[Core]Consolidate RoPE-related parsing into ModelArchitectureConfig

[charlotte12l]

Purpose

We are introducing model_arch_config #28454, which defines explicitly what kind of information vLLM engine need from hugggingface config/ user-defined config, so we could avoid hf_config/getattr(hf_config, xxx) got passing around in engine.

Before this PR, _get_and_verify_max_len in model.py directly accessed multiple HuggingFace config fields:

• rope_parameters
• original_max_position_embeddings
• model_max_length

By moving this logic into the convertor, we:

• Centralize all RoPE-related decisions in one place
• Reduce hf_config access in model.py

Behavior Differences/ Potential Bug Fix

Order of RoPE Scaling vs Caps
Before: RoPE scaling was applied AFTER sliding_window and tokenizer caps

1. Get base max_len from config
2. Apply sliding_window cap
3. Apply tokenizer cap
4. Apply RoPE scaling (yarn override + scaling_factor)

After: RoPE scaling is applied BEFORE caps

1. Get base max_len from config
2. Apply RoPE scaling (yarn override + scaling_factor) <- Moved here
3. Apply sliding_window cap
4. Apply tokenizer cap

Rationale: The new order is more correct. Sliding_window is a hard architectural limit that shouldn't be scaled up. The old behavior could result in incorrect values like sliding_window(2048) * factor(8) = 16384.

Test Plan

Select several models, run and get max_model_len groundtruth before this PR, compare if the values after this PR is the same.

pytest  tests/config/test_model_arch_config.py

Test Result

passed

---

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 is a well-executed refactoring that consolidates RoPE-related configuration parsing into the ModelArchitectureConfig. By moving this logic from vllm/config/model.py to vllm/transformers_utils/model_arch_config_convertor.py, it centralizes all RoPE-related decisions, improving maintainability and reducing direct access to HuggingFace configs in model.py. A key improvement is the correction of the application order for RoPE scaling and context length caps, which resolves a potential bug where hard architectural limits like sliding_window were incorrectly scaled. The changes are clean, logical, and enhance the robustness of the configuration handling. The introduction of the MaxModelLenInfo named tuple provides a clear and well-documented interface for maximum length information. Overall, this is a high-quality contribution.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @charlotte12l.

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