torchtalk-trace

torchtalk-trace is a skill for Claude Code from opendatahub-io/ai-helpers. It costs 24 tokens per session (963 once invoked), scanned A, original, Apache-2.0.

A tracing tool for following a PyTorch function from its Python interface through C++ code to CUDA or other hardware-specific implementations.

In plain words
What is it for?
It helps locate bindings, backend implementations, automatic differentiation code, and related file and line references.
Why use it?
It makes the path through PyTorch's multiple programming layers easier to understand when debugging or modifying code.

Skill for Claude Code

Written for Claude Code: argument-hint in frontmatter. Also seen: mentions Claude Code.

Part of the odh-pytorch plugin — 3 skills shipped together

Good fit It helps locate bindings, backend implementations, automatic differentiation code, and related file and line references.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/opendatahub-io/ai-helpers/torchtalk-trace
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add opendatahub-io/ai-helpers --skill torchtalk-trace
Clone the repo
git clone --depth 1 https://github.com/opendatahub-io/ai-helpers

Made for: Claude Code.

Or install odh-pytorch, the plugin that ships this one along with the rest of its 3 skills.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for torchtalk-trace

README.md
[![agentmods](https://agentmods.dev/badge/skills/opendatahub-io/ai-helpers/torchtalk-trace/github.svg)](https://agentmods.dev/skills/opendatahub-io/ai-helpers/torchtalk-trace)
Your own site
<a href="https://agentmods.dev/skills/opendatahub-io/ai-helpers/torchtalk-trace"><img src="https://agentmods.dev/badge/skills/opendatahub-io/ai-helpers/torchtalk-trace/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for torchtalk-trace

Your own site · 80×15
<a href="https://agentmods.dev/skills/opendatahub-io/ai-helpers/torchtalk-trace"><img src="https://agentmods.dev/badge/skills/opendatahub-io/ai-helpers/torchtalk-trace.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 24 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 963 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector pass 7 Sept 2026
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00024 $0.00963
Opus 5 $0.00012 $0.00481
Sonnet 5 $0.00005 $0.00193
Haiku 4.5 $0.00002 $0.00096

Measured 7d ago against content hash eff5eb0d7bcf, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-11, from the pricing page.

Security

Grade A, and why

torchtalk-trace scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 7d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugins/odh-pytorch/skills/torchtalk-trace/SKILL.md · 86 lines

How it starts

The opening of the file, as written. The whole thing — 86 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Name

odh-ai-helpers:torchtalk-trace

Synopsis

/torchtalk:trace <function-name>
/torchtalk:trace <function-name> full
/torchtalk:trace <function-name> dispatch
/torchtalk:trace <function-name> yaml

Description

The torchtalk:trace command traces a PyTorch function's complete binding chain from Python through C++ to CUDA implementations. It uses the TorchTalk MCP server to look up operator definitions in native_functions.yaml, find pybind11 and TORCH_LIBRARY bindings, and map dispatch keys to backend implementations.

After gathering the raw binding data, the command analyzes the results to explain:

  • How the Python API connects to the native implementation
  • Which dispatch keys route to which backends (CPU, CUDA, MPS, etc.)
  • Where each layer is implemented with exact file:line references
  • The autograd integration and backward pass formula (if applicable)
  • Architectural context about why the dispatch is structured this way

This command requires the TorchTalk MCP server to be running. Run mcp__torchtalk__get_status to verify availability.

Prerequisites

  • TorchTalk MCP server must be running and registered with Claude Code
  • Run /torchtalk:setup if not yet installed

Implementation

  1. Verify MCP server: Call mcp__torchtalk__get_status to confirm the TorchTalk server is running and has indexed data available
  2. Trace binding chain: Call mcp__torchtalk__trace with the function name and optional focus parameter to retrieve the full binding chain
  3. Get internal dependencies: Call mcp__torchtalk__calls with the function name to understand what the function invokes internally
  4. Analyze dispatch architecture: Examine the dispatch keys, backend routing, and structured binding patterns returned by the trace
  5. Synthesize explanation: Combine the trace data, call graph, and dispatch information into a coherent explanation that covers:
    • The Python entry point and how it connects to native code
    • The native_functions.yaml definition and its dispatch configuration
    • Each backend implementation with file:line references
    • The autograd formula (from derivatives.yaml) if applicable
    • Architectural observations about the operator's design
  6. Provide file references: Format all implementation locations as file:line references so the developer can navigate directly to the source

Read the full file on GitHub · 86 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 7d ago First seen · 86 lines · 24 tokens per session scan A eff5eb0d7bcf

Subscribe to this mod's changes

torchtalk-trace is a skill published in the GitHub repository opendatahub-io/ai-helpers (37 stars, last pushed 3d ago), licensed Apache-2.0. It adds 24 tokens to every session and 963 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.

Related

Other skills, from other repositories

cuopt-numerical-optimization-api

LP, MILP, and QP (beta) with cuOpt — Python, C, and CLI. Use when the user is solving LP, MILP, or QP with any cuOpt interface.

NVIDIA/skills · 51 tokens

ampl-python

Expert in amplpy and Python integration with AMPL. Use when writing or refactoring amplpy scripts, notebooks, FastAPI services, data pipelines (pandas/polars), solver configuration via Python, result extraction, or Colab/MO-Book Python workflows. Knows amplpy.ampl.com and dev.ampl.com amplpy best practices.

marcos-dv/ampl-agents · 75 tokens

temporal-python-testing

Test Temporal workflows with pytest, time-skipping, and mocking strategies. Covers unit testing, integration testing, replay testing, and local development setup. Use when implementing Temporal workflow tests or debugging test failures.

wshobson/agents · 45 tokens

fastapi-templates

Create production-ready FastAPI projects with async patterns, dependency injection, and comprehensive error handling. Use when building new FastAPI applications or setting up backend API projects.

wshobson/agents · 37 tokens

manimgl-best-practices

Trigger when: (1) User mentions "manimgl" or "ManimGL" or "3b1b manim", (2) Code contains from manimlib import , (3) User runs manimgl CLI commands, (4) Working with InteractiveScene, self.frame, self.embed(), ShowCreation(), or ManimGL-specific patterns. Best practices for ManimGL (Grant Sanderson's 3Blue1Brown…

calesthio/OpenMontage · 167 tokens

cuopt-developer

Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.

NVIDIA/cuopt · 47 tokens