Motivation

Part of Ngram series #21052
Following #22538. Verify that allowing dynamic spec token allocation across Trie and SAMs bring benefit

Modifications

Add the benchmark

• Added an end-to-end regression benchmark that compares three settings on the
  same prompt set: trieOnly (no external corpus loaded, draft tokens come only from Trie), samOnly (load strong matching external suffix-automaton corpus and rerun the same prompts), and samPlusDistractors (keep that matching SAM, then add extra irrelevant SAM corpora). The metric is avg_spec_accept_length, i.e. the average number of speculative tokens
  accepted per verify step.
• On the benchmark workload, samOnly improves accept length from 2.13 to 6.64 (3.12x vs. trieOnly).
• With 2 or 4 distractor SAM corpora, accept length stays at 5.90 / 5.92 (~89% of samOnly), showing the new trie/SAM ranking preserves the strong matching corpus instead of collapsing back toward the trie-only baseline.

Accuracy Tests

This is the accuracy test for #22538

Speed Tests and Profiling

This is the speed test for #22538

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.