Commit graph

12 commits

Author SHA1 Message Date
Snider
754d6e2f93 fix(metal): error handling audit — propagate MLX errors instead of swallowing
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>
2026-02-19 20:59:37 +00:00
Snider
ff01175a62 bench(metal): add 29 benchmarks baselined on M3 Ultra
MatMul (128² to 4096², token projection), Softmax, element-wise
ops, fused Metal kernels (RMSNorm, LayerNorm, RoPE, SDPA), Linear,
Embedding, reductions, and full sampler chain. CGO floor ~170μs.

Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-19 20:47:25 +00:00
Snider
ca6b16eaf2 feat(metal): bind memory diagnostics and device info
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>
2026-02-19 20:39:51 +00:00
Snider
f39126f6bd feat(metal): bind CumSum, implement TopP and MinP sampling
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>
2026-02-19 20:39:44 +00:00
Snider
df0b300b1a fix(metal): auto-contiguous data access for non-contiguous arrays
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>
2026-02-19 20:39:36 +00:00
Snider
bff97ccf19 feat(api): migrate to go-inference shared interfaces
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>
2026-02-19 20:15:42 +00:00
Snider
4d1bff3d78 refactor(api): clean root package — interfaces only, metal auto-registered
Root package now contains only:
- mlx.go: package doc + go:generate directives
- textmodel.go: TextModel, Token, Message interfaces
- options.go: GenerateOption, LoadOption functional options
- backend.go: Backend interface, Register/Get/Default/LoadModel
- register_metal.go: build-tagged init() + adapter + memory delegates
- mlx_stub.go: non-darwin fallback

internal/metal/ has its own Token, GenerateConfig, Model types.
register_metal.go adapts between the two via metalAdapter.

Co-Authored-By: Virgil <virgil@lethean.io>
2026-02-19 20:04:19 +00:00
Snider
c612c3e060 refactor(metal): move all tests to internal/metal (148 tests passing)
Co-Authored-By: Virgil <virgil@lethean.io>
2026-02-19 20:00:02 +00:00
Snider
08976aa504 refactor(metal): flatten model, tokenizer, sample, cache into internal/metal
Co-Authored-By: Virgil <virgil@lethean.io>
2026-02-19 19:51:14 +00:00
Snider
a669d1d9c1 refactor(metal): move nn, io, grad, lora, optim to internal/metal
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-19 19:39:58 +00:00
Snider
d6a49544bd refactor(metal): move ops, slice, random, fast, compile to internal/metal
Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-19 19:39:49 +00:00
Snider
1cf5178c80 refactor(metal): move dtype, array, metal, stream to internal/metal
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>
2026-02-19 19:34:38 +00:00