aprender-serve 0.64.0

//! T-COV-95 Synthetic Falsification: gguf/loader.rs deep coverage
//!
//! Uses GGUFBuilder to exercise loader code paths with Pygmy Models.

use crate::gguf::test_factory::{
    build_minimal_llama_gguf, build_minimal_phi2_gguf, create_f16_data, create_q2_k_data,
    create_q4_0_data, create_q4_1_data, create_q4_k_data, create_q5_0_data, create_q5_1_data,
    create_q5_k_data, create_q6_k_data, create_q8_0_data, GGUFBuilder,
};
use crate::gguf::{GGUFModel, GGUFTransformer};

// ============================================================================
// GGUFModel::from_bytes coverage
// ============================================================================

#[test]
fn test_from_bytes_minimal_empty() {
    let data = GGUFBuilder::new().build();
    let model = GGUFModel::from_bytes(&data);
    assert!(model.is_ok());
    let model = model.expect("test value should be present");
    assert_eq!(model.tensors.len(), 0);
    assert_eq!(model.metadata.len(), 0);
}

#[test]
fn test_from_bytes_with_metadata_only() {
    let data = GGUFBuilder::new()
        .architecture("test")
        .add_u32("test.value", 42)
        .add_f32("test.float", 3.14)
        .add_string("test.string", "hello")
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert!(model.metadata.len() >= 4);
}

#[test]
fn test_from_bytes_llama_arch() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.architecture(), Some("llama"));
}

#[test]
fn test_from_bytes_phi2_arch() {
    let data = build_minimal_phi2_gguf(32, 64, 128, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.architecture(), Some("phi2"));
}

#[test]
fn test_from_bytes_with_q4_0_tensors() {
    let q4_data = create_q4_0_data(1024);
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_q4_0_tensor("test.weight", &[32, 32], &q4_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 2); // Q4_0
}

#[test]
fn test_from_bytes_with_q8_0_tensors() {
    let q8_data = create_q8_0_data(1024);
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_q8_0_tensor("test.weight", &[32, 32], &q8_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 8); // Q8_0
}

#[test]
fn test_from_bytes_with_q4_k_tensors() {
    let q4k_data = create_q4_k_data(256);
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_q4_k_tensor("test.weight", &[16, 16], &q4k_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 12); // Q4_K
}

#[test]
fn test_from_bytes_with_q5_k_tensors() {
    let q5k_data = create_q5_k_data(256);
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_q5_k_tensor("test.weight", &[16, 16], &q5k_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 13); // Q5_K
}

#[test]
fn test_from_bytes_with_q6_k_tensors() {
    let q6k_data = create_q6_k_data(256);
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_q6_k_tensor("test.weight", &[16, 16], &q6k_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 14); // Q6_K
}

#[test]
fn test_from_bytes_with_f32_tensors() {
    let f32_data: Vec<f32> = (0..1024).map(|i| i as f32 * 0.001).collect();
    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_f32_tensor("test.weight", &[32, 32], &f32_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 1);
    assert_eq!(model.tensors[0].qtype, 0); // F32
}

// ============================================================================
// Metadata accessor coverage
// ============================================================================

#[test]
fn test_metadata_accessors_llama() {
    let data = build_minimal_llama_gguf(100, 128, 256, 8, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    assert_eq!(model.architecture(), Some("llama"));
    assert_eq!(model.embedding_dim(), Some(128));
    assert_eq!(model.num_layers(), Some(1));
    assert_eq!(model.num_heads(), Some(8));
    assert_eq!(model.num_kv_heads(), Some(4));
    assert_eq!(model.context_length(), Some(256));
}

#[test]
fn test_metadata_accessors_phi2() {
    let data = build_minimal_phi2_gguf(100, 128, 256, 8);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    assert_eq!(model.architecture(), Some("phi2"));
    assert_eq!(model.embedding_dim(), Some(128));
    assert_eq!(model.num_heads(), Some(8));
}

#[test]
fn test_rope_freq_base_accessor() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .rope_freq_base("llama", 10000.0)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let rope_base = model.rope_freq_base();
    assert!(rope_base.is_some());
    assert!((rope_base.expect("test value should be present") - 10000.0).abs() < 0.01);
}

#[test]
fn test_rms_epsilon_accessor() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .rms_epsilon("llama", 1e-5)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let eps = model.rms_epsilon();
    assert!(eps.is_some());
    assert!((eps.expect("test value should be present") - 1e-5).abs() < 1e-10);
}

#[test]
fn test_ffn_hidden_dim_accessor() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .ffn_hidden_dim("llama", 512)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    // FFN hidden dim is accessed via metadata
    let val = model.metadata.get("llama.feed_forward_length");
    assert!(val.is_some());
}

// ============================================================================
// GGUFTransformer::from_gguf coverage
// ============================================================================

#[test]
fn test_transformer_from_llama_pygmy() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let transformer = GGUFTransformer::from_gguf(&model, &data);
    assert!(
        transformer.is_ok(),
        "Transformer load failed: {:?}",
        transformer.err()
    );

    let transformer = transformer.expect("test value should be present");
    assert_eq!(transformer.config.num_layers, 1);
    assert_eq!(transformer.config.hidden_dim, 64);
}

#[test]
fn test_transformer_from_phi2_pygmy() {
    let data = build_minimal_phi2_gguf(32, 64, 128, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let transformer = GGUFTransformer::from_gguf(&model, &data);
    // May fail if Phi2 QKV fusion isn't handled - that's OK, we exercise the path
    let _ = transformer;
}

#[test]
fn test_transformer_token_embedding_size() {
    let vocab = 64;
    let hidden = 32;
    let data = build_minimal_llama_gguf(vocab, hidden, 64, 2, 2);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let transformer =
        GGUFTransformer::from_gguf(&model, &data).expect("test value should be present");
    // Token embedding should be vocab_size * hidden_dim
    assert_eq!(transformer.token_embedding.len(), vocab * hidden);
}

// ============================================================================
// get_tensor_f32 coverage (dequantization paths)
// ============================================================================

#[test]
fn test_get_tensor_f32_from_f32() {
    let f32_data: Vec<f32> = (0..64).map(|i| i as f32 * 0.1).collect();
    let data = GGUFBuilder::new()
        .add_f32_tensor("test.weight", &[8, 8], &f32_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let tensor = model.get_tensor_f32("test.weight", &data);
    assert!(tensor.is_ok());
    let tensor = tensor.expect("test value should be present");
    assert_eq!(tensor.len(), 64);
}

#[test]
fn test_get_tensor_f32_from_q4_0() {
    let q4_data = create_q4_0_data(1024);
    let data = GGUFBuilder::new()
        .add_q4_0_tensor("test.weight", &[32, 32], &q4_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let tensor = model.get_tensor_f32("test.weight", &data);
    // Dequantization should work
    assert!(tensor.is_ok());
    let tensor = tensor.expect("test value should be present");
    assert_eq!(tensor.len(), 1024);
}

#[test]
fn test_get_tensor_f32_from_q8_0() {
    let q8_data = create_q8_0_data(1024);
    let data = GGUFBuilder::new()
        .add_q8_0_tensor("test.weight", &[32, 32], &q8_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let tensor = model.get_tensor_f32("test.weight", &data);
    assert!(tensor.is_ok());
}

#[test]
fn test_get_tensor_f32_from_q4_k() {
    let q4k_data = create_q4_k_data(256);
    let data = GGUFBuilder::new()
        .add_q4_k_tensor("test.weight", &[16, 16], &q4k_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let tensor = model.get_tensor_f32("test.weight", &data);
    assert!(tensor.is_ok());
}

#[test]
fn test_get_tensor_f32_nonexistent() {
    let data = GGUFBuilder::new().architecture("llama").build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let tensor = model.get_tensor_f32("nonexistent.tensor", &data);
    assert!(tensor.is_err());
}

// ============================================================================
// Tokenizer coverage (encode/decode)
// ============================================================================

#[test]
fn test_decode_empty_tokens() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let decoded = model.decode(&[]);
    assert_eq!(decoded, "");
}

#[test]
fn test_decode_single_token() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    // Token 0 - may decode to something or empty depending on vocab
    let decoded = model.decode(&[0]);
    // Just check it doesn't panic
    let _ = decoded;
}

#[test]
fn test_decode_multiple_tokens() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let decoded = model.decode(&[0, 1, 2, 3, 4]);
    let _ = decoded;
}

#[test]
fn test_decode_out_of_range_token() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    // Token 99999 is out of range for vocab_size=32
    let decoded = model.decode(&[99999]);
    // Should handle gracefully
    let _ = decoded;
}

#[test]
fn test_encode_empty_text() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let encoded = model.encode("");
    // May return None or Some([]) depending on impl
    let _ = encoded;
}

#[test]
fn test_encode_simple_text() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let encoded = model.encode("hello");
    // May return None if no vocab, or Some(tokens)
    let _ = encoded;
}

// ============================================================================
// BOS/EOS token coverage
// ============================================================================

#[test]
fn test_bos_token_id() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let bos = model.bos_token_id();
    // May or may not be set depending on metadata
    let _ = bos;
}

#[test]
fn test_eos_token_id() {
    let data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");

    let eos = model.eos_token_id();
    let _ = eos;
}

// ============================================================================
// Error path coverage
// ============================================================================

#[test]
fn test_from_bytes_too_small() {
    let data = vec![0u8; 4]; // Way too small
    let result = GGUFModel::from_bytes(&data);
    assert!(result.is_err());
}

#[test]
fn test_from_bytes_invalid_magic() {
    let mut data = vec![0u8; 100];
    // Invalid magic
    data[0..4].copy_from_slice(b"XXXX");
    let result = GGUFModel::from_bytes(&data);
    assert!(result.is_err());
}

#[test]
fn test_from_bytes_truncated_header() {
    // Valid magic but truncated
    let mut data = vec![0u8; 16];
    data[0..4].copy_from_slice(&0x46554747u32.to_le_bytes()); // GGUF magic
    let result = GGUFModel::from_bytes(&data);
    assert!(result.is_err());
}

// ============================================================================
// Mixed tensor types
// ============================================================================

#[test]
fn test_model_with_mixed_tensor_types() {
    let f32_data: Vec<f32> = (0..64).map(|i| i as f32 * 0.1).collect();
    let q4_data = create_q4_0_data(1024);
    let q4k_data = create_q4_k_data(256);

    let data = GGUFBuilder::new()
        .architecture("llama")
        .add_f32_tensor("f32_tensor", &[8, 8], &f32_data)
        .add_q4_0_tensor("q4_0_tensor", &[32, 32], &q4_data)
        .add_q4_k_tensor("q4_k_tensor", &[16, 16], &q4k_data)
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 3);

    // Verify we can dequant each
    let _ = model.get_tensor_f32("f32_tensor", &data);
    let _ = model.get_tensor_f32("q4_0_tensor", &data);
    let _ = model.get_tensor_f32("q4_k_tensor", &data);
}

include!("vocabulary_accessor.rs");