Pin user-supplied floats to canonical dtype at every API boundary

benchmarks/bench_aca_baseline.py

@@ -47,14 +47,30 @@
 
 
 def _build() -> tuple[object, object, object]:
-    """Build the aca-baseline model, params, and initial conditions."""
+    """Build the aca-baseline model, params, and initial conditions.
+
+    aca_model and lcm imports are deferred to the function body — ASV's
+    forkserver runs `preimport` to discover benchmarks across every
+    `bench_*.py` module before forking workers. Importing JAX at module
+    top loads the multithreaded XLA backend into the forkserver; every
+    subsequent `os.fork()` inherits a corrupted CUDA context and the
+    first device op in the worker aborts with
+    `CUDA_ERROR_NOT_INITIALIZED`. Per-call imports keep JAX out of the
+    forkserver and confine it to the worker process.
+    """
+    from aca_model.agent.preferences import BenchmarkPrefType
     from aca_model.benchmark import (
         create_benchmark_model,
         get_benchmark_initial_conditions,
         get_benchmark_params,
     )
 
-    model = create_benchmark_model(n_subjects=_N_SUBJECTS)
+    from lcm import DiscreteGrid
+
+    model = create_benchmark_model(
+        n_subjects=_N_SUBJECTS,
+        pref_type_grid=DiscreteGrid(BenchmarkPrefType),
+    )
     _, model_params = get_benchmark_params(model=model)
     initial_conditions = get_benchmark_initial_conditions(
         model=model, n_subjects=_N_SUBJECTS, seed=0

pyproject.toml

@@ -98,7 +98,7 @@ tests-cuda13 = { features = [ "tests", "cuda13" ], solve-group = "cuda13" }
 tests-metal = { features = [ "tests", "metal" ], solve-group = "metal" }
 type-checking = { features = [ "type-checking", "tests" ], solve-group = "default" }
 [tool.pixi.feature.benchmarks.pypi-dependencies]
-aca-model = { git = "https://github.com/OpenSourceEconomics/aca-model.git", rev = "f09b5e34102ff42f739b95be5a9d388795b734a1" }
+aca-model = { git = "https://github.com/OpenSourceEconomics/aca-model.git", rev = "9ac20430f499a8b1cdb056af85bc2a26e850bad2" }
 [tool.pixi.feature.cuda12]
 platforms = [ "linux-64" ]
 system-requirements = { cuda = "12" }
@@ -242,6 +242,15 @@ per-file-ignores."tests/*" = [
   "S301",    # Use of pickle
   "SLF001",  # Private member access
 ]
+per-file-ignores."tests/test_dtypes.py" = [
+  "ARG001", # Unused function argument (x64_enabled / x64_disabled fixtures)
+]
+per-file-ignores."tests/test_explicit_dtype_filter.py" = [
+  "ARG001", # Unused function argument (x64_disabled fixture)
+]
+per-file-ignores."tests/test_float_dtype_invariants.py" = [
+  "ARG001", # Unused function argument (x64_disabled fixture)
+]
 per-file-ignores."tests/test_next_state.py" = [
   "ARG001", # Unused function argument
   "ARG005", # Unused lambda argument
@@ -294,7 +303,15 @@ ini_options.addopts = [
   "--dist",
   "loadfile",
 ]
-ini_options.filterwarnings = []
+ini_options.filterwarnings = [
+  # JAX emits this UserWarning when user code asks for a dtype wider
+  # than the active x64 setting allows. Under `--precision=32` it
+  # surfaces every stray `jnp.int64` / `jnp.float64` / `dtype=int64`
+  # literal in src/ — the only files that legitimately trigger it are
+  # the dtype-invariant test modules, which opt out via a local
+  # `pytestmark` filter.
+  "error:Explicitly requested dtype.*:UserWarning",
+]
 ini_options.markers = [
   "illustrative: Tests are designed for illustrative purposes",
   "gpu: Tests that require a GPU (skipped on CPU-only machines)",

src/lcm/ages.py

@@ -10,7 +10,7 @@
 import jax.numpy as jnp
 
 from lcm.exceptions import GridInitializationError, format_messages
-from lcm.typing import Age, Float1D, Int1D
+from lcm.typing import Float1D, Int1D
 
 STEP_UNITS: MappingProxyType[str, Fraction] = MappingProxyType(
     {
@@ -129,7 +129,7 @@ def exact_step_size(self) -> int | Fraction | None:
         """
         return self._exact_step_size
 
-    def period_to_age(self, period: int) -> Age:
+    def period_to_age(self, period: int) -> int | float:
         """Convert a period index to the corresponding age.
 
         Args:
@@ -151,7 +151,7 @@ def period_to_age(self, period: int) -> Age:
             return int(self._values[period])
         return float(self._values[period])
 
-    def age_to_period(self, age: Age) -> int:
+    def age_to_period(self, age: float) -> int:
         """Convert an age to the corresponding period index.
 
         Args:
@@ -172,12 +172,14 @@ def age_to_period(self, age: Age) -> int:
             raise ValueError(msg) from None
 
     @functools.cached_property
-    def _age_to_period_map(self) -> dict[Age, int]:
+    def _age_to_period_map(self) -> dict[int | float, int]:
         if self._is_integer:
             return {int(v): i for i, v in enumerate(self._exact_values)}
         return {float(v): i for i, v in enumerate(self._exact_values)}
 
-    def get_periods_where(self, predicate: Callable[[Age], bool]) -> tuple[int, ...]:
+    def get_periods_where(
+        self, predicate: Callable[[int | float], bool]
+    ) -> tuple[int, ...]:
         """Get period indices where predicate is True.
 
         Args:
@@ -187,7 +189,7 @@ def get_periods_where(self, predicate: Callable[[Age], bool]) -> tuple[int, ...]
             Tuple of period indices where predicate(age) is True.
 
         """
-        _convert: Callable[[object], Age] = int if self._is_integer else float  # ty: ignore[invalid-assignment]
+        _convert: Callable[[object], int | float] = int if self._is_integer else float  # ty: ignore[invalid-assignment]
         return tuple(
             period
             for period in range(self.n_periods)

src/lcm/dtypes.py

@@ -3,22 +3,32 @@
 Used at every API boundary that accepts user data (params, initial
 conditions, regime-id arrays) — always called from Python, never inside
 JIT. Each helper validates that the value fits the target dtype and
-raises a clearly-named error if not.
-
-Casts further down the simulate stack (e.g. transition outputs landing
-in the state pool) use plain `.astype` and rely on the boundary cast
-above them having already pinned the canonical dtype.
+raises a clearly-named error if not. Once an input has crossed the
+boundary it carries the canonical dtype unchanged through the simulate
+stack; downstream code does not re-cast.
 """
 
+import jax
 import jax.numpy as jnp
 import numpy as np
 from jax import Array
 
 _INT32_MIN = int(np.iinfo(np.int32).min)
 _INT32_MAX = int(np.iinfo(np.int32).max)
+_FLOAT32_MAX = float(np.finfo(np.float32).max)
+
+
+def canonical_float_dtype() -> jnp.dtype:
+    """Return pylcm's canonical float dtype, derived from `jax_enable_x64`.
+
+    Returns `jnp.float64` if `jax.config.jax_enable_x64` is True,
+    otherwise `jnp.float32`. The value is read at call time, not at
+    import, so toggling the JAX config (e.g. between tests) is honoured.
+    """
+    return jnp.float64 if jax.config.read("jax_enable_x64") else jnp.float32
 
 
-def safe_to_int32(value: object, *, name: str) -> Array:
+def safe_to_int_dtype(value: object, *, name: str) -> Array:
     """Cast a scalar, sequence, or array to `jnp.int32`, checking int32 range.
 
     Args:
@@ -46,3 +56,41 @@ def safe_to_int32(value: object, *, name: str) -> Array:
             )
             raise ValueError(msg)
     return jnp.asarray(np_value, dtype=jnp.int32)
+
+
+def safe_to_float_dtype(value: object, *, name: str) -> Array:
+    """Cast a scalar, sequence, or array to the canonical float dtype.
+
+    Range check fires only on a down-cast:
+
+    - Down-cast (float64 → float32 under `jax_enable_x64=False`): raise
+      `OverflowError` if any element exceeds float32 magnitude rather
+      than letting JAX silently saturate to ``±inf``.
+    - Up-cast or same-width cast: skip the range check. Precision loss
+      within range is not an error — it is an inherent consequence of
+      `jax_enable_x64=False`.
+
+    Args:
+        value: A Python float, numpy/JAX scalar, or array-like.
+        name: Qualified name of the leaf — surfaced in the error message.
+
+    Returns:
+        A JAX array at `canonical_float_dtype()` (0-d if `value` was a
+        scalar).
+
+    Raises:
+        OverflowError: If down-casting to `float32` would saturate any
+            element to `±inf`. The message names the leaf via `name`.
+
+    """
+    target_dtype = canonical_float_dtype()
+    np_value = np.asarray(value)
+    if target_dtype == jnp.float32 and np_value.size > 0:
+        max_mag = float(np.max(np.abs(np_value)))
+        if max_mag > _FLOAT32_MAX:
+            msg = (
+                f"{name}: float32 overflow — max |value| {max_mag:g} "
+                f"exceeds float32 max {_FLOAT32_MAX:g}."
+            )
+            raise OverflowError(msg)
+    return jnp.asarray(np_value, dtype=target_dtype)