core/go-mlx
8
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions
go-mlx/cpp/TODO.md
Snider f507620e85 docs(cpp): complete mlx-c API research — all tasks done 
CLion Claude session 1 completed all orientation and priority tasks:
- Full API surface map (180 ops, Go binds ~40)
- mlx_contiguous found (fixes Floats() bug)
- mlx_cumsum found (unblocks TopP sampling)
- Memory management fully documented (refcounted, safe to free)
- Error model documented (free-form strings only)
- Device info API documented (mlx_metal_device_info)

Co-Authored-By: Virgil <virgil@lethean.io>
2026-02-19 20:40:00 +00:00

3.1 KiB
Raw Blame History

TODO.md — go-mlx C++ Task Queue

Tasks for the CLion Claude session. Written by GoLand Claude or Virgil.

Orientation (First Session)

  • Map the mlx-c API surface — Read all 27 headers. Full map in FINDINGS.md: ~180 ops functions, Go binds ~40. Identified 8 high-priority unbound functions. (Done 2026-02-19)
  • Understand the error model — Free-form strings only ("<msg> at <file>:<line>"). No error codes or categories. Handler stores string, Go checks return code. Details in FINDINGS.md. (Done 2026-02-19)
  • Check memory management patterns — Arrays are refcounted via shared_ptr<ArrayDesc>. Double-free is UB. Free during async is safe. NULL-free is safe. Details in FINDINGS.md. (Done 2026-02-19)

Priority Tasks (from GoLand Claude)

  • Find mlx_contiguous or equivalent — FOUND: mlx_contiguous(res, a, allow_col_major, stream) at ops.h:220. Plus _mlx_array_is_row_contiguous() for checking. GoLand Claude: see FINDINGS.md for recommended pattern. (Done 2026-02-19)
  • Verify mlx_array_data_* eval semantics — Does NOT auto-evaluate. Returns raw buffer pointer (crash/garbage if unevaluated). Materialise() before data access is essential. item() auto-evaluates but data() does not. (Done 2026-02-19)
  • Check if mlx_cumsum exists — FOUND: mlx_cumsum(res, a, axis, reverse, inclusive, stream) at ops.h:344. GoLand Claude can now implement proper TopP sampling. (Done 2026-02-19)
  • Survey mlx_contiguous / mlx_flatten / mlx_copy — All three exist. mlx_contiguous is the correct tool (forces row-major). mlx_copy may preserve non-contiguous layout. mlx_flatten works but changes shape semantics. (Done 2026-02-19)

Memory Management Research (from Backend Abstraction Design)

  • What does mlx_clear_cache() release? — Releases allocator pool cache back to system. Does NOT touch active memory. Safe mid-generation. (Done 2026-02-19)
  • Is mlx_array_free() safe on graph-referenced arrays? — Yes, safe. Arrays use shared_ptr<ArrayDesc>. Freeing the C handle just decrements refcount. Graph computation continues normally. (Done 2026-02-19)
  • MLX allocator pool behaviour — mlx_array_free() returns memory to internal pool (not system). Pool reuses allocations. Under sustained inference, memory should plateau. Call mlx_clear_cache() to release pool to system if needed. (Done 2026-02-19)
  • Research structured error info — No structured info available. Free-form string only. Format is stable: "<message> at <file>:<line>". GoLand Claude should use return code (0/1) + stored error string pattern. (Done 2026-02-19)

Standing Tasks

  • API gap analysis — When the GoLand Claude needs a C function that isn't exposed by mlx-c, document the gap here and research if upstream mlx-c supports it or if a patch is needed.

Workflow

  1. GoLand Claude or Virgil writes tasks here
  2. Pick up in order, mark [x] when done
  3. New findings → cpp/FINDINGS.md
  4. If Go changes needed → note in FINDINGS.md for GoLand Claude