`SlicesSplit` rule makes `optimizer.optimize` quadratic in the number of `Slice` nodes

未关闭
  • 哪里不对？
#3,154 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 4 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
68/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python
领域
machine-learning, performance

调研方向

Start by reading SimplePatternMatcher.match and _multi_match in onnxscript.rewriter, then inspect the SlicesSplit rule in _DEFAULT_REWRITE_RULES. Use the provided repro to compare its runtime with the other default rules as the number of Slice nodes grows. Done means SlicesSplit no longer exhibits quadratic scaling while preserving rewrite behavior.

由索引模型根据 Issue 内容生成。

描述

module: optimizer

SimplePatternMatcher.match handles a pattern with several output nodes by fixing the first output to the candidate node and trying every combination (itertools.product) of graph nodes with the same op type for the other outputs. It also rebuilds an index of all the graph's nodes for each candidate. The only multi-output rule in _DEFAULT_REWRITE_RULES is SlicesSplit, so on a graph with many Slices that one rule takes far longer than all the others, and its time grows with the square of the number of Slices. optimize() runs it on every iteration, and optimize_for_ort calls optimize() three times.

We hit this exporting a mixture-of-experts model whose experts were unrolled into a Slice per expert (24 layers × 32 experts, about 3,900 Slices). torch.onnx.export(optimize=True) took 11 minutes per graph, and optimize_for_ort 13 minutes, against seconds for a dense model of the same size. A py-spy profile put 97% of the time in SimplePatternMatcher.match / _multi_match.

Repro (onnxscript 0.7.2, onnx_ir 1.0.0): n independent Slices of one input, timing SlicesSplit alone against the other 45 default rules together.

"""onnxscript's `SlicesSplit` rewrite rule takes time quadratic in the number of `Slice` nodes in the graph.

The graph is n independent `Slice`s of one input. `SlicesSplit` alone is timed against every other default rewrite
rule together, on the same graph.
"""

import time

import numpy as np
import onnx_ir as ir
import onnxscript.rewriter


def constant(name, values):
    value = ir.Value(name=name, const_value=ir.tensor(np.array(values, dtype=np.int64)))
    value.dtype, value.shape = ir.DataType.INT64, ir.Shape([len(values)])
    return value


def build(n: int) -> ir.Model:
    x = ir.Value(name="x", type=ir.TensorType(ir.DataType.FLOAT), shape=ir.Shape([4 * n, 8]))
    nodes, initializers, outputs = [], [], []
    for i in range(n):
        arguments = [constant(f"starts{i}", [4 * i]), constant(f"ends{i}", [4 * i + 4]), constant(f"axes{i}", [0])]
        part = ir.node("Slice", [x, *arguments])
        relu = ir.node("Relu", [part.outputs[0]])
        relu.outputs[0].name = f"y{i}"
        relu.outputs[0].dtype, relu.outputs[0].shape = ir.DataType.FLOAT, ir.Shape([4, 8])
        nodes += [part, relu]
        initializers += arguments
        outputs.append(relu.outputs[0])
    graph = ir.Graph([x], outputs, nodes=nodes, initializers=initializers, opset_imports={"": 18})
    return ir.Model(graph, ir_version=10)


rules = list(onnxscript.rewriter._DEFAULT_REWRITE_RULES)
slices_split = [rule for rule in rules if rule.name == "SlicesSplit"]
others = [rule for rule in rules if rule.name != "SlicesSplit"]
for n in (250, 500, 1000, 2000):
    times = []
    for selected in (slices_split, others):
        model = build(n)
        start = time.perf_counter()
        onnxscript.rewriter.RewriteRuleSet(selected).apply_to_model(model)
        times.append(time.perf_counter() - start)
    print(f"{n:5d} Slice nodes: SlicesSplit {times[0]:6.2f} s, the {len(others)} other default rules {times[1]:6.2f} s")

  250 Slice nodes: SlicesSplit   0.24 s, the 45 other default rules   0.09 s
  500 Slice nodes: SlicesSplit   0.91 s, the 45 other default rules   0.18 s
 1000 Slice nodes: SlicesSplit   3.56 s, the 45 other default rules   0.36 s
 2000 Slice nodes: SlicesSplit  14.11 s, the 45 other default rules   0.72 s

Each doubling of n makes SlicesSplit 4x slower and the other rules together 2x slower.

Possible fixes:

  • for the other outputs, only try nodes that share an input with the first one, since that's what SlicesSplit requires anyway;
  • build the op_to_nodes index once per pass, not once per candidate;
  • let optimize() take a rule set, so callers can leave a rule out.
主要语言
Python
星标
457
派生
149
平均合并
3 天 20 小时
30 天内合并 PR
22

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

microsoft/onnxscript 的其他 Issue

查看 microsoft/onnxscript 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。