📐gparse-model

The groupant beside gparse/src/model.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

Components

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.

Parts

• gparse-model-configuration
• gparse-model-attention
• gparse-model-heads

Model text

descriptant

Source-bound description

Cfg loads defaults, valid numeric replacements and strict erf/tanh GELU; Net::new applies GP_GELU then initializes embedding/transformer/head parameters. forward_impl sums word/pre/suf/shape/position/prag embeddings; pre-LN blocks use softmax(QKᵀ/sqrt(dh)+keybias)V and residual feedforward. Arc scores dep_i W head_j+head_j u+keybias+role pull. rels gathers flat word/head states and sums bilinear and concatenated-linear logits; tags gather word states.

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

Cfg loads defaults, valid numeric replacements and strict erf/tanh GELU; Net::new applies GP_GELU then initializes embedding/transformer/head parameters. forward_impl sums word/pre/suf/shape/position/prag embeddings; pre-LN blocks use softmax(QKᵀ/sqrt(dh)+keybias)V and residual feedforward. Arc scores dep_i W head_j+head_j u+keybias+role pull. rels gathers flat word/head states and sums bilinear and concatenated-linear logits; tags gather word states.

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

LayerNorm uses mean, variance and epsilon 1e-5 through differentiable ops. Training fused route calls custom differentiable softmax; fused inference uses upstream no-backward operation. heads>0, d divisible heads, valid batch dimensions and embedding indices are assumed, not cfg-validated. Padding uses finite -1e9 bias, not mathematical -infinity.

Cost

Let B,T,d,L,H,f,a,r,N,X,U,K denote batch, padded length, encoder width, layers, heads, FF width, arc width, relation width, real rows and XPOS/UPOS/relation class counts. Forward encoder costs O(BTd + L*(BTdd+BTTd+BTdf)); arc projection/scoring O(BTda+BTaa+BTTa); relation gather/projections/classifier O(Ndr+NrrK+NrK), tags O(Nd*(X+U)). Memory includes BHTT attention cells, BTd encoder states and retained training graphs plus parameters/optimizer state. Embedding/final-LN BTd remains when L=0; two AFFIXd tables remain parameter allocations. Dimension products and backend costs are not measured timing claims.

Allocation, finite arithmetic, platform dimensions and external failures are not universally certified.

composition

Composition

Cfg loads defaults, valid numeric replacements and strict erf/tanh GELU; Net::new applies GP_GELU then initializes embedding/transformer/head parameters. forward_impl sums word/pre/suf/shape/position/prag embeddings; pre-LN blocks use softmax(QKᵀ/sqrt(dh)+keybias)V and residual feedforward. Arc scores dep_i W head_j+head_j u+keybias+role pull. rels gathers flat word/head states and sums bilinear and concatenated-linear logits; tags gather word states.

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

LayerNorm uses mean, variance and epsilon 1e-5 through differentiable ops. Training fused route calls custom differentiable softmax; fused inference uses upstream no-backward operation. heads>0, d divisible heads, valid batch dimensions and embedding indices are assumed, not cfg-validated. Padding uses finite -1e9 bias, not mathematical -infinity.

Cfg::save identifies fields but input text is many-to-one due unknown keys, invalid numeric values and repeats. Network outputs admit parameter symmetries and many input feature collisions; invertible reconstruction is not claimed.

models

Typed model catalog

behavior_complement

Training never selects the upstream no-backward softmax via public forward; train=false ignores GP_TRAIN_FUSED_SOFTMAX. Unknown GP_GELU fails before parameter allocation. Invalid numeric cfg strings preserve previous values rather than report error.

reverse_reconstruction

Cfg::save identifies fields but input text is many-to-one due unknown keys, invalid numeric values and repeats. Network outputs admit parameter symmetries and many input feature collisions; invertible reconstruction is not claimed.

gradient_transport

Custom layer norm and training softmax retain Q/K/embedding gradients; inference fusion intentionally excludes Q/K gradient while value route survives.

configuration_precedence

Defaults then ordered lines then GP_GELU override; invalid saved GELU aborts immediately even if a later line is valid.

tensor_coordinates

QKV reshape B,T,3,heads,d/heads; transpose to B,heads,T,dh. Relations require flattened heads k*T+h.

finite_arithmetic

Variance epsilon avoids zero denominator for finite constant rows; nonfinite weights, overflow and all-masked behavior have no universal finite guarantee.

tests

Checks and limits

model::tests::embeddings_get_gradients; model::fused::which_ops_backprop; activation_configuration_roundtrips; CUDA gradient/GELU/synthetic optimizer checks; CPU prediction

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/model.rs

• model::tests::embeddings_get_gradients: twice debug, twice release, all four exits 0.
• model::fused::which_ops_backprop: twice debug, twice release, all four exits 0.
• activation_configuration_roundtrips: twice debug, twice release, all four exits 0.
• cuda_values_and_derivatives: twice debug, twice release, all four exits 0.
• training_attention_keeps_gradients_and_inference_control_breaks_them: twice debug, twice release, all four exits 0.
• synthetic_cpu_prediction_preserves_slots_and_fallback: 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; 3 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/model.rs.groupant/