Manual control of numba caching

Author: ricardoV94

doc/extending/creating_a_numba_jax_op.rst

@@ -228,7 +228,7 @@ Here's an example for :class:`DimShuffle`:
                     # E       No match.
                     # ...(on this line)...
                     # E           shuffle_shape = res.shape[: len(shuffle)]
-                    @numba_basic.numba_njit(inline="always")
+                    @numba_basic.numba_njit
                     def dimshuffle(x):
                         return dimshuffle_inner(np.asarray(x), shuffle)
 

pytensor/bin/pytensor_cache.py

@@ -1,5 +1,6 @@
 import logging
 import os
+import shutil
 import sys
 from pathlib import Path
 
@@ -74,7 +75,10 @@ def main():
                     'You can also call "pytensor-cache purge" to '
                     "remove everything from that directory."
                 )
-                _logger.debug(f"Remaining elements ({len(items)}): {', '.join(items)}")
+                _logger.debug(f"Remaining elements ({len(items)}): {items}")
+            numba_cache_dir: Path = config.base_compiledir / "numba"
+            shutil.rmtree(numba_cache_dir, ignore_errors=True)
+
         elif sys.argv[1] == "list":
             pytensor.compile.compiledir.print_compiledir_content()
         elif sys.argv[1] == "cleanup":
@@ -86,6 +90,8 @@ def main():
             print("Lock successfully removed!")
         elif sys.argv[1] == "purge":
             pytensor.compile.compiledir.compiledir_purge()
+            numba_cache_dir: Path = config.base_compiledir / "numba"
+            shutil.rmtree(numba_cache_dir, ignore_errors=True)
         elif sys.argv[1] == "basecompiledir":
             # Simply print the base_compiledir
             print(pytensor.config.base_compiledir)

pytensor/link/numba/cache.py

@@ -0,0 +1,114 @@
+import weakref
+from collections.abc import Callable
+from pathlib import Path
+from tempfile import NamedTemporaryFile
+from typing import Any
+
+from numba.core.caching import CacheImpl, _CacheLocator
+
+from pytensor.configdefaults import config
+
+
+NUMBA_CACHE_PATH = config.base_compiledir / "numba"
+NUMBA_CACHE_PATH.mkdir(exist_ok=True)
+CACHED_SRC_FUNCTIONS = weakref.WeakKeyDictionary()
+
+
+class NumbaPyTensorCacheLocator(_CacheLocator):
+    """Locator for Numba functions defined from PyTensor-generated source code.
+
+    It uses an internally-defined hash to disambiguate functions.
+
+    Functions returned by the PyTensor dispatchers are cached in the CACHED_SRC_FUNCTIONS
+    weakref dictionary when `compile_numba_function_src` is called with a `cache_key`.
+    When numba later attempts to find a cache for such a function, this locator gets triggered
+    and directs numba to the PyTensor Numba cache directory, using the provided hash as disambiguator.
+
+    It is not necessary that the python functions be cached by the dispatchers.
+    As long as the key is the same, numba will be directed to the same cache entry, even if the function is fresh.
+    Conversely, if the function changed but the key is the same, numba will still use the old cache.
+    """
+
+    def __init__(self, py_func, py_file, hash):
+        self._py_func = py_func
+        self._py_file = py_file
+        self._hash = hash
+
+    def ensure_cache_path(self):
+        """We ensured this when the module was loaded.
+
+        It's too slow to run every time a cache is needed.
+        """
+        pass
+
+    def get_cache_path(self):
+        """Return the directory the function is cached in."""
+        return NUMBA_CACHE_PATH
+
+    def get_source_stamp(self):
+        """Get a timestamp representing the source code's freshness.
+        Can return any picklable Python object.
+
+        This can be used to invalidate all caches from previous PyTensor releases.
+        """
+        return 0
+
+    def get_disambiguator(self):
+        """Get a string disambiguator for this locator's function.
+        It should allow disambiguating different but similarly-named functions.
+        """
+        return self._hash
+
+    @classmethod
+    def from_function(cls, py_func, py_file):
+        """Create a locator instance for functions stored in CACHED_SRC_FUNCTIONS."""
+        if config.numba__cache and py_func in CACHED_SRC_FUNCTIONS:
+            return cls(py_func, Path(py_file).parent, CACHED_SRC_FUNCTIONS[py_func])
+
+
+# Register our locator at the front of Numba's locator list
+CacheImpl._locator_classes.insert(0, NumbaPyTensorCacheLocator)
+
+
+def compile_numba_function_src(
+    src: str,
+    function_name: str,
+    global_env: dict[Any, Any] | None = None,
+    local_env: dict[Any, Any] | None = None,
+    store_to_disk: bool = False,
+    cache_key: str | None = None,
+) -> Callable:
+    """Compile (and optionally cache) a function from source code for use with Numba.
+
+    This function compiles the provided source code string into a Python function
+    with the specified name. If `store_to_disk` is True, the source code is written
+    to a temporary file before compilation. The compiled function is then executed
+    in the provided global and local environments.
+
+    If a `cache_key` is provided the function is registered in a `CACHED_SRC_FUNCTIONS`
+    weak reference dictionary, to be used by the `NumbaPyTensorCacheLocator` for caching.
+
+    """
+    if store_to_disk:
+        with NamedTemporaryFile(delete=False) as f:
+            filename = f.name
+            f.write(src.encode())
+    else:
+        filename = "<string>"
+
+    if global_env is None:
+        global_env = {}
+
+    if local_env is None:
+        local_env = {}
+
+    mod_code = compile(src, filename, mode="exec")
+    exec(mod_code, global_env, local_env)
+
+    res = local_env[function_name]
+    res.__source__ = src  # type: ignore
+
+    if cache_key is not None:
+        CACHED_SRC_FUNCTIONS[res] = cache_key
+
+    return res