ferrotorch-jit-script 0.5.7

//! Conformance suite for `ferrotorch-jit-script::script` — parity against
//! `torch.jit.script` semantics.
//!
//! Tracking issue: #826 (Conformance Buildout C1 — Tier-1 crates).
//!
//! `torch.jit.script` is a code-transformation annotation: it compiles a
//! Python function body into a TorchScript graph.  ferrotorch's `#[script]`
//! is the Rust analogue — a proc-macro that rewrites a Rust function body so
//! it is captured via `ferrotorch_jit::trace` and returned as a
//! `TracedModule<T>`.
//!
//! Parity contract: calling `forward_multi` on the resulting `TracedModule`
//! with input tensors X must produce the same output that calling the
//! original arithmetic on X would produce.  The contract is structural, not
//! numeric in the PyTorch-fixture sense — we compare against direct
//! arithmetic, not against a serialised PyTorch output tensor.  Reference
//! values are recorded in `tests/conformance/fixtures.json` and were
//! generated by `scripts/regenerate_jit_script_fixtures.py` against
//! torch==2.11.0.
//!
//! Coverage dimensions:
//! * **Correct output** — `forward_multi` returns the right values.
//! * **Scalar-type preservation** — `Tensor<f64>` functions produce
//!   `TracedModule<f64>`, not `TracedModule<f32>` (regression for the
//!   silent-f32-fallback bug).
//! * **Multi-arg functions** — 2-arg and 3-arg scripted functions.
//! * **Module reuse** — the same `TracedModule` can be called multiple times.
//! * **Serialisation round-trip** — `to_bytes` / `from_bytes` preserves
//!   graph semantics.
//! * **Spec-only / compile-error paths** — documented via `cascade_skip`
//!   entries in the fixtures; exercised in `tests/ui/` trybuild tests (see
//!   fixture entries with `"op": "script_error"`).

use std::path::PathBuf;

use ferrotorch_core::grad_fns::arithmetic::{add, mul};
use ferrotorch_core::grad_fns::reduction::sum;
use ferrotorch_core::storage::TensorStorage;
use ferrotorch_core::{FerrotorchResult, Tensor};
use ferrotorch_jit::TracedModule;
use ferrotorch_jit_script::script;
use serde::Deserialize;

// ---------------------------------------------------------------------------
// Fixture loading
// ---------------------------------------------------------------------------

#[derive(Debug, Deserialize)]
struct FixtureFile {
    #[allow(dead_code, reason = "metadata kept for forward-compat and diagnostics")]
    metadata: FixtureMetadata,
    fixtures: Vec<Fixture>,
}

#[derive(Debug, Deserialize)]
struct FixtureMetadata {
    #[allow(dead_code, reason = "used for version drift diagnostics")]
    torch_version: String,
    #[allow(dead_code, reason = "metadata kept for forward-compat")]
    python_executable: String,
    #[allow(dead_code, reason = "metadata kept for forward-compat")]
    python_platform: String,
    #[allow(dead_code, reason = "metadata kept for forward-compat")]
    generated_at: String,
    #[allow(dead_code, reason = "metadata kept for forward-compat")]
    description: String,
}

#[derive(Debug, Deserialize)]
struct Fixture {
    case: String,
    op: String,
    description: String,

    // Runtime fixtures (op == "script")
    #[serde(default)]
    input_a: Option<Vec<f64>>,
    #[serde(default)]
    input_b: Option<Vec<f64>>,
    #[serde(default)]
    input_c: Option<Vec<f64>>,
    #[serde(default)]
    input_w: Option<Vec<f64>>,
    #[serde(default)]
    input_a_first: Option<Vec<f64>>,
    #[serde(default)]
    input_w_first: Option<Vec<f64>>,
    #[serde(default)]
    input_a_second: Option<Vec<f64>>,
    #[serde(default)]
    input_w_second: Option<Vec<f64>>,
    #[serde(default)]
    #[allow(dead_code, reason = "kept for fixture-schema parity and diagnostics")]
    dtype: Option<String>,
    #[serde(default)]
    expected_output: Option<Vec<f64>>,
    #[serde(default)]
    expected_output_first: Option<Vec<f64>>,
    #[serde(default)]
    expected_output_second: Option<Vec<f64>>,

    // Spec-only sentinel
    #[serde(default)]
    cascade_skip: Option<String>,
}

fn load_fixtures() -> FixtureFile {
    let p = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
        .join("tests")
        .join("conformance")
        .join("fixtures.json");
    let bytes = std::fs::read(&p).unwrap_or_else(|e| {
        panic!(
            "read {} failed: {e}. Regenerate via \
             scripts/regenerate_jit_script_fixtures.py",
            p.display()
        )
    });
    serde_json::from_slice(&bytes).unwrap_or_else(|e| panic!("parse {}: {e}", p.display()))
}

fn fixtures_for<'a>(file: &'a FixtureFile, case: &str) -> Option<&'a Fixture> {
    file.fixtures.iter().find(|f| f.case == case)
}

// ---------------------------------------------------------------------------
// Tensor construction helpers (mirror script_macro.rs)
// ---------------------------------------------------------------------------

fn t1d_f32(data: &[f64]) -> Tensor<f32> {
    let f32s: Vec<f32> = data.iter().map(|&x| x as f32).collect();
    let n = f32s.len();
    Tensor::from_storage(TensorStorage::cpu(f32s), vec![n], false).unwrap()
}

fn t1d_f64(data: &[f64]) -> Tensor<f64> {
    let n = data.len();
    Tensor::from_storage(TensorStorage::cpu(data.to_vec()), vec![n], false).unwrap()
}

// ---------------------------------------------------------------------------
// Scripted functions under test
//
// These mirror the functions in tests/script_macro.rs but are defined here
// so the conformance test is self-contained — the conformance suite must be
// able to run independently of the unit-test suite.
// ---------------------------------------------------------------------------

#[script]
fn cs_weighted_sum(a: Tensor<f32>, w: Tensor<f32>) -> FerrotorchResult<Tensor<f32>> {
    let prod = mul(&a, &w)?;
    sum(&prod)
}

#[script]
fn cs_three_arg_add(
    a: Tensor<f32>,
    b: Tensor<f32>,
    c: Tensor<f32>,
) -> FerrotorchResult<Tensor<f32>> {
    let ab = add(&a, &b)?;
    add(&ab, &c)
}

#[script]
fn cs_weighted_sum_f64(a: Tensor<f64>, w: Tensor<f64>) -> FerrotorchResult<Tensor<f64>> {
    let prod = mul(&a, &w)?;
    sum(&prod)
}

// ---------------------------------------------------------------------------
// Helper: assert two f32 slices match within a tight tolerance.
// For these fixtures the expected values are exact-arithmetic results so
// 1e-5 relative tolerance is more than sufficient.
// ---------------------------------------------------------------------------

fn assert_close_f32(actual: &[f32], expected: &[f64], label: &str) {
    assert_eq!(
        actual.len(),
        expected.len(),
        "{label}: length mismatch (actual={}, expected={})",
        actual.len(),
        expected.len()
    );
    const TOL: f32 = 1e-5;
    for (i, (&a, &e)) in actual.iter().zip(expected.iter()).enumerate() {
        let e_f32 = e as f32;
        let diff = (a - e_f32).abs();
        let scale = e_f32.abs().max(1.0);
        assert!(
            diff <= TOL * scale,
            "{label}: index {i}: actual={a} expected={e_f32} diff={diff:.3e} tol={:.3e}",
            TOL * scale
        );
    }
}

fn assert_close_f64(actual: &[f64], expected: &[f64], label: &str) {
    assert_eq!(
        actual.len(),
        expected.len(),
        "{label}: length mismatch (actual={}, expected={})",
        actual.len(),
        expected.len()
    );
    const TOL: f64 = 1e-10;
    for (i, (&a, &e)) in actual.iter().zip(expected.iter()).enumerate() {
        let diff = (a - e).abs();
        let scale = e.abs().max(1.0);
        assert!(
            diff <= TOL * scale,
            "{label}: index {i}: actual={a} expected={e} diff={diff:.3e}",
        );
    }
}

// ---------------------------------------------------------------------------
// Sanity: fixture file has all expected cases
// ---------------------------------------------------------------------------

#[test]
fn fixture_file_covers_every_case() {
    let file = load_fixtures();
    let required = [
        "two_arg_weighted_sum_f32",
        "three_arg_add_f32",
        "module_save_load_roundtrip_f32",
        "scalar_type_preservation_f64",
        "module_reuse_f32",
        "unrecognized_return_type_emits_compile_error",
        "missing_return_type_emits_compile_error",
    ];
    for case in required {
        assert!(
            fixtures_for(&file, case).is_some(),
            "fixture file missing case {case:?} — regenerate via \
             scripts/regenerate_jit_script_fixtures.py"
        );
    }
}

// ---------------------------------------------------------------------------
// Layer 3 — Conformance tests
// ---------------------------------------------------------------------------

/// `script` — two-arg weighted sum, f32.
/// PyTorch reference: `(a * w).sum()`.
#[test]
fn script_two_arg_weighted_sum_f32() {
    let file = load_fixtures();
    let fx =
        fixtures_for(&file, "two_arg_weighted_sum_f32").expect("fixture two_arg_weighted_sum_f32");
    assert_eq!(fx.op, "script");
    assert!(fx.cascade_skip.is_none(), "unexpected cascade_skip");

    let input_a = fx.input_a.as_deref().expect("input_a");
    let input_w = fx.input_w.as_deref().expect("input_w");
    let expected = fx.expected_output.as_deref().expect("expected_output");

    let module: TracedModule<f32> = cs_weighted_sum(t1d_f32(input_a), t1d_f32(input_w)).unwrap();
    let result = module
        .forward_multi(&[t1d_f32(input_a), t1d_f32(input_w)])
        .unwrap();
    let actual = result.data().expect("read data");
    assert_close_f32(actual, expected, "script two_arg_weighted_sum_f32");
}

/// `script` — three-argument addition, f32.
/// PyTorch reference: `a + b + c`.
#[test]
fn script_three_arg_add_f32() {
    let file = load_fixtures();
    let fx = fixtures_for(&file, "three_arg_add_f32").expect("fixture three_arg_add_f32");
    assert_eq!(fx.op, "script");
    assert!(fx.cascade_skip.is_none(), "unexpected cascade_skip");

    let input_a = fx.input_a.as_deref().expect("input_a");
    let input_b = fx.input_b.as_deref().expect("input_b");
    let input_c = fx.input_c.as_deref().expect("input_c");
    let expected = fx.expected_output.as_deref().expect("expected_output");

    let module: TracedModule<f32> =
        cs_three_arg_add(t1d_f32(input_a), t1d_f32(input_b), t1d_f32(input_c)).unwrap();
    let result = module
        .forward_multi(&[t1d_f32(input_a), t1d_f32(input_b), t1d_f32(input_c)])
        .unwrap();
    let actual = result.data().expect("read data");
    assert_close_f32(actual, expected, "script three_arg_add_f32");
}

/// `script` — serialisation round-trip.
/// `to_bytes` / `from_bytes` must produce a module that re-executes correctly.
#[test]
fn script_module_save_load_roundtrip_f32() {
    let file = load_fixtures();
    let fx = fixtures_for(&file, "module_save_load_roundtrip_f32")
        .expect("fixture module_save_load_roundtrip_f32");
    assert_eq!(fx.op, "script");
    assert!(fx.cascade_skip.is_none(), "unexpected cascade_skip");

    let input_a = fx.input_a.as_deref().expect("input_a");
    let input_w = fx.input_w.as_deref().expect("input_w");
    let expected = fx.expected_output.as_deref().expect("expected_output");

    let module = cs_weighted_sum(t1d_f32(input_a), t1d_f32(input_w)).unwrap();
    let bytes = module.to_bytes();
    let loaded: TracedModule<f32> = TracedModule::<f32>::from_bytes(&bytes).unwrap();
    let result = loaded
        .forward_multi(&[t1d_f32(input_a), t1d_f32(input_w)])
        .unwrap();
    let actual = result.data().expect("read data");
    assert_close_f32(actual, expected, "script module_save_load_roundtrip_f32");
}

/// `script` — scalar-type preservation, f64.
///
/// This is the regression guard for the silent-f32-fallback bug: prior to
/// the fix, `extract_tensor_param` returned `None` for `Tensor<f64>` and the
/// macro silently substituted `f32`, wrapping the function into a
/// `TracedModule<f32>`.  The fix emits a `compile_error!` for unrecognised
/// return types and recognises `Tensor<f64>` correctly.
///
/// The type ascription `TracedModule<f64>` on the `module` binding is the
/// load-bearing assertion — if the macro produced `TracedModule<f32>` this
/// would fail to compile.
#[test]
fn script_scalar_type_preservation_f64() {
    let file = load_fixtures();
    let fx = fixtures_for(&file, "scalar_type_preservation_f64")
        .expect("fixture scalar_type_preservation_f64");
    assert_eq!(fx.op, "script");
    assert!(fx.cascade_skip.is_none(), "unexpected cascade_skip");

    let input_a = fx.input_a.as_deref().expect("input_a");
    let input_w = fx.input_w.as_deref().expect("input_w");
    let expected = fx.expected_output.as_deref().expect("expected_output");

    // Type ascription is the load-bearing assertion.
    let module: TracedModule<f64> =
        cs_weighted_sum_f64(t1d_f64(input_a), t1d_f64(input_w)).unwrap();
    let result = module
        .forward_multi(&[t1d_f64(input_a), t1d_f64(input_w)])
        .unwrap();
    let actual = result.data().expect("read data");
    assert_close_f64(actual, expected, "script scalar_type_preservation_f64");
}

/// `script` — module reuse across multiple `forward_multi` calls.
/// The same `TracedModule` must return different outputs for different inputs.
#[test]
fn script_module_reuse_f32() {
    let file = load_fixtures();
    let fx = fixtures_for(&file, "module_reuse_f32").expect("fixture module_reuse_f32");
    assert_eq!(fx.op, "script");
    assert!(fx.cascade_skip.is_none(), "unexpected cascade_skip");

    let a_first = fx.input_a_first.as_deref().expect("input_a_first");
    let w_first = fx.input_w_first.as_deref().expect("input_w_first");
    let a_second = fx.input_a_second.as_deref().expect("input_a_second");
    let w_second = fx.input_w_second.as_deref().expect("input_w_second");
    let exp_first = fx
        .expected_output_first
        .as_deref()
        .expect("expected_output_first");
    let exp_second = fx
        .expected_output_second
        .as_deref()
        .expect("expected_output_second");

    // Build once using the first inputs.
    let module: TracedModule<f32> = cs_weighted_sum(t1d_f32(a_first), t1d_f32(w_first)).unwrap();

    // First call.
    let r1 = module
        .forward_multi(&[t1d_f32(a_first), t1d_f32(w_first)])
        .unwrap();
    assert_close_f32(
        r1.data().expect("r1 data"),
        exp_first,
        "module_reuse first call",
    );

    // Second call with different inputs — same module.
    let r2 = module
        .forward_multi(&[t1d_f32(a_second), t1d_f32(w_second)])
        .unwrap();
    assert_close_f32(
        r2.data().expect("r2 data"),
        exp_second,
        "module_reuse second call",
    );
}

/// `script_error` cases — spec-only, cascade-skipped.
/// These are the compile-error paths (unrecognised return type, missing
/// return type). They have no runtime fixture; the test verifies that the
/// fixture entry is present and correctly marks itself for cascade-skip.
/// The actual compile-error assertions live in trybuild UI tests.
#[test]
fn script_error_cases_are_cascade_skipped_in_fixtures() {
    let file = load_fixtures();
    let spec_only_cases = [
        "unrecognized_return_type_emits_compile_error",
        "missing_return_type_emits_compile_error",
    ];

    // Lock the case-list cardinality. If a new `script_error` fixture is
    // added without being explicitly listed here, the count of fixtures
    // with op=="script_error" will exceed `spec_only_cases.len()` and
    // the assertion below will fail — forcing a deliberate test update
    // rather than silent drift.
    assert_eq!(
        spec_only_cases.len(),
        2,
        "spec_only_cases list must enumerate every script_error fixture; \
         update both the list and this count together"
    );
    let fixture_script_error_count = file
        .fixtures
        .iter()
        .filter(|f| f.op == "script_error")
        .count();
    assert_eq!(
        fixture_script_error_count,
        spec_only_cases.len(),
        "fixture file has {fixture_script_error_count} entries with op=script_error \
         but the test enumerates {}; add the new case_name to spec_only_cases",
        spec_only_cases.len()
    );

    let mut matched_count = 0_usize;
    for case_name in spec_only_cases {
        let fx = fixtures_for(&file, case_name)
            .unwrap_or_else(|| panic!("missing fixture for spec-only case {case_name:?}"));
        assert_eq!(
            fx.op, "script_error",
            "{case_name}: expected op=script_error"
        );
        let skip = fx.cascade_skip.as_deref().unwrap_or_else(|| {
            panic!(
                "{case_name}: spec-only fixture must have cascade_skip set; \
                 set cascade_skip = \"spec-only marker, no PyTorch reference\""
            )
        });
        assert!(
            skip.contains("spec-only"),
            "{case_name}: cascade_skip must contain 'spec-only', got: {skip:?}"
        );

        // Tighten the metadata contract: the description must explicitly
        // reference `compile_error` so the fixture can never silently
        // drift to a runtime fixture under the same case name. A stub
        // fixture with `description: ""` or unrelated text would now
        // fail rather than be quietly accepted.
        assert!(
            fx.description.contains("compile_error"),
            "{case_name}: description must mention 'compile_error', got: {:?}",
            fx.description
        );

        matched_count += 1;
    }

    // Belt-and-suspenders: every entry in spec_only_cases must have been
    // matched (loop iteration count == expected count). Catches a
    // duplicate-case-name typo in the list that `spec_only_cases.len()`
    // alone would miss.
    assert_eq!(
        matched_count,
        spec_only_cases.len(),
        "matched only {matched_count} of {} spec-only cases",
        spec_only_cases.len()
    );
}