go-mlx/TODO.md

TODO.md — go-mlx Task Queue

Dispatched from core/go orchestration. Pick up tasks in order.

---

Phase 1: Standalone Package Hardening

• Verify go generate → test round-trip — ✅ 29/29 tests pass. CMake 3.24+, AppleClang 17.0.0, macOS SDK 26.2. Build takes ~2min on M3 Ultra.
• Add missing tests for core operations — ✅ 86 new tests across 4 files: array_test.go (25), ops_test.go (44), nn_test.go (8), fast_test.go (9). Covers: all scalar/array creation, shape ops, element-wise arithmetic, math functions, matrix ops, reductions, indexing, slicing, fused kernels (RMSNorm, LayerNorm, RoPE, SDPA), Linear, Embedding, RepeatKV. Found non-contiguous view bug in Floats()/DataInt32() — see FINDINGS.md.
• Add missing tests for model/tokenizer/sample/cache — ✅ 33 new tests: cache_test.go (10: KVCache + RotatingKVCache lifecycle, update, bounded, reset), sample_test.go (8: greedy, temperature, topK, chain, stub pass-through), tokenizer_test.go (15: Load/error, BOS/EOS, encode/decode, DecodeToken, SentencePiece space, GPT-2 byte maps). model/ still needs tests (requires model files on disk).
• Benchmark suite — ✅ 29 benchmarks in bench_test.go. Covers: MatMul (128² to 4096², token-shaped 1×2048→32000), Softmax (1K to 128K vocab), element-wise (Add, Mul, SiLU at 1M elements), fused kernels (RMSNorm, LayerNorm, RoPE, SDPA at various shapes), Linear, Embedding, reductions (Sum, Argmax), and full sampler chain (greedy, TopK, TopP, combined). Baselined on M3 Ultra. model.Forward and tokenizer benchmarks deferred to Phase 2 (require model files on disk).

Phase 2: Model Support

• Gemma3-1B inference validation — ✅ End-to-end inference works. 4-bit quantised Gemma3-1B loads and generates coherently at 46 tok/s on M3 Ultra. Fixed: model_type: "gemma3_text" not matched in architecture dispatch, GPT-2 BPE false detection on 262K SentencePiece vocab (checked Ġthe instead of bare Ġ). 3 new tests: inference (greedy, timing), chat template, context cancellation.
• Model loading robustness — ✅ 24 new tests in model_test.go covering: missing/invalid config.json, unsupported architecture, gemma3_text dispatch, missing tokenizer, missing safetensors (was a nil-pointer panic — fixed with early error return in both LoadGemma3 and LoadQwen3), config parsing defaults/quantization/nested text_config, isLayerSliding, resolveWeight with prefix fallback.
• Add Qwen2 model support — ✅ Qwen2 architecture (used by DeepSeek R1) now supported. Shares Qwen3 loader with optional Q/K RMS normalization (Qwen3 has it, Qwen2 does not). Auto-detected from weight presence. DeepSeek R1 7B: 27 tok/s on M3 Ultra. 2 new tests.
• Add Llama model support — ✅ Llama 3 architecture shares Qwen3 loader (same decoder: pre-norm, SwiGLU, GQA, no Q/K norm). Model type detected from config.json model_type field. Llama 3 chat template (<|start_header_id|>) and EOS token (<|eot_id|> id=128009) added. Llama 3.1 8B 4-bit: 30 tok/s on M3 Ultra. 2 new tests.

Phase 3: Training Pipeline

• LoRA fine-tuning end-to-end — ✅ Full pipeline validated: load Gemma3-1B → apply LoRA (rank=8, q_proj+v_proj, 745K params) → 5 training steps with cross-entropy loss (7.15→6.31) → save adapter (2.9MB safetensors) → reload and verify weights match. Uses ValueAndGrad + AdamW. 1 new test in train_test.go.
• Gradient checkpointing — ✅ Checkpoint() validates with real model training. Wraps forward pass to recompute activations during backward. Verified: produces correct gradients (loss 7.15→7.08 in 3 steps, matching non-checkpointed initial loss). 2 new tests: unit (grad_test.go) + model (train_test.go).
• Mixed precision training — ✅ LoRAConfig.DType selects training dtype for A/B matrices. BFloat16 validated: loss 7.15→6.29 in 5 steps, matches Float32 accuracy with half param memory. MLX auto-promotes for cross-dtype ops. 1 new test in train_test.go.

Phase 4: Backend Abstraction — ✅ COMPLETE (19 Feb 2026)

Design doc: docs/plans/2026-02-19-backend-abstraction-design.md Implementation plan: docs/plans/2026-02-19-backend-abstraction-plan.md

All Virgil review items implemented:

• context.Context on TextModel.Generate() — Generate(ctx context.Context, prompt string, opts ...GenerateOption) iter.Seq[Token]. Checks ctx.Done() in the decode loop.
• Err() error on TextModel — Distinguishes normal stop (EOS, max tokens) from errors (OOM, ctx cancelled).
• Chat() on TextModel — Model owns its chat template. Gemma3 and Qwen3 templates implemented.
• Memory control functions at root — SetCacheLimit, SetMemoryLimit, GetActiveMemory, GetPeakMemory, ClearCache delegate to internal/metal.
• Backend registration — register_metal.go auto-registers via build-tagged init().
• All CGO moved to internal/metal/ — 19 source files, 10 test files, 148 tests passing.
• Public API: TextModel, Backend, functional options — Clean root package, compiles on all platforms.
• Integration tests — 7 tests for public API (backend registration, options, LoadModel paths).
• Error handling audit — ✅ checkError() replaced with lastError() error (reads + clears C-level error string). Added Eval(...*Array) error and EvalAsync(...*Array) error as error-returning variants of Materialize. Generate loop propagates errors via m.lastErr. LoadAllSafetensors returns (map, error). Model loaders (gemma3, qwen3) check lastError() after safetensors load. grad.go/lora.go now surface real MLX error messages. 4 new tests in error_test.go.
• Memory management — deterministic cleanup — ✅ Model.Close() now walks the full model tree (GemmaModel/Qwen3Model) and explicitly frees all weight arrays via Free(). Helpers: freeLinear, freeEmbedding, freeRMSNorm, freeCaches, closeGemma, closeQwen3 in close.go. Handles tied output weights (skip double-free), nil safety, idempotent Close(). 8 new tests in close_test.go.
• Documentation — Public API has godoc but needs examples for common workflows.

Phase 5: Ecosystem Integration (Virgil wishlist)

• Batch inference API — ✅ Classify (prefill-only, fast path) and BatchGenerate (autoregressive) implemented. Added ForwardMasked to InternalModel interface, threaded attention masks through Gemma3 and Qwen3 decoders. Mask: [N, 1, L, L] combining causal + padding (0=attend, -inf=ignore). Right-padded, sorted by length. Gemma3-1B 4-bit: 152 prompts/s classify (4 prompts), BatchGenerate produces coherent per-prompt output. Types (ClassifyResult, BatchResult, WithLogits) in go-inference. 6 new tests (3 mask unit, 3 model). Design doc: docs/plans/2026-02-19-batch-inference-design.md.
• Inference metrics — ✅ GenerateMetrics type in go-inference with Metrics() on TextModel. Captures: prefill/decode timing, token counts, throughput (tok/s), peak and active GPU memory. Instrumented Generate, Classify, and BatchGenerate. Gemma3-1B 4-bit: prefill 246 tok/s, decode 82 tok/s, peak 6.2 GB. 1 new test.
• Model quantisation awareness — ✅ ModelInfo type in go-inference with Info() on TextModel. Exposes architecture, vocab size, layer count, hidden dimension, quantisation bits and group size. Loader already handles quantised safetensors transparently. 1 new test.
• Embed-friendly model loading — Add Discover(baseDir) that scans for available models and returns metadata.
• mlxlm/ backend — Python subprocess wrapper via core/go/pkg/process. Implements mlx.Backend for mlx_lm compatibility.

Phase 6: Go 1.26 Modernisation

• Evaluate Go 1.26 features — ✅ Documented in FINDINGS.md. Key wins: CGO ~30% faster (free), Green Tea GC default (10-40% less overhead, helps Array finalisers), slice stack alloc.
• Range-over-func for Array — ✅ Array.Iter() iter.Seq[float32] implemented in array.go. Handles non-contiguous arrays via ensureContiguous(). Supports early break. 4 tests: basic, 2D flatten, transposed, early break.

---

go-inference Integration — ✅ COMPLETE (19 Feb 2026)

All types (TextModel, Backend, Token, Message, options) moved to shared forge.lthn.ai/core/go-inference package. go-mlx is now a pure backend implementation — import _ "forge.lthn.ai/core/go-mlx" to register the "metal" backend. See FINDINGS.md for migration details.

Upstream Dependencies

• go-i18n Phase 2a is blocked on this package providing working Gemma3-1B inference
• go-ml/backend_mlx.go needs updating to use inference.LoadModel() + m.Generate() (types from go-inference, _ "go-mlx" for Metal registration)
• go-ai has a replace directive pointing at ../go-mlx. No code changes needed in go-ai itself.
• go-rocm — sibling backend for AMD GPUs, implements same inference.Backend interface
• LEM Lab uses MLXBackend via go-ml. Migration transparent once go-ml updates.

Functional Options Convention

Virgil confirms: the WithMaxTokens(n) functional option pattern is the right call for this package.

core/go/pkg/process (for mlxlm backend, Phase 5)

Virgil confirms: no changes needed. The process package provides everything needed for the mlxlm subprocess backend.

Virgil Code Review — 19 Feb 2026

Full codebase review after Phase 4 completion + go-inference integration. Grouped by priority.

Critical — Fix Before Phase 2

• Error handler thread safety — ✅ last_mlx_error now uses _Atomic(const char*) with atomic_store_explicit (release) / atomic_exchange_explicit (acquire). Thread-safe even if MLX calls the error handler from background threads.

• -mmacosx-version-min=26.0 is wrong — ✅ Changed to 13.3 (MLX's own minimum). No longer locks out macOS 14/15 users.

• LoadOption is ignored in metalBackend.LoadModel() — ✅ Now calls inference.ApplyLoadOpts(). ContextLen passed through to metal.LoadConfig → stored on Model → replaces unbounded KVCache with RotatingKVCache(contextLen) in generate loop. GPULayers=0 logs a warning (Metal always uses full GPU offload). newArray test: TestNewCaches_ContextLen.

Important — Should Fix

• KV cache leak between turns — ✅ Documented in Generate() godoc: each call allocates fresh KV caches released to GC; call ClearCache() between turns for prompt reclaim. Cache reuse across turns deferred to batch inference design (Phase 5).

• RepeatPenalty is accepted but never applied — ✅ Implemented applyRepeatPenalty() in generate.go. Tracks generated token IDs, deduplicates, then for each seen token: divides positive logits by penalty, multiplies negative logits by penalty. Applied before sampling when RepeatPenalty > 1.0. 2 new tests.

• DefaultGPUStream() / DefaultCPUStream() leak and mislead — ✅ Now cached with sync.Once like DefaultStream(). No more allocation on every call.

• Tokenizer Encode() is character-level only — ✅ Implemented bpeMerge() — standard BPE algorithm using merge rank lookup. Both SentencePiece Encode() and GPT-2 encodeGPT2() now split into characters, apply BPE merges, then look up merged symbols. Merge ranks built during tokenizer load. 3 new tests.

• CompileShapeless is dead code — ✅ Removed C closure, callback, sync.Map, and nextID infrastructure. CompiledFunc is now a plain function wrapper with mutex. CompileShapeless() and Call() signatures unchanged (gemma3.go GELU still works).

Minor — Nice to Have

• Rename New() → newArray() — ✅ Renamed via IDE refactoring (112 usages updated). Unexported, signals internal-only intent.

• Collect() is unused — ✅ Removed function and its test. Dead code eliminated.

• qwen3.go — second json.Unmarshal error discarded — ✅ Now checks and returns the error. gemma3.go already handled it correctly.

• Document AsStrided stride formula — ✅ Added comment explaining the stride derivation for the [B,L,H*D] → [B,H,L,D] virtual transpose.

Questions for You to Consider

1. Per-step intermediate freeing: The design doc mentions freeIntermediates(logits) per decode step to reduce GC pressure. This isn't implemented — the generate loop creates ~500 intermediate arrays per forward pass that rely on GC finalizers. Is Go 1.26 Green Tea GC considered sufficient, or is explicit per-step freeing still planned?

2. SentencePiece BPE: The merges field is parsed but never used. For Gemma3's SentencePiece tokenizer, is character-level encoding sufficient (because the vocab contains full token strings), or is merge application a known gap for Phase 2?

3. nextID in compile.go: nextID is a uintptr used as unsafe.Pointer key into sync.Map. This works but uintptr(0) is never valid (starts at 1 after first increment). If CompileShapeless is kept, consider using atomic.AddUint64 instead of mutex + plain increment.

Workflow

1. Virgil in core/go writes tasks here after research
2. This repo's session picks up tasks in phase order
3. Mark [x] when done, note commit hash
4. newArray discoveries → add tasks, flag in FINDINGS.md