mirror of
https://github.com/ollama/ollama.git
synced 2026-03-28 03:08:44 +07:00
33ee7168ba
* WIP - MLX backend with gemma3 * MLX: add cmake and go tag build toggles To build the new MLX backend code: cmake --preset MLX cmake --build --preset MLX --parallel cmake --install build --component MLX go build -tags mlx . Note: the main.go entrypoint for the MLX engine will change in a follow up commit. * add experimental image generation runtime * add experimental image generation runtime * MLX: wire up cuda build for linux * MLX: get dependencies correct and dedup This is still too large for a unified github artifact, but is now "correct" for the mlx_cuda_v13 directory. * fix relative link bug in dedup * Add darwin build and readme * add go build tag for mlx dependent code and wire up build_darwin.sh * lint cleanup * macos: build mlx for x86 This will be CPU only. * cuda build instructions and fix drift from mlx bump * stale comment * Delete agent helper doc * Clean up readme.md * Revise README for tokenizer clarity and details Updated README to clarify tokenizer functionality and removed correctness section. --------- Co-authored-by: jmorganca <jmorganca@gmail.com>
47 lines
1.3 KiB
Markdown
47 lines
1.3 KiB
Markdown
# MLX Memory Management
|
|
|
|
| This package will get consolidated with `x/ml/backend/mlx` in the future.
|
|
|
|
## Automatic Tracking
|
|
|
|
All arrays are automatically tracked when created. On `Eval()`, non-kept arrays are freed.
|
|
|
|
### API
|
|
|
|
```go
|
|
result := mlx.Matmul(x, w) // arrays automatically tracked
|
|
mlx.Eval(result) // free non-kept, eval result (auto-kept)
|
|
```
|
|
|
|
### Key Functions
|
|
|
|
- `mlx.Eval(outputs...)` - free non-kept arrays, then evaluate (outputs auto-kept)
|
|
- `mlx.AsyncEval(outputs...)` - async version of Eval (outputs auto-kept)
|
|
- `mlx.Keep(arrays...)` - mark arrays to survive cleanup (for weights, caches)
|
|
- `array.Free()` - mark array for cleanup on next Eval
|
|
|
|
### Loop Pattern
|
|
|
|
```go
|
|
for step := 0; step < maxTokens; step++ {
|
|
logits := model.Forward(token, caches)
|
|
oldToken := token
|
|
token = sample(logits)
|
|
|
|
// Keep cache state across iterations
|
|
for _, c := range caches {
|
|
mlx.Keep(c.State()...)
|
|
}
|
|
|
|
oldToken.Free() // mark for cleanup
|
|
mlx.AsyncEval(token) // frees old, evals new
|
|
}
|
|
```
|
|
|
|
### Notes
|
|
|
|
- `Eval()` and `AsyncEval()` auto-keep their outputs
|
|
- `Free()` marks for cleanup - actual free happens during next Eval
|
|
- Use `Keep()` for weights and cache state that must survive multiple Eval cycles
|
|
- Arrays created inside compiled closures are managed by MLX, not tracked
|