- defer freeCaches() in Generate and InspectAttention
- Free orphaned arrays during KVCache growth and slice updates
- Free per-token scalar intermediates in samplers and ops
- Free intermediate arrays in applyRepeatPenalty
Found by 3-way review: Claude explorer, Codex (gpt-5.3), Gemini Ultra.
Gemini implemented the fixes.
Co-Authored-By: Gemini <noreply@google.com>
Co-Authored-By: Virgil <virgil@lethean.io>
Instrument Generate, Classify, and BatchGenerate with:
- Prefill/decode timing (separate phases)
- Token counts (prompt + generated)
- Throughput (tok/s for each phase)
- Peak and active GPU memory via Metal allocator
Wire through metalAdapter.Metrics() to go-inference interface.
Test validates all fields populated after generation.
Gemma3-1B 4-bit on M3 Ultra: prefill 246 tok/s, decode 82 tok/s,
peak 6.2 GB GPU memory.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add ForwardMasked to InternalModel, Gemma3 and Qwen3 architectures
- Thread attention mask through decoder layers and SDPA calls
- Use ScaledDotProductAttentionWithMask when explicit mask provided
- Create batch.go with padded batching, mask construction, Classify
(prefill-only) and BatchGenerate (autoregressive) implementations
- Wire Classify/BatchGenerate through metalAdapter to go-inference
- Tests: mask unit tests (shape, values, multi-batch), Classify with
4 prompts (152 prompts/s), WithLogits, BatchGenerate with 2 prompts
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LoRA A/B matrices can now be created in BFloat16 or Float16 for mixed
precision training. DType field added to LoRAConfig, passed through
ApplyLoRA and NewLoRALinear. MLX auto-promotes for cross-dtype ops.
BFloat16 validated: loss 7.15→6.29, matches Float32 accuracy with
half param memory.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Checkpoint() wraps forward pass to recompute activations during
backward, trading compute for memory. Verified with Gemma3-1B LoRA
training: produces correct gradients (loss 7.15→7.08, matches
non-checkpointed initial loss). Unit test confirms gradient
correctness on simple function (sum(x^2), grad=[2,4,6]).
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validates full pipeline: load Gemma3-1B → apply LoRA (rank=8, 745K
params across 52 layers) → train 5 steps with cross-entropy loss
(7.15→6.31) → save adapter to safetensors → reload and verify all
weights match. Uses ValueAndGrad for autograd + AdamW optimiser.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Llama shares the Qwen3 loader (same decoder: pre-norm, SwiGLU, GQA).
Model type now detected from config.json model_type field instead of
weight-only heuristic. Llama 3 chat template and EOS token added.
Model tests now clear Metal GPU cache between runs.
Llama 3.1 8B Instruct 4-bit: 30 tok/s on M3 Ultra.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Qwen2 and Qwen3 share the same architecture — Qwen3 adds Q/K RMS
normalization which Qwen2 lacks. The loader auto-detects the variant
from weight presence and reports the correct ModelType().
- Add "qwen2" to architecture dispatch in model.go
- Make Q/K norm optional in attention forward (nil-safe check)
- Store detected model type on Qwen3Model struct
- Add "qwen2" to chat template routing
- DeepSeek R1 7B (4-bit): 27 tok/s on M3 Ultra
- 2 new tests: inference + chat
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
24 new tests covering error paths in model loading:
- Missing/invalid config.json, unsupported architecture
- Missing tokenizer.json for both Gemma3 and Qwen3
- Missing safetensors: was a nil-pointer panic in precomputeScaledWeights,
fixed with early error return in both LoadGemma3 and LoadQwen3
- Config parsing: defaults, quantization, nested text_config
- isLayerSliding sliding window pattern logic
- resolveWeight with language_model. prefix fallback
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Rename New() → newArray() to signal internal-only intent (112 usages)
- Remove unused Collect() function and its test
- Fix discarded json.Unmarshal error in qwen3.go
- Document AsStrided stride formula in gemma3.go
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1. RepeatPenalty: implemented applyRepeatPenalty() — tracks generated
token IDs, deduplicates, divides positive logits by penalty and
multiplies negative logits by penalty. 2 new tests.
2. DefaultGPUStream/DefaultCPUStream: now cached with sync.Once,
no more C stream allocation on every call.
3. CompileShapeless: removed dead C closure, callback, sync.Map,
and nextID infrastructure. CompiledFunc is now a plain function
wrapper with mutex. API unchanged.
4. Tokenizer BPE: implemented bpeMerge() — standard BPE algorithm
using merge rank lookup. Both SentencePiece and GPT-2 Encode paths
now apply merges instead of falling back to character-level lookup.
3 new tests.
5. KV cache lifecycle: documented in Generate() godoc — fresh caches
per call, ClearCache() between turns for prompt Metal reclaim.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1. Error handler thread safety: last_mlx_error now uses _Atomic(const char*)
with atomic_store_explicit/atomic_exchange_explicit (release/acquire).
2. macOS version minimum: -mmacosx-version-min changed from 26.0 to 13.3
(MLX's own minimum), no longer locks out macOS 14/15 users.
3. LoadOption applied in metalBackend.LoadModel(): calls ApplyLoadOpts(),
passes ContextLen through to Model which replaces unbounded KVCache
with RotatingKVCache when set. GPULayers=0 logs a warning.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Model.Close() now walks the full model tree (Gemma3/Qwen3) and
explicitly frees all weight arrays. Handles tied output weights,
nil safety, idempotent double-close. Helpers: freeLinear,
freeEmbedding, freeRMSNorm, freeCaches, closeGemma, closeQwen3.
Array.Iter() returns iter.Seq[float32] for range-over-func iteration.
Handles non-contiguous arrays and supports early break.
192 tests passing (12 new: 8 close, 4 iter).
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace checkError() log+swallow with lastError() that returns real MLX
error messages. Add Eval/EvalAsync as error-returning variants of
Materialize. Generate loop now propagates GPU errors via model.Err().
LoadAllSafetensors returns (map, error). Model loaders check lastError()
after safetensors load. 180 tests passing.
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New bindings from mlx-c memory.h and metal.h:
- GetCacheMemory() — current allocator cache size
- ResetPeakMemory() — reset high-water mark
- SetWiredLimit() — control wired memory limit
- GetDeviceInfo() — GPU architecture, max buffer, memory size
All exposed at root package level via register_metal.go delegates.
Co-Authored-By: Virgil <virgil@lethean.io>
New ops: CumSum, Sort, Argsort, Greater, MaxAxis — all bound to mlx-c.
TopP (nucleus) sampling now fully implemented: sorts probabilities
descending, computes cumulative sum, masks tokens beyond the threshold,
and scatters the mask back to original positions via argsort.
MinP sampling now fully implemented: computes softmax, finds max
probability, masks tokens below min_p * max_prob.
Both were previously stubs that passed through logits unchanged.
10 new tests (CumSum variants, Sort, Argsort, Greater, MaxAxis,
TopP, MinP). 176 total tests passing.
Co-Authored-By: Virgil <virgil@lethean.io>
Bind mlx_contiguous and _mlx_array_is_row_contiguous from mlx-c.
Floats(), DataInt32(), and Ints() now automatically handle non-contiguous
arrays (from Transpose, BroadcastTo, SliceAxis, etc.) by checking
IsRowContiguous() and making a contiguous copy when needed.
Previously these methods returned silently wrong data for view arrays.
The old workaround of Reshape(arr, totalSize) is no longer needed.
7 new tests for contiguous handling (transpose, broadcast, slice views).
Co-Authored-By: Virgil <virgil@lethean.io>
Replace local TextModel, Backend, Token, Message, and option types with
forge.lthn.ai/core/go-inference. go-mlx is now a pure backend that
registers "metal" into the shared inference registry via init().
Deleted: textmodel.go, options.go, backend.go
Updated: register_metal.go (implements inference.Backend with Available()),
mlx_test.go (uses inference.* types, 4 new tests), go.mod,
internal/metal/generate.go (added RepeatPenalty)
159 tests passing (148 internal/metal + 11 root).
Co-Authored-By: Virgil <virgil@lethean.io>
Move foundation CGO files from root package to internal/metal/ package.
Changes package declaration from `package mlx` to `package metal`.
Updates CGO SRCDIR paths to account for new location (two levels deeper).
Extracts go:generate directives into root generate.go.
Co-Authored-By: Virgil <virgil@lethean.io>
