[MagicMTP] New Feature.

[chenaoxuan]

What this PR does / why we need it?

MagicMTP Key Project: Introduce the 'block verify' algorithm into the draft token rejection sampling logic, taking into account the correlation between generated draft tokens to improve draft token acceptance rate and increase vllm_ascend runtime throughput.

• vLLM version: v0.11.0rc3
• vLLM main: vllm-project/vllm@83f478b

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request introduces a 'block verify' algorithm for rejection sampling to improve performance. The changes are mainly in rejection_random_sample_pytorch.

My review identified two critical issues in the new implementation that could lead to runtime errors:

1. A potential UnboundLocalError due to an uninitialized variable last_accepted_token_pos.
2. A potential ZeroDivisionError when calculating pi as draft_prob can be zero.

I have provided code suggestions to fix both issues. Please review them carefully.

[gemini-code-assist]

There is a potential ZeroDivisionError at line 430. draft_prob can be 0, but it is used as a divisor before being checked. The division should be guarded by a check for draft_prob > 0.

            if draft_prob > 0:
                pi = min(pi * target_prob / draft_prob, 1.0)
            else:
                pi = 0.0

            if pi >= uniform_prob:
                last_accepted_token_pos = pos
                rejected = False
            else:
                rejected = True

[chenaoxuan]

MTP>=3时强制开启MagicMTP优化逻辑：块校验（Block Verify）。
吞吐提升10%，保持测试精度不变。
绕过greedy校验逻辑（当前版本greedy校验导致性能劣化）。