frink-models 0.49.0

//! The per-layer sliding-window ARRAY and the NextN/MTP blocks, checked
//! against llama.cpp itself: two seams (`frink_models::swa_layers`,
//! `frink_models::mtp_blocks`) that three refusals named and every
//! real EXAONE-4 32B, EXAONE-MoE, Olmo-3, MiMo-V2, Step-3.5 and Mellum
//! export carries.
//!
//! # The array
//!
//! `{arch}.attention.sliding_window_pattern` is read upstream with
//! `get_key_or_arr`, and that name hides three behaviours decided by
//! which overload each graph's `load_arch_hparams` calls:
//!
//! | mode | arrays are | scalars are | rows |
//! |---|---|---|---|
//! | scalar overload, `required = false` | **IGNORED** (returns false, seed stands; `llama-model-loader.cpp:502-507`) | the period | `exaone4`, `exaone-moe`, `olmo2`, `gemma3`, ... (15) |
//! | array overload | the per-layer truth, length = `n_layer()` at the time, which is `block_count` | broadcast as a BOOL | `mimo2`, `step35`, `gemma4`, `gemma4-assistant`, `dflash` |
//! | scalar then array | the per-layer truth | the period | `mellum`, `cohere2moe` |
//!
//! frink refused the array form for every architecture. Three fixtures
//! evidence the two branches a generic-path row can reach:
//!
//! - `exaone_moe_array`: the base `exaone_moe` fixture plus the array
//!   `conversion/exaone.py:84` writes ([T, T, T, F], the seeded
//!   period). **libllama's logits are byte-identical to the base's.**
//! - `exaone_moe_array_flipped`: the same with the array INVERTED.
//!   **libllama's logits are STILL byte-identical to the base's**,
//!   which is the measurement that upstream never reads it; frink
//!   answers the seeded period for both.
//! - `mellum`: the one generic-path graph that HONOURS the array. Its
//!   array [T, T, F, T] disagrees with the seeded period-4 [T, T, T, F]
//!   on two layers, the window is narrower than the prompt, and the
//!   golden is the file's layout: `load_tensors` reports `is_swa =
//!   1, 1, 0, 1`.
//!
//! # The MTP block
//!
//! `exaone_moe_mtp` is the base fixture with ONE NextN block appended
//! inside `block_count` (5 blocks, `nextn_predict_layers = 1`), the
//! shape of every K-EXAONE export (`exaone.py:132,146`). llama.cpp
//! reports `n_layer = 4, n_layer_all = 5` and "model has unused
//! tensor blk.4.*" fifteen times; its trunk is the same graph as the
//! base with the output head redrawn, and the golden is that. The
//! block's weights are drawn 8x wider than the trunk's, so a loader
//! that ran it as a fifth layer would not be a near miss -- and the
//! two sabotages below show that no config reaches that: counting it
//! as a layer asks it for experts a dense NextN block does not have,
//! and skipping it without saying so leaves fifteen tensors for the
//! consumption gate to name.
//!
//! **Where the numbers come from.** Each golden was produced by running
//! llama.cpp's own graph over its fixture through
//! `scripts/gptoss_reference_logits.cpp` linked against a real
//! `libllama` built from `.scratch/llama.cpp`.
//!
//! Regenerating:
//!
//! ```text
//! PYTHONPATH=$LLAMA/gguf-py python3 scripts/make_exaone_moe_fixture.py \
//!     crates/frink-models/tests/fixtures/exaone_moe_array_tiny.gguf --window-array agree
//! PYTHONPATH=$LLAMA/gguf-py python3 scripts/make_exaone_moe_fixture.py \
//!     crates/frink-models/tests/fixtures/exaone_moe_array_flipped_tiny.gguf --window-array disagree
//! PYTHONPATH=$LLAMA/gguf-py python3 scripts/make_exaone_moe_fixture.py \
//!     crates/frink-models/tests/fixtures/exaone_moe_mtp_tiny.gguf --window-array agree --mtp
//! PYTHONPATH=$LLAMA/gguf-py python3 scripts/make_mellum_fixture.py \
//!     crates/frink-models/tests/fixtures/mellum_tiny.gguf
//! /tmp/ref_logits <fixture> 3 7 11 19 23 5
//! ```

mod common;
use common::{
    assert_all_three_paths_match, graph_caches, graph_fixture_path, kl_vs_golden,
    load_graph_fixture, worst_vs, EXAONE_MOE_GOLDEN, GRAPH_PROMPT,
};
use frink_models::loader::LoadError;
use frink_models::swa_layers::SwaLayers;
use frink_models::{Decoder, ModelConfig};

const EXAONE_MOE_MTP_GOLDEN: [f32; 48] = [
    -0.0399573,
    -0.26254788,
    0.112740405,
    -0.45957953,
    0.20378731,
    0.09908521,
    0.0650342,
    0.03963881,
    0.027113974,
    0.04426959,
    0.2971161,
    0.44927487,
    -0.08063339,
    -0.31896305,
    0.12336944,
    -0.26059663,
    0.23168293,
    -0.07957372,
    -0.041728646,
    -0.06822635,
    -0.012256693,
    -0.06147337,
    0.058683395,
    0.018150419,
    0.34285027,
    -0.17899069,
    0.35574993,
    0.005653359,
    0.21688741,
    0.18249817,
    0.015758194,
    0.19431275,
    -0.5292666,
    0.089496456,
    -0.27419174,
    0.6649489,
    -0.22339234,
    -0.28366813,
    0.23308644,
    0.13166983,
    -0.035332873,
    -0.45804775,
    0.030781724,
    0.18007353,
    0.16690017,
    -0.18101907,
    -0.04980211,
    -0.14753367,
];

const MELLUM_GOLDEN: [f32; 48] = [
    0.060756013,
    -0.023824722,
    0.20276277,
    -0.08673769,
    -0.17752253,
    0.16671993,
    -0.049065586,
    -0.14727482,
    -0.23216516,
    0.018855155,
    -0.15827064,
    -0.13220614,
    0.017678097,
    0.44898403,
    0.18785386,
    -0.060248606,
    0.18946445,
    0.09642051,
    0.009634852,
    -0.22297539,
    -0.1649667,
    0.3061224,
    -0.31932747,
    -0.063845895,
    -0.42996672,
    -0.00028830767,
    0.4683508,
    -0.33105162,
    0.21087095,
    -0.16849121,
    -0.39124924,
    0.124670506,
    0.019993221,
    -0.48215437,
    -0.43157175,
    -0.56464714,
    0.34288388,
    0.11505209,
    0.18613203,
    -0.30828017,
    0.1446941,
    -0.335675,
    -0.043562084,
    -0.14170834,
    0.27802056,
    0.03002052,
    0.1744021,
    -0.067254916,
];

/// Every (fixture, golden) pair this suite audits, for the KL report.
const ROWS: [(&str, &[f32]); 4] = [
    ("exaone_moe_array", &EXAONE_MOE_GOLDEN),
    ("exaone_moe_array_flipped", &EXAONE_MOE_GOLDEN),
    ("exaone_moe_mtp", &EXAONE_MOE_MTP_GOLDEN),
    ("mellum", &MELLUM_GOLDEN),
];

fn prefill(decoder: &Decoder) -> Vec<f32> {
    let mut kv = graph_caches(decoder);
    decoder.forward_batch_last(&GRAPH_PROMPT, 0, &mut kv)
}

// --- the array, ignored where llama.cpp ignores it -------------------

/// The array a real converter writes, on a graph that reads the scalar
/// overload: the same golden as the base fixture, on all three paths.
#[test]
fn exaone_moe_with_the_converter_array_matches_llama_cpp_on_all_three_paths() {
    assert_all_three_paths_match("exaone_moe_array", &EXAONE_MOE_GOLDEN);
    let d = load_graph_fixture("exaone_moe_array");
    assert_eq!(
        d.config.swa_layers,
        SwaLayers::period(4, false),
        "exaone-moe.cpp:6-8 seeds 4 and :7's scalar overload returns false on an array"
    );
}

/// The INVERTED array: llama.cpp's logits do not move, so neither may
/// frink's. This is the over-refusal, lifted, and the proof that
/// lifting it by IGNORING rather than by honouring is llama.cpp's
/// answer and not a guess.
#[test]
fn exaone_moe_with_a_disagreeing_array_is_still_the_seeded_period() {
    assert_all_three_paths_match("exaone_moe_array_flipped", &EXAONE_MOE_GOLDEN);
    let d = load_graph_fixture("exaone_moe_array_flipped");
    assert_eq!(d.config.swa_layers, SwaLayers::period(4, false));
    // The file says [F, F, F, T]; the graph runs [T, T, T, F].
    assert_eq!(d.config.layer_sliding_window(0), Some(3));
    assert_eq!(d.config.layer_sliding_window(3), None);
}

// --- the array, honoured where llama.cpp honours it ------------------

/// `mellum.cpp:12-17`: the array is the layout, and it is NOT the
/// seeded period.
#[test]
fn mellum_matches_llama_cpp_on_all_three_paths() {
    assert_all_three_paths_match("mellum", &MELLUM_GOLDEN);
    let d = load_graph_fixture("mellum");
    assert_eq!(
        d.config.swa_layers,
        SwaLayers::PerLayer(vec![true, true, false, true].into())
    );
    assert_eq!(d.config.sliding_window, Some(3));
    let windowed: Vec<bool> = (0..4)
        .map(|il| d.config.layer_sliding_window(il).is_some())
        .collect();
    assert_eq!(
        windowed,
        [true, true, false, true],
        "libllama's load_tensors reports is_swa = 1, 1, 0, 1"
    );
}

/// Sabotage: keeping the seeded period-4 layout in place of the file's
/// array -- what `exaone-moe` correctly does and `mellum` must not --
/// windows layer 2 and unwindows layer 3, and the logits see it.
#[test]
fn mellum_with_the_seeded_period_instead_of_its_array_diverges_from_llama_cpp() {
    let path = graph_fixture_path("mellum");
    let file = frink_gguf::GgufFile::open(&path).expect("opens");
    let mut config = ModelConfig::from_gguf(&file).expect("parses");
    config.swa_layers = SwaLayers::period(4, false);
    let d = Decoder::from_gguf(&path, config).expect("loads");
    assert_eq!(d.config.layer_sliding_window(2), Some(3), "sabotage landed");
    assert_eq!(d.config.layer_sliding_window(3), None, "sabotage landed");
    let worst = worst_vs(&prefill(&d), &MELLUM_GOLDEN);
    assert!(
        worst > 1e-2,
        "the seeded period must be visible against a 3-token window: {worst}"
    );
}

/// Sabotage: windowing every layer -- the answer a missing array used
/// to fall through to -- diverges too.
#[test]
fn mellum_windowing_every_layer_diverges_from_llama_cpp() {
    let path = graph_fixture_path("mellum");
    let file = frink_gguf::GgufFile::open(&path).expect("opens");
    let mut config = ModelConfig::from_gguf(&file).expect("parses");
    config.swa_layers = SwaLayers::All;
    let d = Decoder::from_gguf(&path, config).expect("loads");
    let worst = worst_vs(&prefill(&d), &MELLUM_GOLDEN);
    assert!(worst > 1e-2, "windowing layer 2 must be visible: {worst}");
}

// --- the MTP block -----------------------------------------------------

/// One NextN block inside `block_count`: the trunk runs, the block does
/// not, and the golden is llama.cpp's over the same file.
#[test]
fn exaone_moe_with_an_mtp_block_matches_llama_cpp_on_all_three_paths() {
    assert_all_three_paths_match("exaone_moe_mtp", &EXAONE_MOE_MTP_GOLDEN);
    let d = load_graph_fixture("exaone_moe_mtp");
    assert_eq!(
        (d.config.n_layers, d.config.n_mtp_blocks),
        (4, 1),
        "llama.cpp: n_layer = 4, n_layer_all = 5"
    );
    assert_eq!(d.layers.len(), 4, "the block is not a layer");
    // The trunk still has its window and its NoPE layer: the block
    // did not shift either rule.
    assert_eq!(d.config.layer_sliding_window(3), None);
    assert!(!d.config.layer_rotates(3));
}

/// Sabotage: a config that counts the block as a fifth layer cannot
/// load at all. A NextN block is DENSE upstream (`exaone-moe.cpp:73`
/// takes the dense branch for `i >= n_layer`) while the trunk's rule
/// makes every layer past the leading one MoE, so the loader asks the
/// block for expert tensors it does not have and stops there -- before
/// the consumption gate would have named the unread `nextn.*` head
/// (the test below shows that gate firing on its own). Running the
/// block is therefore not reachable by any config, which is stronger
/// than "diverges".
#[test]
fn counting_the_mtp_block_as_a_layer_cannot_load() {
    let path = graph_fixture_path("exaone_moe_mtp");
    let file = frink_gguf::GgufFile::open(&path).expect("opens");
    let mut config = ModelConfig::from_gguf(&file).expect("parses");
    config.n_layers = 5;
    config.n_mtp_blocks = 0;
    match Decoder::from_gguf(&path, config) {
        Err(LoadError::Gguf(frink_gguf::GgufError::TensorNotFound(name))) => {
            assert!(
                name.starts_with("blk.4.ffn_") && name.contains("_exps"),
                "the block has no experts: {name}"
            );
        }
        Err(other) => panic!("the block is not a layer, got {other:?}"),
        Ok(_) => panic!("the block is not a layer, yet it loaded as one"),
    }
}

/// Sabotage the other way: a config that skips the block but marks
/// nothing as skipped trips the same gate on all fifteen tensors. The
/// mark and the layer loop take their range from the same two config
/// fields, so this is the shape that CANNOT happen from the loader; it
/// is here to show the gate sees a skipped block that nobody declared.
#[test]
fn an_undeclared_skip_is_refused_by_the_consumption_gate() {
    let path = graph_fixture_path("exaone_moe_mtp");
    let file = frink_gguf::GgufFile::open(&path).expect("opens");
    let mut config = ModelConfig::from_gguf(&file).expect("parses");
    config.n_mtp_blocks = 0;
    match Decoder::from_gguf(&path, config) {
        Err(LoadError::UnconsumedTensors(n, _)) => assert_eq!(n, 15),
        Err(other) => panic!("fifteen unread tensors must refuse, got {other:?}"),
        Ok(_) => panic!("fifteen unread tensors must refuse, yet it loaded"),
    }
}

// --- the number this suite speaks in ------------------------------------

#[test]
fn report_kl_against_llama_cpp() {
    for (name, want) in ROWS {
        let d = load_graph_fixture(name);
        let got = prefill(&d);
        let kl = kl_vs_golden(&got, want);
        let worst = worst_vs(&got, want);
        eprintln!("{name}: KL(llama.cpp || frink) = {kl:.3e} nats, max |delta| = {worst:.3e}");
        assert!(kl < 1e-8, "{name}: KL {kl}");
    }
}