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

[PyTorch] fp8_cs_quantize fake implementation returns a vector inverse scale instead of a scalar

未关闭 适合新手
  • 哪里不对？
#3,636 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 2 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
82/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python, pytorch
领域
machine-learning, testing

调研方向

阅读 transformer_engine/pytorch/onnx_extensions.py 中 onnx_cs_quantize_fp8_op 及其 register_fake 实现附近的代码。修改 fake 逆缩放输出，使其与 eager 标量形状一致，然后在兼容的 TE/CUDA 环境中运行建议的 torch.library.opcheck FakeTensor 检查；当 eager 和 fake 输出元数据匹配时即为完成。

由索引模型根据 Issue 内容生成。

描述

Problem

At current main 9d4bd38678f29a29dc0989e3a679dc14f3538e6a, onnx_cs_quantize_fp8_op computes amax = tensor.abs().max() without a dimension, then scale_inv = 1 / scale. This produces a zero-dimensional float32 tensor. Its register_fake implementation instead returns torch.ones(1, ...) for that output, describing a one-dimensional tensor.

This violates the custom operator's eager/FakeTensor metadata contract. PyTorch's torch.library.opcheck(..., test_utils=("test_faketensor",)) reports:

found mismatched tensor metadata for output[1]:
Shapes torch.Size([]) and torch.Size([1]) are not equal!
Narrow reproduction and limitation

I isolated the exact upstream current-scaling function and its fake registration on Linux with PyTorch 2.11.0+cu130, CUDA hidden. I replaced only the lower tex::fp8_quantize operation with a CPU uint8 shape producer. The separately computed inverse scale does not depend on that producer's output. Across FP32/FP16/BF16 inputs with shapes (4,), (2, 3), and (2, 3, 4), all nine original metadata checks fail with the mismatch above.

Changing only the fake inverse-scale allocation to torch.ones((), dtype=torch.float32, device=tensor.device) makes all nine isolated metadata checks pass. This is a shape-contract reproduction, not a test of native TE FP8 kernels, numerical quantization, ONNX Runtime, or TensorRT export.

Suggested fix and native follow-up

Return a scalar inverse scale in the fake implementation:

return torch.empty(tensor.shape, dtype=torch.uint8, device=tensor.device), torch.ones(
    (), dtype=torch.float32, device=tensor.device
)

On a compatible installed TE/CUDA environment, a focused regression can compare eager and FakeTensor metadata using:

import torch
import transformer_engine.pytorch.onnx_extensions

x = torch.randn(16, 16, device="cuda", dtype=torch.float32)
torch.library.opcheck(torch.ops.tex.fp8_cs_quantize.default, (x,),
                     test_utils=("test_faketensor",))

The native follow-up command above is proposed, not executed in my isolated reproduction.

主要语言
Python
星标
3.6k
派生
844
平均合并
4 天 55 分钟
30 天内合并 PR
51

环境准备

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/TransformerEngine 的其他 Issue

查看 NVIDIA/TransformerEngine 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。