Refactor warmup logic and generator mock in OnnxDiscrepancyCheck

[tadani3]

Summary

Fix a warmup pollution bug in OnnxDiscrepancyCheck.compare_generation
(olive/passes/onnx/discrepancy_check.py) that caused a systematic off-by-one
misalignment in the ONNX Runtime GenAI generated-token list.

Bug

The GenAI warmup loop (introduced in fe69db6) reused the genai_tokens
accumulator across both the warmup and the real generation run. The generator
is correctly re-created for the real run, but the accumulator was not reset, so
warmup-generated tokens were already present in genai_tokens when the real run
started.

Because generation is deterministic, this shifted every generated token by one
position and produced misleading metrics:

• first_token_matches → True (warmup token equals the real first token)
• second_token_matches → False (real first token lands in the second slot)
• longest_common_token_sequence → collapsed to 1

The warmup break condition (len(genai_tokens) >= 3) also counted prompt tokens,
under-warming on prompts of 3 or more tokens.

Fix

• Use a throwaway warmup_tokens accumulator during warmup so genai_tokens is
  never touched until the real run — structurally mirroring how the transformers
  side already discards its warmup output.
• Change the warmup break condition to len(warmup_tokens) >= genai_prompt_token_count + 1
  so it counts only generated tokens.

The real generation block immediately following was already correct and is
unchanged.

Tests

• Updated test/passes/onnx/test_discrepancy_check.py TestCompareGeneration
  mocks so each og.Generator call (warmup + real run) replays the token stream
  from the start via an og.Generator.side_effect factory. This models real
  deterministic re-generation, where warmup and the real run produce the same
  tokens and the warmup output is discarded. Previously the mocks shared a single
  call_count across both generators, which encoded the buggy behavior.
• The two append_tokens assertions now target a shared mock so the "prompt is
  seeded for both warmup and real run" checks still hold.
• test_compare_generation_with_zero_max_new_tokens was left unchanged — its
  generator reports is_done immediately and never appended warmup tokens, so it
  was unaffected by the bug.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

Release notes

Fixed a bug in the OnnxDiscrepancyCheck pass where GenAI warmup tokens polluted
the generated-token comparison, causing incorrect first_token_matches,
second_token_matches, and longest_common_token_sequence results.

(Optional) Issue link

[Copilot]

Pull request overview

This PR fixes a GenAI warmup “token pollution” bug in OnnxDiscrepancyCheck.compare_generation that could misalign generated-token comparisons (systematically shifting GenAI tokens by one) and thereby corrupt first_token_matches, second_token_matches, and longest_common_token_sequence results.

Changes:

• Refactored GenAI warmup to append into a throwaway warmup_tokens list instead of genai_tokens, preventing warmup output from affecting the real run.
• Adjusted the warmup break condition to be based on generated-token count (via genai_prompt_token_count + 1 threshold) rather than a fixed total-token length.
• Updated unit tests to mock og.Generator with a side_effect factory so each generator instance (warmup + real) replays the deterministic token stream from the start, matching real behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
olive/passes/onnx/discrepancy_check.py	Prevents warmup-generated tokens from being appended into the “real run” accumulator and fixes warmup stopping criteria.
test/passes/onnx/test_discrepancy_check.py	Updates generator mocking to model deterministic re-generation per generator instance and keeps prompt-seeding assertions valid.