Gradient of batched block_diag falls back to numba object mode (Blockwise ARange)

オープン
#2,702 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@saksham10arora-dotcom がすでに取り組んでいます。

2026年11月21日 から。

  • #2703 @saksham10arora-dotcom による — オープン

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
35/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
numpy, python
領域
performance

調査の方向性

Reproduce with the snippet in mode="NUMBA" and confirm the Blockwise{ARange} object-mode warning and the timing gap. Then read BlockDiagonal.pullback and how Blockwise calls vectorize_shape, since that is where core shapes become batched; compare against the concatenate-based construction to see what a non-batched pullback would look like. Done means the gradient runs in nopython mode, timings approach the concatenate figure, and a regression test covers batched plus unbatched inputs.

索引モデルが issue の本文から書いたものです。

説明

When one input to block_diag is batched, its gradient compiles a Blockwise{ARange} that numba runs in object mode. In a small example the gradient is about 20x slower than building the same matrix with concatenate.

import pytensor
import pytensor.tensor as pt

A = pt.tensor3("A")  # batch of 6x6 blocks
B = pt.matrix("B")   # 2x2, not batched
out = (pt.linalg.block_diag(A, B) ** 2).sum()
f = pytensor.function([A, B], pytensor.grad(out, [A, B]), mode="NUMBA")
# UserWarning: Numba will use object mode to run Blockwise{ARange{dtype='int64'}, (),(),()->(o00)}'s perform method

Timings with A of shape (1000, 6, 6), pytensor 3.3.8, M-series Mac:

gradient of ms per call
block_diag(A, B) as it is now 3.2
same, with the pullback using slices instead of arange 1.1 (still object mode, now Blockwise{Subtensor})
the same matrix built with concatenate and zeros 0.14

What I think happens: BlockDiagonal.pullback indexes gout with arange(index_begin, index_end), computed from i.shape. When Blockwise vectorizes the pullback, vectorize_shape broadcasts each core shape over the batch (which is correct in general), so the indices and the arange become batched even though they are identical for every batch element. Swapping arange for slices removes the ARange, but the slice bounds are still batched, so it moves the problem to Blockwise{Subtensor}.

Where I hit it: pymc-extras structural models. TimeSeasonality(duration > 1) has a time-varying transition matrix, and adding any other component (e.g. Autoregressive) combines them with block_diag(T, o_T), so the Kalman filter gradient picks this up (pymc-devs/pymc-extras#833).

I'm not sure what the preferred fix is: a pullback that avoids batched indices, a rewrite that drops batch dims from Blockwise inputs that only come from broadcast shapes, or something else. Happy to work on it with some direction.

主要言語
Python
スター
654
フォーク
258
平均マージ
2日 7時間
マージ済み PR(30日)
39

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

pymc-devs/pytensor のほかの issue

pymc-devs/pytensor の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。