Skip to content

RISKS.md

Latest commit

 

History

History
74 lines (60 loc) · 3.42 KB

RISKS.md

File metadata and controls

74 lines (60 loc) · 3.42 KB

Cube Memory — known risks ledger

Captured from the Phase 1→2 transition adversarial sweep (2026-04-26). Each item is something the random-input parity tests do not exercise but production will. Severity tags reflect "how likely is this to bite real users."

High — must guard before Phase 2 deployment

  • top_k > n_slots is silently clamped on GPU but unrestricted in PyTorch. Layer constructor should validate top_k <= n_slots and top_k <= MAX_TOP_K. Add to phase1/cube_memory_layer.py::__init__.

  • top_k = 0 crashes softmax in the PyTorch retrieve and produces uniform 1/k weights with garbage indices on GPU. Constructor should refuse top_k < 1.

Medium — test before betting on it

  • Cleanup tie-break diverges between CPU argmax (first-wins) and GPU > comparison (last-wins). Random unit phasors don't tie at fp32 precision; deterministic codebooks with duplicate entries do. Either standardize the comparison (>= vs >) on both sides or document that the layer is not bit-deterministic across backends.

  • STE gradient through cleanup vanishes at exact codebook boundaries. Gradient on role_proj is zero whenever the query exactly aligns with a snapped phasor. Practically rare with random init, but the test set never explicitly covers it. Add a unit test that initializes a query at a codebook entry and asserts the forward+backward path still moves the loss.

  • Distillation test fixture's FakeLoader resets seed every epoch, so all 10 steps see identical batches. Hides shuffling bugs that would matter at real distillation scale. Fix by using a per-epoch RNG reseed or asserting batch variety.

  • GGUF buffer layout for slot_keys and slot_values is unspecified. PyTorch is row-major (n_slots, 2*d_codebook); the ggml-vulkan backend's tensor stride conventions need to match exactly. Test before claiming Phase 2 is integrated. If a silent transpose creeps in, the dot products compute on the wrong axis and retrieve returns plausible-looking garbage.

  • Unitize on Vec2::ZERO returns Vec2::ZERO (the eps floor prevents division-by-zero but does not produce a unit vector). Algebraically wrong but doesn't crash. Mitigate by making the cleanup path map Vec2::ZERO to a canonical phasor (1, 0).

Low — flag if scaling up

  • Cleanup with d ≤ 4 has insufficient noise margin to be reliable. Document a d >= 64 floor in the layer.
  • n_slots = 1 with top_k > 1 picks the same slot k times and softmax-weights it equally. Mathematically correct, semantically pointless. Validate n_slots >= top_k at init.
  • Naive dot product in the GPU retrieve loop could accumulate fp32 error at very large d_key. Probably fine to d_key ≈ 4096; re-check before going beyond.

Phase 2 implications

Three of the items above gate Phase 2 acceptance:

  1. The top_k/n_slots invariants must be enforced in the GGUF loader (refuse to load a model that violates them).
  2. The buffer-layout assumption must be verified by a round-trip parity test: PyTorch forward → export to GGUF → load in ggml → ggml forward → compare to PyTorch output. Same input, same weights, same output to fp32 tolerance.
  3. The shader push-constant order must match the GGUF tensor order exactly. Today both are written by us; in a future where ggml reorders tensors during graph optimization, that assumption breaks. Add a runtime assert.