Madgraph5 GPU development

Tuesday 15 Sept 2026, 16:00 → 17:00 Europe/Zurich

513/R-070 - Openlab Space (CERN)

Meeting minutes 2026-09-15

Quick recap

The meeting focused on updates and discussions related to the MadGraph7 project, including a presentation by Kenneth on a new recognition framework for research software contributions using ORCID accounts.
The team discussed the compatibility and performance differences between MadGraph 5 and MadGraph 7, emphasizing that the new version is an evolution rather than a complete break.
Daniele shared results from recent performance tests, highlighting some discrepancies in cross-sections and throughput that need further investigation.
Frantisek announced the merging of a PR for mixed precision and proposed changes to the UI prompt color for better visibility.
Olivier provided updates on ongoing work with MLM and polarization, noting progress and remaining challenges.

Next steps

Frantisek

• Create a PR to modify the MadGraph 7 UI prompt (e.g., color it blue) and make it configurable.
• Update and push the code splitting PR for merging as soon as possible.

Daniele

• Organise a presentation (by Stefan or Kenneth) about the Everse recognition framework at the Cyprus meeting or the next monthly meeting.
• Rerun the MadGraph 7 performance and cross-section tests, starting with a subset of smaller processes, and investigate discrepancies.

Olivier

• Merge MadGraph 5.3.8 into MadGraph 7 and port the quarkonia/leptonium features.

Summary

Everse Project Update Presentation

Kenneth presented an update on the Everse project, which aims to provide best practices and guidelines for improving research software quality through a recognition framework.
The framework would require contributors to have an ORCID account and register on a special web platform to credit their contributions to projects like MadGraph.
Olivier expressed interest in the approach but emphasized the need for a smooth workflow and detailed implementation.
The group discussed presenting this information to other contributors, potentially at an upcoming meeting in Cyprus, and agreed to add Kenneth to the invitation list for the Cyprus meeting.

Talk about MadGraph7 at CMS Physics Days by Daniele

The discussion focused on the evolution from MadGraph 5 to MadGraph 7, where Olivier explained that while they are not completely breaking compatibility, there are changes to default settings that make it difficult to replicate the exact same code as before.
The team acknowledged the importance of validation processes and noted that while many users may assume breaking changes due to the jump from version 5 to 7, the syntax remains the same and the changes represent improvements rather than complete compatibility breaks.

Benchmarking results MadGraph5 vs MadGraph7

Daniele presented test results showing significant differences in cross-sections and PDF calculations between MadGraph 5 and MadGraph 7, particularly in processes like FCC e+ e- > z h (a factor of 4 difference), while for other processes the cross-section was anyway off of few factors even of 100% for processes that before where within uncertainty.
He also reported that cross-section measurements showed stable results across different backends.
However, throughput measurements revealed some unexpected results, with some performance metrics being below MadGraph 5 and identical performance between CPU and GPU configurations in certain cases, which Daniele planned to investigate further by running tests sequentially rather than in parallel.
The team identified discrepancies in performance comparisons between different grid packs that required further investigation.

PDFs and Error Investigation

The team discussed issues with PDFs and cross-section errors in their analysis.
Olivier suggested investigating potential problems with negative contributions in the PDFs, while also considered the error quoted by MadGraph 5 (on the HTML) to be quite small and that it should be considered at least 3 times it: the one quoted is very optimistic.
Daniele planned to run a subset of processes taking less than a few minutes each to understand throughput differences within half a day.

Mixed-precision PR

Frantisek reported that his mixed precision PR had been successfully merged.

Code-splitting PR

Now it's time to push this PR as much as possible, and we would need to merge it as soon as possible.
Frantisek noting that significant rework would be needed to update the current implementation due to changes in the repository structure.

UI Prompt Color Enhancement Proposal

Frantisek proposed changing the UI appearance by coloring the prompt blue as a visual indicator, with the ability to toggle it off in settings.
Frantisek agreed to create a pull request to implement these changes, including updating references from MG5 to MG7 throughout the interface.
The team decided to start with a blue color and opt-out feature.

Code Updates and Implementation Progress

Olivier provided updates on his work on MLM and Polarization, reporting good progress with W+Jets but identifying issues with mixed expansions where the code fails to identify diagrams correctly, creating a chicken-and-egg problem with weight evaluation and merging.

16:00 → 16:10 News

16:10 → 16:30 Topical discussion

Speaker: Dr All

Agenda

Actions

Daniele

• Complete the MG5/MG7 validation runs:
  • use MadNIS train_batches;
  • include the previously missing processes and V+jets;
  • investigate the 3W discrepancy, including sde_strategy to 2 and/or FD gauge.
• Repeat the CPU benchmarks on hardware with native 512-bit AVX-512 support.
• Update the validation/performance plots, showing MG5 and MG7 uncertainties separately and considering alternative representations.
• Coordinate the merging of the large renaming PR.

Theo

• Finalize and send the missing MadSpace and Madboard sections for the paper.
• Investigate the MadNIS training convergence issue and, once resolved, open the reproducibility PR.
• Prepare the final performance plots for the ML4Jets plenary by mid-September.

Olivier

• Finish the MadSpin paper work for MG5 and then continue with the MG7 part.

Collaboration

• Close the PRs (should be 2 of them) blocking code splitting, then merge the code-splitting PR.
• We will use the repository for collaboration of EVERSE for software recognition → Kenneth is gonna talk about that on behalf of Stefan.

Open points

• Daniele, Olivier and Theo: decide the default configuration for systematics, in particular scale and PDF variations.

16:30 → 16:50 Round table

16:50 → 17:00 AoB