[Perf] Compiling drafter model for spec decoding

[chelsea0x3b]

Purpose

Compiles the drafter model used for speculative decoding. Also cleaning up the compilation logic a bit

Benchmark speculative decoding model with 4 spec tokens vllm bench serve on B200:

vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct -tp 4 --tokenizer-mode auto --speculative-config '{"method": "qwen3_next_mtp", "num_speculative_tokens": 4}' --no-enable-chunked-prefill

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct   --endpoint /v1/completions --dataset-name random --random-input 2048  --random-output 1024 --max-concurrency 256 --num-prompt 1000

Ran the benchmarks 4 separate times and took averages:
Main branch: 25912.5375 tok/s
This PR: 26184.05tok/s

For a consistent 1% boost

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request introduces a performance optimization by compiling the drafter model used for speculative decoding. This is achieved by refactoring the model compilation logic from load_model into a new _compile_model helper method, which is then applied to both the main model and the drafter model. The refactoring improves code organization and readability. The changes appear correct and well-implemented, and the provided benchmarks indicate a consistent performance boost.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @coreylowman.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[benchislett]

Would this happen to enable full cuda graphs for the draft model? If so, how would that work?

I believe we currently force piecewise for the drafter, but I still don't fully understand how many obstacles remain to unblock full-graphs for the draft model. See here for more context #23679.