open-standards/jouleclaw/jouleclaw-rs/crates/jouleclaw-loader-gguf/tests/real_granite_oracle.rs

//! End-to-end dense-Granite (`granite` arch) GGUF decode oracle. Ignored by
//! default — needs the real ~8 GB Q8_0 checkpoint on disk.
//!
//! Unlike the safetensors decoders, this exercises the substrate's first
//! **GGUF-file** forward ([`jouleclaw_loader_gguf::granite::Granite`]). The
//! reference is `llama.cpp`: for the raw prompt "The capital of France is"
//! (token ids `[791, 6864, 315, 9822, 374]`, Granite adds no BOS), greedy
//! decode yields " Paris" (token id 12366) — confirmed with
//! `llama-simple` and `llama-tokenize` on the same file.
//!
//! Run: `cargo test -p jouleclaw-loader-gguf --release --features rayon \
//!   --test real_granite_oracle -- --ignored --nocapture`

use jouleclaw_loader_gguf::bpe_tokenizer::BpeTokenizer;
use jouleclaw_loader_gguf::granite::Granite;
use jouleclaw_loader_gguf::read_gguf_file;

const GG: &str =
    "/Volumes/macos_4TB_external_1/neuronexus_ai/model_library/granite-4.1-8b-gguf/granite-4.1-8b-Q8_0.gguf";

/// "The capital of France is" (raw, no BOS) per `llama-tokenize`.
const PROMPT: [u32; 5] = [791, 6864, 315, 9822, 374];
/// " Paris" per `llama-tokenize`; the llama.cpp greedy next token.
const PARIS: u32 = 12366;

fn argmax(v: &[f32]) -> usize {
    v.iter()
        .enumerate()
        .fold((0usize, f32::NEG_INFINITY), |(bi, bv), (i, &x)| {
            if x > bv {
                (i, x)
            } else {
                (bi, bv)
            }
        })
        .0
}

#[test]
#[ignore]
fn real_granite_4_1_predicts_paris() {
    if !std::path::Path::new(GG).exists() {
        eprintln!("skip: granite-4.1 GGUF not present");
        return;
    }
    let m = Granite::load(GG).expect("load granite gguf");
    eprintln!(
        "granite: {} layers, d={}, heads={}/{}, ffn={}, vocab={}, scales(emb={},attn={},res={},logit={})",
        m.cfg.n_layers, m.cfg.d_model, m.cfg.n_heads, m.cfg.n_kv, m.cfg.ffn, m.cfg.vocab,
        m.cfg.embedding_scale, m.cfg.attention_scale, m.cfg.residual_scale, m.cfg.logit_scale
    );
    let logits = m.forward(&PROMPT);
    let got = argmax(&logits);
    eprintln!("argmax = {got} (expected {PARIS} = ' Paris')");
    assert_eq!(got, PARIS as usize, "greedy next token diverges from llama.cpp reference");
}

/// Greedy continuation should read as a coherent " Paris..." completion, matching
/// llama.cpp's raw greedy ("The capital of France is Paris.").
#[test]
#[ignore]
fn real_granite_4_1_greedy_starts_with_paris() {
    if !std::path::Path::new(GG).exists() {
        eprintln!("skip: granite-4.1 GGUF not present");
        return;
    }
    let m = Granite::load(GG).expect("load granite gguf");
    let out = m.generate(&PROMPT, 4);
    eprintln!("greedy +4 ids = {out:?}");
    assert_eq!(out[0], PARIS, "first generated token must be ' Paris'");
}

/// Full text-in / text-out through the GGUF-embedded tokenizer: the prompt must
/// tokenize to the llama.cpp ids, and greedy decode must read " Paris…".
#[test]
#[ignore]
fn real_granite_4_1_text_roundtrip() {
    if !std::path::Path::new(GG).exists() {
        eprintln!("skip: granite-4.1 GGUF not present");
        return;
    }
    let tok = BpeTokenizer::from_gguf(&read_gguf_file(GG).expect("gguf")).expect("tokenizer");
    let ids = tok.encode("The capital of France is");
    eprintln!("encode = {ids:?}");
    assert_eq!(ids, PROMPT.to_vec(), "tokenization diverges from llama.cpp");

    let m = Granite::load(GG).expect("load");
    let gen_ids = m.generate(&ids, 4);
    let text = tok.decode(&gen_ids);
    eprintln!("generated text = {text:?}");
    assert!(text.starts_with(" Paris"), "decode should start with ' Paris', got {text:?}");
}