Uploaded image for project: 'Red Hat Enterprise Linux AI'
  1. Red Hat Enterprise Linux AI
  2. RHELAI-2669

(phase 1) Productize the 128k context window Granite v3.1

XMLWordPrintable

    • False
    • Hide

      None

      Show
      None
    • False
    • Not Selected
    • 100% To Do, 0% In Progress, 0% Done
    • Hide

      Targetting at least TP for RHEL AI 1.4 is conditional on the availability of the 128k base models.

      Show
      Targetting at least TP for RHEL AI 1.4 is conditional on the availability of the 128k base models.

      Feature Overview:

      This is phase 1 for the Support Granite 3.1 model with 128k context window

      Goal:

      • Fine-tune the 128k model using the default 8b 4k dataset (this might affect the effective context windows size)
      • Identify the effective context window to be used as the supported context window
        • The expectation is at least 32K context window to match the context window of the teacher model.
        • The ideal range would be at least 64K.

       

      Requirements:

      • Fine-tune the Granite 3.1 128k context model (Dec 2024) with the 8b 4k dataset
      • Validate and document effective context window
      • Identify any deviation in the performance of the final model
        • To move as GA, it should be within the margin of error 
      • Identify optimal batch size for training 

       
       
      Done - Acceptance Criteria:

       

      • [ ] InstructLab can fine-tune the 128k model using the 8b 4k dataset
      • [ ] Document the effective context window of the resulting model
      • [ ] Evaluate and compare the performance of the final model to a 4k fine-tuned model
      • [ ] Document and use optimal batch size during training (if required)

       

      Use Cases:

      Enterprise use cases that would benefit from a large context window include the following:

      • RAG
      • Summarization
      • Code generation
      • Tools use
      • Advanced reasoning

      Out of Scope :

      For phase 1 the creation of the 8b 128k dataset or SDG optimizations for large context windows are out of scope.

       

      Documentation Considerations:

      Document the support of large context window limits based on effective context window size.

       

      Questions to Answer:

      • Can the fine-tuning of the 128k be done with the existing batch size for phase 1 and then updated for phase 2, or does it require modification for phase 1?{}{}
      • Can we default to the 128k model, or are there circumstances in which we should maintain the 4k model?

       

      Background and Strategic Fit:

      To support the enterprise use cases required by customers, we need at least a 32k effective context window.

       

      Customer Considerations:

      • These changes should be transparent to the user-facing CLI flow

              wcabanba@redhat.com William Caban
              wcabanba@redhat.com William Caban
              Mustafa Eyceoz, Oleg Silkin
              Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

                Created:
                Updated: