Every premium and members-only resource for $9.99/month, as a no-commitment subscription. See the plans

PalateKit

Sign in
Menu

Request a resource

Describe the resource you are missing in one sentence. We will email you when it is published.

0 / 280

A page, a video or an image online.

Unlock PalateKit

Unlock everything

Plan

Paying means accepting our terms of sale.

Secure payment by Stripe. Receipts and invoices sent by email.

Welcome back

Sign in to install, download or save this resource. A free account is enough for free resources.

Debug a MAX model whose output diverges v1.0.0

Use this resource to find the cause of an output divergence between a MAX model and its reference

New

  • 324 downloads
  • 101 likes
  • 124 saves

Free

Install with the CLI

In a terminal, from the folder you want:

npx --yes palatekit@latest install model-parity-debugging

With no option, the skill goes to ~/.agents/skills/; add --agent claude-code (or codex, cursor…) to target an agent.

First time: connect the CLI with

npx --yes palatekit@latest login

The agent builds a comparator that dumps tensors layer by layer, then checks the MAX model against the reference implementation. It ranks the gaps and proposes a verified fix before recompiling. You get a step-by-step method to fix a model that generates wrong tokens.

Mostly picked by

Updated November 2, 2026

Card for the “Debug a MAX model whose output diverges” skill

Details

Works in
Install guides
Generates with
Your agent's model
Licence
Apache-2.0

Similar resources

View all

What you get

  • SKILL.md: six-step protocol, from the HF sanity check to the fix
  • references/agent-workflow.md: how to split the work across investigation agents
  • references/stacked-failures.md: cases where several failures overlap

Good fit for

  • Your architecture port serves the model, but its tokens diverge from the reference
  • A quantization or multi-GPU conversion gives incoherent answers

Not for

    Tags

    • parity
    • debugging
    • model
    • tensors
    • max

    About

    You say

    “My MAX model gives the same logits but a different top-1 than PyTorch, help me find the gap.”

    Card for the “Debug a MAX model whose output diverges” skill

    Skill published by modular under the Apache-2.0 license, delivered in its original version.

    Show more

    How to use it

    1. Install the skill in your agent's skills folder
    2. Ask your agent to find the cause of an output divergence between a MAX model and its reference
    3. Review and adjust the result

    Content preview

    Files

    • LICENSE15.1 KB
    • NOTICE-PALATEKIT.md0.4 KB
    • references/agent-workflow.md2.7 KB
    • references/comparator-build.md12.7 KB
    • references/comparator-output-patterns.md4 KB
    • references/stacked-failures.md2.9 KB
    • SKILL.md7.3 KB

    Origin and licence

    Listed unmodified

    Author
    modular
    Source
    github.com/modular/skills/tree/9837a3af93e26456a6d05a9c05e8167cfb6e4f18/debug-model
    Licence
    Apache-2.0

    Notice

    « debug-model » (modular, Apache-2.0) livré dans sa version d'origine : fiche en français et en anglais par PalateKit.

    User reviews

    –

    No ratings yet

    Loading reviews…

    Delete your review?

    Your rating and comment will be removed from the page.