- [ ]**Benchmark suite** — No benchmarks exist. Add: MatMul (various sizes), Softmax, model.Forward (single token), tokenizer.Encode/Decode, full Generate (tokens/sec). Baseline on M3 Ultra.
## Phase 2: Model Support
- [ ]**Gemma3-1B inference validation** — The go-i18n Phase 2a needs 1B model inference for domain classification at ~5K sentences/sec. Validate Gemma3-1B loads and generates correctly via `model.LoadModel()` + `model.Generate()`. Report tokens/sec.
- [ ]**Model loading robustness** — Test with missing files, corrupted safetensors, wrong dtype. Currently no error handling tests for `io.go`.
- [ ]**Add Llama model support** — Only Gemma3 and Qwen3 exist. Llama architecture would cover Meta's model family (Llama 3, CodeLlama).
## Phase 3: Training Pipeline
- [ ]**LoRA fine-tuning end-to-end** — `lora.go` has the adapter but no integration test showing: load base model → apply LoRA → train on small dataset → save adapter → reload. Critical for LEM Lab.
- [ ]**Gradient checkpointing** — `grad.go` has VJP but large models will OOM without checkpointing. Add selective recomputation.
- [ ]**Mixed precision training** — MLX supports BFloat16/Float16. Add dtype selection for training (currently inference uses model's native dtype).
## Phase 4: API Polish
- [ ]**Error handling audit** — `checkError()` only logs. Should return errors to callers instead of silent logging. The C error handler stores last error but Go code doesn't propagate it.
- [ ]**Memory management audit** — Array finalizers use `runtime.SetFinalizer`. Verify no leaks under sustained inference (1000+ tokens). Check C-side deallocation.
- [ ]**Documentation** — Public API has minimal godoc. Add examples for common workflows: load model, generate text, fine-tune with LoRA.
- [ ]**Streaming token iterator** — `model.Generate()` should return an `iter.Seq[Token]` (or channel) for token-by-token streaming. LEM Lab chat UI needs this for real-time output. Currently the generate path is batch-only.
- [ ]**Batch inference API** — go-i18n Phase 2a wants ~5K sentences/sec through Gemma3-1B. Single-request inference won't hit that. Add a batch API that processes N prompts concurrently on the GPU. Even batching 8-16 at a time would help throughput significantly.
- [ ]**Inference metrics** — Expose tokens/sec, peak memory, GPU utilisation as structured data. LEM Lab dashboard and go-ai scoring engine both want this. A simple `Stats` struct returned alongside generation results.
- [ ]**Model quantisation awareness** — MLX supports 4-bit and 8-bit quantised models. The `model.LoadModel()` path should detect and handle quantised safetensors. This matters for running 27B models on 32GB Macs.
- [ ]**Embed-friendly model loading** — For the Core native app (FrankenPHP + Wails), models are on disk but the path discovery needs to be clean. Add `model.Discover(baseDir)` that scans for available models and returns metadata (name, params, quant level, size).
- [ ]**Range-over-func for Array** — If 1.26 stabilises range-over-func, `Array.Iter()` returning `iter.Seq[float32]` (or typed variant) would be cleaner than the current index-based access. Measure overhead vs direct C pointer access.
