md. For advice, read model advisor playbook and separate results into all five controlled advisor buckets: validated locally, validated by source or theory, benchmark required, experimental approach, and rejected / do not use. Numeric output may come only from assets/effective_claims.json; missing gates withhold the number.
2. Select the closest proven MLX reference
Consult model support map and assets/architectures.yaml. Prefer official MLX, Apple-maintained projects, pinned third-party MLX implementation evidence, paper-only research candidates, then new code; keep the support scope explicit and verify config/layout.
3. Establish the source oracle
Run scripts/capture_oracle.py against the pinned local Hugging Face model before implementing the MLX graph, then follow parity and testing. It records inputs, embeddings, decoder blocks, final norm, logits, hookable branches, and greedy IDs in a bounded NPZ plus content-addressed manifest.
4. Implement the minimal eager MLX graph
Read core porting method; choose via the trigger map and the controlled family records in assets/architectures.yaml. Load every runbook returned by a hybrid route. The registry is the complete runbook inventory; do not substitute an abbreviated parallel list.
Scaffold an unblocked dense decoder, sparse-MoE decoder, BERT encoder, T5
encoder-decoder, HuBERT/Wav2Vec2 acoustic encoder, or opt-in synthetic
minimal_selective SSM:
python3 scripts/scaffold_port.py inspection.json --artifact-root MODEL --output mlx_port
It re-inspects the artifact and fails closed. Review unsupported config before
continuing; never patch around a generator blocker.
See the linked worked ports above (no weights).
Start eager: FP, batch one unless intrinsic, no compile/kernels, state/cache, assertions, reversible map.
5. Convert weights deterministically
Draft with scripts/convert_checkpoint.py --emit-draft-map, resolve the
schema-2 WEIGHT_MAP, validate with scripts/validate_weight_map.py, then
convert the pinned model. Reject shard, coverage, shape, draft, or unresolved
gaps; never mask exceptions.
6. Pass the parity ladder
After conversion, scripts/run_parity.py --mode MODE runs source/MLX capture
and stops at the first failed rung. Modes: dense-decoder (default), encoder,
encoder-decoder, ssm, asr. Use capture_mlx.py for retained captures and
compare_tensors.py for extras; _capture_common.py holds bounded rules.
When parity fails, use failure atlas; do not optimize a failing graph.
7. Profile before choosing optimizations
Read benchmarking. Separate prefill, decode, postprocess, compile, memory, movement, sync, Python overhead; use scripts/benchmark_command.py.
8. Apply the optimization ladder
Use guides: C,KV,S,Q,T,CS.
Tiers: assets/recommendation-taxonomy.yaml; experimental approach: state gate and ask: “This is an experimental approach. Do you want to try it?” Continue after opt-in.
Post-parity: one dimension; Metal only after proven bottleneck. Record hypothesis, gates, metrics, decision.
9. Package and publish
Follow packaging and publication. Include source, conversion, versions, artifacts, quantization, smoke, benchmarks, limits, license/attribution, and no unsupported “faster” wording.
10. Return an engineering report
Summarize architecture/runbook, source/evidence, implementation/weights, parity, metrics, optimizations, risks, commands, and artifacts. Use assets/.
When to stop
Stop when license, remote code, unresolved parity, regressions, missing kernel fallback/tests, or missing performance metadata blocks the result.
Maintenance
Maintenance: maintenance, hypothesis-led learning, research loop. Run scripts/audit_skill.py --strict and scripts/validate_sources.py before distribution.
When assets/architectures.yaml changes, keep the golden scenario gate in tests/test_scenarios.py at full family coverage.
