Elamur Groupantsgparse-softmax
The groupant beside gparse/src/softmax.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 |
| composition | partial | composition |
| cost | symbolic source derived | invariant |
| descriptant | present | descriptant |
| evidence | executed finite checks | evidence |
| invariant | partial | invariant |
| tests | executed 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 |
gradient_transport | gradient_transport | partial source derived model | gradient_transport |
layout_transport | layout_transport | partial source derived model | layout_transport |
failure_precedence | failure_precedence | partial source derived model | failure_precedence |
numerical_observation | numerical_observation | partial source derived model | numerical_observation |
Model text
descriptant
Source-bound description
last_dim dispatches nonempty CUDA F32 to contiguous custom Softmax; otherwise Candle reference last-dimension softmax. cuda_fwd validates contiguous offsets, row dimension/nonempty, i32 columns/u32 rows, allocates output and launches existing softmax_f32 with retained source offset and grid(rows,1,1), block(1,32,1). bwd returns p*(upstream-sum_last(p*upstream)).
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
last_dim dispatches nonempty CUDA F32 to contiguous custom Softmax; otherwise Candle reference last-dimension softmax. cuda_fwd validates contiguous offsets, row dimension/nonempty, i32 columns/u32 rows, allocates output and launches existing softmax_f32 with retained source offset and grid(rows,1,1), block(1,32,1). bwd returns p*(upstream-sum_last(p*upstream)).
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
Output shape matches input; finite real row relation sums probabilities to one, while numerical agreement is only finite-tested. Backward row-sum zero mathematically; singleton derivative zero. CPU custom kernel rejects; public wrapper routes CPU to reference.
Cost
O(elements) forward and backward plus reductions; contiguous conversion may allocate/copy O(elements). CUDA allocation/launch/device errors propagate; no speed measurement or all-domain error cap.
Allocation, finite arithmetic, platform dimensions and external failures are not universally certified.
composition
Composition
last_dim dispatches nonempty CUDA F32 to contiguous custom Softmax; otherwise Candle reference last-dimension softmax. cuda_fwd validates contiguous offsets, row dimension/nonempty, i32 columns/u32 rows, allocates output and launches existing softmax_f32 with retained source offset and grid(rows,1,1), block(1,32,1). bwd returns p*(upstream-sum_last(p*upstream)).
Scoped children
The byte-scoped children project distinct internal relations through the parent syntax file. The parent combines them in actual source order. Remaining declarations/tests outside child ranges belong to this parent. Their correctness does not establish external caller preconditions.
Interfaces and limits
Output shape matches input; finite real row relation sums probabilities to one, while numerical agreement is only finite-tested. Backward row-sum zero mathematically; singleton derivative zero. CPU custom kernel rejects; public wrapper routes CPU to reference.
For positive normalized row p, compatible finite real logits are log(p)+constant; zero probabilities may be finite underflow/masking in f32 and do not uniquely identify -infinity. Gradient output admits upstream shifts by a row constant.
models
Typed model catalog
behavior_complement
CPU/F64 inputs cannot use the custom CUDA F32 kernel through last_dim. Noncontiguous input is copied before custom invocation; raw custom rejects noncontiguous. CUDA error path does not silently use CPU fallback.
reverse_reconstruction
For positive normalized row p, compatible finite real logits are log(p)+constant; zero probabilities may be finite underflow/masking in f32 and do not uniquely identify -infinity. Gradient output admits upstream shifts by a row constant.
gradient_transport
VJP is diag(p)g-p(pᵀg); unlike no-backward upstream fused op, custom propagates into logits.
layout_transport
contiguous_offsets start..end preserves narrow-view offset; transpose becomes contiguous before launch.
failure_precedence
Raw CUDA path checks contiguous -> last dimension -> nonempty -> integer conversions -> dtype slice -> kernel -> alloc -> launch.
numerical_observation
Selected widths 1,3,17,33,257 test abs forward/gradient/finite-difference error<2e-6; this is not relative percentage or universal guarantee.
tests
Checks and limits
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.
See source execution evidence for exact names and unexecuted paths.
evidence
Execution evidence — 2026-10-31
Source: gparse/src/softmax.rs
- cpu_f64_uses_reference: twice debug, twice release, all four exits 0.
- fused_softmax_matches_reference_gradients_and_finite_differences: twice debug, twice release, all four exits 0.
- synthetic_single_optimizer_step: twice debug, twice release, all four exits 0.
Controls
1 source-local compiled assertion mutants; 2 observed control tests including dependency mutations. Mutant compile exit0, assertion exit101, byte-exact restoration and positive rerun exit0. Per-source controls are listed in external source-test-map.json, not implied for every source.
Limits
Finite synthetic CPU/CUDA observations; no trained-corpus quality or all-input numeric guarantee.
CUDA softmax comparisons use absolute tolerance2e-6. GELU grid maxima observed: values0.0004734993, derivatives0.00086838007, finite difference0.0000005534244635896357; respective strict caps5e-4/1e-3/1e-5. These are finite observations, not percentage profiles or throughput claims.
Receipts: external wave30 gparse run native-commands.json, results.json, controls.json and source-test-map.json. Sorted exact test names/outcomes match per profile; numeric diagnostics and timings need not be byte-identical. Strict scoped library Clippy reports ten pre-existing production lints; appended test blocks were formatted independently without production reformatting.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/gparse/src/softmax.rs.groupant/