Ricardo Vieira ricardoV94

How `Scan.L_op` Works

Summary

Scan.L_op computes the gradient of a Scan operation by constructing a new backward Scan that runs in reverse, propagating gradients through time. The method:

Differentiates the inner function symbolically to get per-step gradients
Adds accumulation terms for recurrent states
Constructs a new backward Scan with reversed sequences and mit-mot states (initialized with output gradients, accumulate total gradients after evaluation)
Re-orders the backward Scan's outputs to match the expected gradient layout

pytensor-suite-feedstock

This is a conda-forge feedstock for PyTensor. The recipe produces two packages: pytensor-base (core, lightweight) and pytensor (adds compiler + BLAS deps).

Reviewing bot version-bump PRs

When the autotick bot opens a version-bump PR, verify that PyPI dependency changes are reflected in the recipe before merging.

	from functools import wraps

	import numpy as np

	import pytensor.tensor as pt
	from pytensor import function
	from pytensor.tensor import tensor
	from pytensor.compile.function.types import Function

	import numpy as np

	import pymc as pm
	import pytensor.tensor as pt
	import pytensor.xtensor as px

	N = 100
	seed = sum(map(ord, "xarray>=numpy?"))
	rng = np.random.default_rng(seed)

Ricardo Vieira ricardoV94

How Scan.L_op Works

Summary

pytensor-suite-feedstock

Reviewing bot version-bump PRs

How `Scan.L_op` Works