Porting Whisper model to n300

[saichandax]

• Measure and record current performance.
• Rebase the model to main, ensure the PCC = 0.99
• Port functionality to n300 card (single device)
• Provide Op Report
• Check Model into CI
• Implement Demo
• Implement Data Parallel on n300 #13400
• Implement perf with trace+2cqs for Whisper model #14728

Executive summary (as of Nov 12):

• Single device implementation is complete: (BLOCKED) due to pending approvals and CIs for #8339

  • PCC = 0.98
  • BS = 8
  • Torch ops: Conv1D (#13529), Slice (#13672), Reshape (#13889)
  • Perf numbers
    • E2E perf is 0.97 samples/sec
    • Device perf hangs on latest main, debugging in progress - draft PR #15257

• Data parallel implementation is complete: (BLOCKED) due to failing CIs #14334

  • PCC = 0.85 (To Debug)
  • BS = 16
  • Torch ops are same as in single device implementation.
  • Unable to load Model's conv1d weights on device - #14952
  • Check device perf and report numbers.

• Trace 2cqs implementation is blocked due to torch ops in the model.

ToDo:

• Debug hang issue in device perf for single device (n300) - #14800
• Data parallel:
  • Debug PCC drop (create an issue).
  • Unable to load Model's conv1d weights on device - #14952
  • Check device perf and report numbers
• To implement perf with trace+2cqs (currently blocked due to torch ops)

[kkeerthana0573]

Whisper Model is ported to n300 with the following torch ops:

1. Slice op - #13672
2. Add op - #13464
3. Conv1d op - #13529

Model Pcc is 0.989.
Demo Results :

• Accuracy for WhisperForAudioClassification - 0.75
• Accuracy for WhisperForConditionalGeneration - 0.73

Single Card Demo Pipeline - #8339