Reduction rewrites discard explicit accumulator dtype and intermediate narrowing

オープン
#2,518 コメント 2 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

まだ誰も着手していません。

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
43/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
numpy, python, pytorch

調査の方向性

Start by reading the six named rewrites: local_sum_prod_all_to_none, local_useless_reduce, local_reduce_broadcastable, local_reduce_chain, local_reduce_join, and local_careduce_of_alloc; the issue identifies each rewrite’s failing precision case. Run the supplied repro.py with the stated PyTensor and NumPy versions, comparing FAST_RUN with the unoptimized and six-exclusion modes. Done means the rewrites preserve explicit accumulator dtype and intermediate narrowing across the listed cases and linkers.

索引モデルが issue の本文から書いたものです。

説明

Problem

Several reduction rewrites discard an explicitly selected acc_dtype or an intermediate narrowing boundary. This is independent of the PyTorch lowering problem in https://github.com/pymc-devs/pytensor/issues/2516.

Reproduced with PyTensor 3.3.2 / NumPy 2.5.3 on Apple Silicon. The script below imports only upstream PyTensor and NumPy. Unoptimized and six-exclusion runs produce the expected values.

Rewrite / case Linker Expected FAST_RUN
local_sum_prod_all_to_none: uint16 [65535, 1], axis=0, output=float64, accumulator=uint16 Python 0 65536
local_useless_reduce: int64 [16777217], axis=(), output=int64, accumulator=float32 Python 16777216 16777217
local_reduce_broadcastable: same value, shape (1, None), axis=0 Python 16777216 16777217
local_reduce_chain: inner uint16 reduction followed by float64 reduction Python 0 65536
local_reduce_join: float32 vectors [1e8], [1], [-1e8], stacked and summed with float64 accumulator C 1 0
local_careduce_of_alloc: repeat uint16 65535 twice, output=float64, accumulator=uint16 Python 65534 131070

The Join case happens to return 1 with the Python linker; the C linker demonstrates the lost precision boundary. It also fails with our native PyTorch execution of the rewritten graph. This report does not claim that every linker fails in every row.

The all-axes rewrite reconstructs the reduction without acc_dtype; empty/broadcast axes can remove a required cast through the accumulator; chain/join/Alloc transformations can remove a required rounding or modular-overflow step. The current main local_careduce_of_alloc still explicitly says to ignore acc_dtype when replacing a reduction by multiplication; that is not valid for the unsigned example.

Reproducer

Save as repro.py and run python repro.py --output results.json (a working C compiler is needed for the three Join C-linker runs).

import argparse
import json
from pathlib import Path

import numpy as np
import pytensor
import pytensor.tensor as pt
from pytensor.compile.mode import Mode
from pytensor.tensor.type import TensorType

EXCLUDED = (
    "local_sum_prod_all_to_none",
    "local_useless_reduce",
    "local_reduce_broadcastable",
    "local_reduce_chain",
    "local_reduce_join",
    "local_careduce_of_alloc",
)
MODES = {
    "unoptimized_python": Mode(linker="py", optimizer=None),
    "fast_run_python": Mode(linker="py", optimizer="fast_run"),
    "six_exclusions_python": Mode(linker="py", optimizer="fast_run").excluding(*EXCLUDED),
}
parser = argparse.ArgumentParser()
parser.add_argument("--output", required=True)
args = parser.parse_args()
records = []


def run(name, inputs, output, values, expected):
    modes = dict(MODES)
    if name == "joined_reduction_accumulator":
        modes.update(
            unoptimized_c=Mode(linker="c", optimizer=None),
            fast_run_c=Mode(linker="c", optimizer="fast_run"),
            six_exclusions_c=Mode(linker="c", optimizer="fast_run").excluding(*EXCLUDED),
        )
    for label, mode in modes.items():
        fn = pytensor.function(inputs, output, mode=mode)
        record = {
            "case": name,
            "mode": label,
            "expected": np.asarray(expected).tolist(),
            "expected_dtype": str(np.asarray(expected).dtype),
            "optimized_ops": [str(node.op) for node in fn.maker.fgraph.toposort()],
        }
        try:
            actual = np.asarray(fn(*values))
            record.update(
                actual=actual.tolist(),
                actual_dtype=str(actual.dtype),
                matches=bool(actual.dtype == np.asarray(expected).dtype and np.array_equal(actual, expected)),
            )
        except Exception as exc:
            record.update(error_type=type(exc).__name__, error=str(exc), matches=False)
        records.append(record)


x = pt.vector("u16", dtype="uint16")
run("all_axes_accumulator", [x], pt.sum(x, axis=0, dtype="float64", acc_dtype="uint16"), [np.array([65535, 1], dtype="uint16")], np.array(0, dtype="float64"))
x = pt.lvector("integer")
run("empty_axes_accumulator", [x], pt.sum(x, axis=(), dtype="int64", acc_dtype="float32"), [np.array([16777217], dtype="int64")], np.array([16777216], dtype="int64"))
x = TensorType("int64", shape=(1, None))("broadcast")
run("broadcast_axis_accumulator", [x], pt.sum(x, axis=0, dtype="int64", acc_dtype="float32"), [np.array([[16777217]], dtype="int64")], np.array([16777216], dtype="int64"))
x = pt.matrix("chain", dtype="uint16")
run("nested_reduction_dtype", [x], pt.sum(pt.sum(x, axis=0, dtype="uint16", acc_dtype="uint16"), dtype="float64", acc_dtype="float64"), [np.array([[65535, 0], [1, 0]], dtype="uint16")], np.array(0, dtype="float64"))
x, y, z = [pt.fvector(name) for name in ("x", "y", "z")]
run("joined_reduction_accumulator", [x, y, z], pt.sum(pt.stack([x, y, z]), axis=0, dtype="float32", acc_dtype="float64"), [np.array([value], dtype="float32") for value in (1e8, 1, -1e8)], np.array([1], dtype="float32"))
x = pt.scalar("allocated", dtype="uint16")
run("allocated_reduction_accumulator", [x], pt.sum(pt.alloc(x, 2), dtype="float64", acc_dtype="uint16"), [np.array(65535, dtype="uint16")], np.array(65534, dtype="float64"))

result = {"pytensor": pytensor.__version__, "numpy": np.__version__, "excluded": EXCLUDED, "records": records}
Path(args.output).write_text(json.dumps(result, indent=2) + "\n")
print(json.dumps(result), flush=True)
assert all(item["matches"] for item in records if item["mode"] not in ("fast_run_python", "fast_run_c"))

Downstream handling

We currently exclude these six rewrites in our native PyTorch mode, including nested OpFromGraph/Scan compilation, rather than changing global registrations or installed dependencies. Tracking downstream: https://github.com/pymc-labs/pytensor.cpp/issues/100. We can remove the exclusions when an upstream release preserves these explicit precision boundaries and passes these cases.

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。