Uploaded image for project: 'Red Hat Enterprise Linux AI'
  1. Red Hat Enterprise Linux AI
  2. RHELAI-2370

Add GradLib Support to ROCm Build

XMLWordPrintable

    • Icon: Bug Bug
    • Resolution: Done
    • Icon: Critical Critical
    • rhelai-1.3
    • None
    • Accelerators - AMD
    • None
    • Approved

      Goal: 

      Build Gradlib with packages from ROCm.

       

      Acceptance Criteria:

      • Gradlib builds

       

      Background:

      Gradlib is a ROCm only dependency in the AMD fork  for vLLM.  It has no entry in the pyproject.toml file and no requirements or version tracking.  AMD copy-pasted the code whole into the project as a way to improve gradient calculation performance with optimizations for hipblas and rocblas GEMMs. This was found very late in the cycle and was only delivered to the Builder team a week before the freeze date.

      We must add it or we will get runtime failures for the ROCm fork of vLLM.

       

      In the future we may take a different integration approach but for the urgency faced in RHEL AI 1.3, we will take the early work for this feature and integrate a simple port of the build from AMD and deliver this as a wheel, allowing the Python interfaces to be called at runtime.

              rh-ee-jgroenen Joseph Groenenboom
              rh-ee-jgroenen Joseph Groenenboom
              Votes:
              0 Vote for this issue
              Watchers:
              4 Start watching this issue

                Created:
                Updated:
                Resolved: