[Bug] [Relax][ONNX] Constant INT64 Concat data becomes ShapeExpr and breaks Add

未关闭
  • 哪里不对？
#20,889 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
1-2 天
新手友好度
38/100
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
compilers

调研方向

The failure starts in the Relax ONNX frontend entry point from_onnx when a Concat of two constant INT64 tensors feeds an Add node named add_data, which raises "cannot handle ShapeExpr inputs". Run the reproduction script from the issue to confirm that only the all-constant variant fails, then trace where the Concat output is classified as a shape rather than tensor data. Done means both variants import and the Add returns [11, 22], matching the ReferenceEvaluator result.

由索引模型根据 Issue 内容生成。

描述

needs-triage type: bug
Expected behavior

Concatenating INT64 tensors [1] and [2], then adding [10, 20], should return the INT64 tensor [11, 22], whether the operands are constants or runtime inputs.

Actual behavior

Making both Concat operands constant causes the frontend to treat ordinary tensor data as a shape. Its Add consumer fails:

ValueError: Node add_data cannot handle ShapeExpr inputs.

Keeping one operand as runtime data imports successfully. Both ONNX variants pass full checking and reference evaluation. Unlike the mixed shape/tensor case in #19498, this graph does not construct a shape.

Environment

TVM 0.28.dev0 (local build; build commit unavailable), Python 3.12.3, Linux x86_64; ONNX 1.22.0, NumPy 2.5.4.

Steps to reproduce

Run with python repro.py:

import numpy as np
import onnx
from onnx import TensorProto as T, helper as h, numpy_helper as nh
from onnx.reference import ReferenceEvaluator
from tvm.relax.frontend.onnx import from_onnx

for constant_a in [False, True]:
    nodes = [h.make_node("Constant", [], ["b"], value=nh.from_array(np.array([2], dtype="int64")))]
    inputs = [h.make_tensor_value_info("x", T.INT64, [2])]
    feeds = {"x": np.array([10, 20], dtype="int64")}
    if constant_a:
        nodes.append(h.make_node("Constant", [], ["a"], value=nh.from_array(np.array([1], dtype="int64"))))
    else:
        inputs.append(h.make_tensor_value_info("a", T.INT64, [1]))
        feeds["a"] = np.array([1], dtype="int64")
    nodes += [h.make_node("Concat", ["a", "b"], ["c"], axis=0),
              h.make_node("Add", ["c", "x"], ["y"], name="add_data")]
    model = h.make_model(h.make_graph(nodes, "concat", inputs,
        [h.make_tensor_value_info("y", T.INT64, [2])]),
        opset_imports=[h.make_opsetid("", 18)], ir_version=10)
    onnx.checker.check_model(model, full_check=True)
    np.testing.assert_array_equal(ReferenceEvaluator(model).run(None, feeds)[0], [11, 22])
    try:
        from_onnx(model, keep_params_in_input=False)
        print("constant_a", constant_a, "imported")
    except Exception as error:
        print("constant_a", constant_a, type(error).__name__, str(error).splitlines()[-1])
Triage
  • needs-triage
主要语言
Python
星标
13.9k
派生
4k
平均合并
7 小时 59 分钟
30 天内合并 PR
159

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。

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。

apache/tvm 的其他 Issue

查看 apache/tvm 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。