Elamur Groupantsgparse-test-softmax-backward-bounded-check
The groupant beside gparse/tests/softmax_backward.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial | approximant |
| ast | partial | ../syntax.json |
| cost | symbolic source derived | invariant |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | partial bounded checks | tests |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| id | type | label | where |
|---|---|---|---|
behavior_complement | behavior_complement | partial source derived model | behavior_complement |
reverse_reconstruction | reverse_reconstruction | partial source derived model | reverse_reconstruction |
layout_cases | layout_cases | partial source derived model | layout_cases |
finite_difference | finite_difference | partial source derived model | finite_difference |
stable_mask | stable_mask | partial source derived model | stable_mask |
negative_reference | negative_reference | partial source derived model | negative_reference |
Model text
descriptant
Source-bound description
CUDA gradient test builds three rows for widths1,3,17,33,257; compares fused/reference probabilities and weighted gradients, f64 central differences, narrow offsets/transposes, and stable masked logits.
Epoch 2026-10-31
Prospective model was recorded before edits. Added tests use bounded synthetic fixtures; production algorithms unchanged. Coverage is partial, not proof. No trained corpus, live teacher or throughput experiment was executed.
approximant
Forward approximant
CUDA gradient test builds three rows for widths1,3,17,33,257; compares fused/reference probabilities and weighted gradients, f64 central differences, narrow offsets/transposes, and stable masked logits.
Domain: valid source-specific dimensions and resources stated in invariant; no automatic production substitution. Numerical observations are finite checks, not an end-to-end relative bound. Executable Rust surrogate: unsupported gap; existing exact tests are not surrogates. Future metric-specific caps may be 1/2/5/10%, never above 10%.
invariant
Invariant and cost
Every checked forward/backward discrepancy<2e-6; Ā±1000 logits stay finite and selected -1e9 columns exactly zero; upstream fused gradient absent.
Cost
O(sum widths) tensor storage and a fixed number of full-row objective evaluations per width.
Allocation, finite arithmetic, platform dimensions and external failures are not universally certified.
models
Typed model catalog
behavior_complement
Reference-gradient absence or threshold exceedance fails; no fallback to CPU is accepted because CUDA creation unwraps.
reverse_reconstruction
Probabilities admit row-shifted logits; derivative tolerances admit intervals around finite-difference observations.
layout_cases
narrow(0,1,2) tests nonzero contiguous offset; transpose tests contiguous conversion and backward through view.
finite_difference
Four indices per width, possibly duplicated at width1; epsilon1e-4. Do not count these as four distinct coordinates at width1.
stable_mask
[1000,999,-1e9] and [-1000,-1001,-1e9] produce finite rows and exact masked zeros on tested cells.
negative_reference
ops::softmax_last_dim graph must lack input gradient; this is distinct from custom derivative correctness.
tests
Checks and limits
Selected parent checks relevant to this fragment (not exhaustive): cpu_f64_uses_reference; fused_softmax_matches_reference_gradients_and_finite_differences
Selected commands run twice debug and twice release; exact receipts and source-to-test mapping retained in wave30 gparse external run directory. Tests do not certify allocation failure, all NaN/overflow inputs, trained-model accuracy or corpus timing. Compiled mutants cover selected boundaries; per-source coverage is recorded in the external map, not presumed.
This page in Markdown ā https://md.groupants.elamur.ai/free/en/mova/gparse/tests/softmax_backward.rs.groupant/bounded-check.groupant/