Keyword arguments are silently ignored when a decorated function is called without positional arguments

未关闭 适合新手
  • 哪里不对？
#88 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
2/5
预计耗时
1-3 小时
新手友好度
76/100
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
performance

调研方向

从 nsight/collection/core.py 开始，追踪 _sanitize_configs 周围以及 if len(args) > 0 检查处的调用时配置处理。阅读 test_config_at_call_time_with_kwargs，并为仅使用关键字参数的调用（例如 bench(n=4096)）补充测试覆盖。完成标准是：关键字参数要么被用作配置，要么始终引发文档所述的 ProfilerException，且不会静默回退到默认值。

由索引模型根据 Issue 内容生成。

描述

Description

Calling a function decorated with @nsight.analyze.kernel with only keyword arguments drops them. The call is profiled with the parameter default (or with the decorator-time configs), and the results report that configuration instead of the requested one. No error or warning is raised.

import nsight
import torch

@nsight.analyze.kernel
def bench(n: int = 64) -> None:
    a = torch.randn(n, n, device="cuda")
    b = torch.randn(n, n, device="cuda")
    with nsight.annotate("add"):
        _ = a + b

df = bench(n=4096).to_dataframe()
print(df[["Annotation", "n", "Metric"]])   # n is 64, not 4096

The same happens with @nsight.analyze.kernel(configs=[(64,)]) and bench(n=4096).

bench(42, y=23) already raises ProfilerException: Keyword arguments are not supported yet: ['y'] (_sanitize_configs in nsight/collection/core.py, tested by test_config_at_call_time_with_kwargs). That check sits inside if len(args) > 0:, so it is skipped when there are no positional arguments.

Expected

The same ProfilerException, or the keyword arguments being used as the configuration.

Environment

nsight-python main at c89d4b5, Python 3.12.3, Linux x86_64.

How this was checked: no GPU was available, so the profiling backend was replaced by a mock (NVTX ranges and NCUCollector.collect). Config handling, the run loop, extraction and aggregation were the real nsight code. An equivalent script with the kernel launch stubbed out shows n=64 in the results for both cases (requested n=4096, profiled n=64). It has not been run on a GPU with NCU or CUPTI.

A small fix is ready if you approve it: move the existing check out of the if len(args) > 0: block, so keyword arguments raise the same ProfilerException on every call, and add a test next to test_config_at_call_time_with_kwargs. That turns today's silent wrong result into an error for callers who use bench(n=...). Supporting keyword arguments instead would be your call.

This was found with an AI coding assistant (Claude Code, claude-fable-5-2), which also wrote and ran the mocked repro and drafted this report.

主要语言
Python
星标
301
派生
23
平均合并
2 天 20 小时
30 天内合并 PR
12

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/nsight-python 的其他 Issue

查看 NVIDIA/nsight-python 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。