Fix Robometer RFM tokenizer and processor parity

[Hynek Kydlíček (hynky1999)]

Summary

Fix Robometer/RFM parity by:

• using the canonical Qwen3-VL tokenizer plus Robometer special tokens before tokenizer caching
• defaulting the RFM pooler step_tag_id to the official <|prog_token|> id 151673
• using vLLM's cached processor loader instead of constructing Qwen3VLProcessor per request

This keeps the fix scoped to RFM/Robometer and removes the per-request processor reload that caused slow /pooling bursts.

Validation

• uvx ruff check vllm/model_executor/models/config.py vllm/model_executor/models/rfm.py vllm/tokenizers/hf.py vllm/tokenizers/registry.py --ignore SIM910
• git diff --check HEAD
• Modal parity run against fresh official Robometer:
  • official run: ap-VMBp6Wp2qPMVte91vZKIqW
  • patched vLLM run: ap-zc4qWvAwmBLYz9byAaUdbp
  • progress max abs diff: 0.0018651148415796959
  • success max abs diff: 0.0029832334522436367
• Modal H100 perf smoke:
  • run: ap-l1doOmDvRacBKeqDVDHDe7
  • warmup: 16.83s
  • post-warmup 4-request burst mean: 90ms
  • post-warmup max: 122ms

Focused pytest was not run locally because this checkout's .venv does not have pytest installed.

Duplicate Check

Checked open PRs in macrodata-labs/vllm for:

• RFM Robometer special tokens
• Robometer step_tag_id

No matching open PRs were found.

AI assistance was used to prepare this change.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3694502539

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Handle null architectures before tokenizer caching

For models whose HF config sets architectures explicitly to None, this check raises TypeError: argument of type 'NoneType' is not iterable before any tokenizer can be loaded; getattr(..., ()) only covers the attribute being absent, not the common explicit-null case (there are existing vLLM test configs with architectures=None). This makes unrelated non-RFM models fail during cached_tokenizer_from_config; coerce with getattr(..., None) or () before the membership test.

Useful? React with 👍 / 👎.