Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

onnx_graphsurgeon: Constant with float8_e4m3fnuz values is exported as FLOAT8E4M3FN

未关闭 适合新手
  • 哪里不对？
#4,856 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

@luicarus 已经在做这个了。

开始于 2026年9月21日。

  • #4857 来自 @luicarus —— 未关闭

评估

难度
1/5
预计耗时
1 小时以内
新手友好度
92/100
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

调研方向

从 onnx_graphsurgeon/exporters/onnx_exporter.py 中的 dtype_to_onnx 和 ml_dtype_to_onnx_name 表开始。运行最小的 Constant/export_tensor_proto 复现，然后验证 float8_e4m3fnuz 是否导出为 FLOAT8E4M3FNUZ，以及导入的 initializer 在值被具体化并重新导出后是否仍保留该类型。

由索引模型根据 Issue 内容生成。

描述

Module:GraphSurgeon
Description

dtype_to_onnx in onnx_graphsurgeon/exporters/onnx_exporter.py currently maps ml_dtypes.float8_e4m3fnuz to FLOAT8E4M3FN instead of FLOAT8E4M3FNUZ:

ml_dtype_to_onnx_name = {
    np.dtype(ml_dtypes.float8_e4m3fn):   "FLOAT8E4M3FN",
    np.dtype(ml_dtypes.float8_e4m3fnuz): "FLOAT8E4M3FN",  # should be FLOAT8E4M3FNUZ
    np.dtype(ml_dtypes.float8_e5m2):     "FLOAT8E5M2",
    np.dtype(ml_dtypes.float8_e5m2fnuz): "FLOAT8E5M2FNUZ",
    ...
}

FLOAT8E4M3FN and FLOAT8E4M3FNUZ are distinct ONNX data types with the same storage width. As a result, the exported raw_data has the expected size and bytes, but the tensor is tagged with the wrong type. Consumers that decode those bytes according to the ONNX type can therefore interpret the values differently.

A minimal reproduction:

import ml_dtypes
import numpy as np
import onnx

from onnx_graphsurgeon.ir.tensor import Constant
from onnx_graphsurgeon.exporters.onnx_exporter import OnnxExporter

values = np.random.random_sample((2, 4)).astype(
    ml_dtypes.float8_e4m3fnuz
)
proto = OnnxExporter.export_tensor_proto(
    Constant("c", values=values)
)

print(proto.data_type)                  # 17: FLOAT8E4M3FN
print(onnx.TensorProto.FLOAT8E4M3FNUZ)  # 18: expected

The issue also breaks import/export round trips once a lazy initializer is materialized. Import already handles FLOAT8E4M3FNUZ correctly, but after accessing .values, re-exporting the tensor changes its type from 18 to 17:

graph = gs.import_onnx(model_with_fnuz_initializer)
tensor = graph.tensors()["w"]

_ = tensor.values  # materialize LazyValues

OnnxExporter.export_tensor_proto(tensor).data_type
# 17 (FLOAT8E4M3FN), originally 18 (FLOAT8E4M3FNUZ)

float8_e5m2fnuz is mapped correctly to FLOAT8E5M2FNUZ in the same table, so the problem appears to be isolated to the float8_e4m3fnuz entry.

Expected behavior

ml_dtypes.float8_e4m3fnuz values should export with:

onnx.TensorProto.FLOAT8E4M3FNUZ

and an imported FLOAT8E4M3FNUZ initializer should preserve its data type after materialization and re-export.

Environment
  • onnx_graphsurgeon 0.6.1
  • also reproducible on current main
  • onnx 1.23.0
  • ml_dtypes 0.6.0
  • Python 3.13
Provenance

The mapping appears to have been introduced in commit a180e081 (10.8 release changes, 2025-01-31), which added the ml_dtype_to_onnx_name table. The affected entry currently matches the float8_e4m3fn mapping immediately above it.

主要语言
C++
星标
13.4k
派生
2.4k
平均合并
15 分钟
30 天内合并 PR
2

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。

NVIDIA/TensorRT 的其他 Issue

查看 NVIDIA/TensorRT 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。