Edge SpAIce CERN Technical Meeting

Tuesday 28 Jan 2025, 10:00 → 11:00 Europe/Zurich

42/R-001 (CERN)

10:00 → 11:00 Updates

Speakers: Nicolo Ghielmetti (CERN), Noemi D'Abbondanza (Sapienza Universita e INFN, Roma I (IT)), Sioni Paris Summers (CERN), Stylianos Tzelepis (National Technical Univ. of Athens (GR))

Nicolo's updates

• Edge SpAIce Gitlab Pipeline:
  • 4 stages implemented:
    • check: test if the commit come from edge-spaice branch, otherwise abort the execution of the pipeline
    • fetch: retrieve QONNX models from the submodule edge-spaice-models which have to be tested in the pipeline. Generated QONNX channels-first models as artifacts.
    • convert: transform the QONNX models in channels-last format and check whether the modification creates any numerical mismatch. Generate QONNX channels-last models as artifacts.
    • generate: produce hls4ml models from the QONNX channels-last models, check whether there are mismatches between hls4ml and QONNX predictions. Generate some images as artifacts
  • The pipeline is triggered everytime that:
    • github.com/nghielme/hls4ml@edge-spaice receive a push
    • github.com/nghielme/QONNX@edge-spaice receive a push
    • gitlab.cern.ch/edge-spaice/edge-spaice-pipeline@master receive a push
    • gitlab.cern.ch/edge-spaice/edge-spaice-models@master receive a push
  • Nicolas has now access to the pipeline and can push the new models directly into edge-spaice-models 
  • To be implemented:
    • Execute the pipeline in parallel for each fetched model: parallel:matrix require an array of string for the parameters, no evaluation can be executed before running the pipeline
    • Add build stage in which the hls4ml generated model is built using vitis_hls and vivado 

 

• Idea to discuss: profile the weights and avoid to repeat the same values multiple times across different layers. 
  For example,
  • with 4 bits only 2⁵ - 1 = 31 different values can exist
  • with 8 bits, 2⁹ - 1 = 511
    Each layer uses pointers to a central common storage of ap_fixed<> values to represent the weights, and need to know the scale of the value

 

• Draft PR: Support for multigraph build, slides
  • Models with branches are supported if you don't cut the model within the branch

 

Stelios' Updates

• Vladimir to review the open PRs soon
• Conferences->check shared codimd
• Gitlab wikis in progress
• PixESL
  • meeting on Thursday
  • Need ROOT to get the clustered data
  • presentation