fix(qwen3-vl): fall back to hf_hub_download for mmraw preprocessor_config.json

[hubert-marek]

Problem

The raw-image (mmraw) layout path (db5058e) resolves the model's image geometry from preprocessor_config.json, but _load_preprocessor_config_json only checks local paths and the local HF cache (try_to_load_from_cache). Hosted env workers render models they never loaded locally, so hub-style ids always miss the cache and every image rollout fails:

RuntimeError: Qwen raw image layout could not find preprocessor_config.json for
'Qwen/Qwen3.6-35B-A3B'. Ensure the model is cached locally or set all image_*
layout fields explicitly in the renderer config.

…even when the file is publicly available on the Hub. Hit in production on worldsims gflights-booking RL runs (Qwen3.6-35B-A3B); currently worked around by passing the five explicit image_* layout fields through chat_template_kwargs in the run config.

Fix

On cache miss for a hub-style id, fall back to hf_hub_download(model_id, "preprocessor_config.json") — a few hundred bytes, lands in the HF cache, then memoized by the existing lru_cache. Offline/no-network workers fall through to the existing RuntimeError, whose message now also mentions Hub reachability alongside the explicit image_* config escape hatch.

Test

test_qwen3_vl_preprocessor_config_hub_download_fallback: cache-miss → fallback download succeeds and parses; offline download failure → RuntimeError preserved.

tests/test_client.py -k 'hub_download_fallback or raw_mode_render' → 2 passed

🤖 Generated with Claude Code

---

Note

Low Risk
Narrow change to config resolution for hub model IDs with existing error paths when offline; adds optional Hub network fetch on cache miss only.

Overview
Fixes Qwen3-VL mmraw image rollouts on workers that never cached the model locally: when preprocessor_config.json is missing from disk and try_to_load_from_cache misses, _load_preprocessor_config_json now falls back to hf_hub_download (small file, then existing lru_cache). Offline or failed downloads still raise the same RuntimeError, with an updated message that also points to Hub reachability and explicit image_* layout overrides.

Adds test_qwen3_vl_preprocessor_config_hub_download_fallback to cover successful hub fallback and preserved failure when download is unavailable.

^{Reviewed by Cursor Bugbot for commit f146935. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fall back to hf_hub_download in _load_preprocessor_config_json for Qwen3-VL when cache misses

Previously, _load_preprocessor_config_json in qwen3_vl.py would raise a RuntimeError if no cached preprocessor_config.json was found for a hub-style model ID. It now attempts hf_hub_download as a fallback before raising. The error message is also updated to mention that the model must be reachable on the Hugging Face Hub.

Changes since #82 opened

• Reformatted code style and documentation in _load_preprocessor_config_json utility function [f146935]

^{Macroscope summarized 91f3040.}

[macroscopeapp]

Approvability

Verdict: Approved

Straightforward bug fix adding a fallback to download preprocessor_config.json from Hugging Face Hub when local cache misses. The change is small, self-contained with proper error handling, and includes comprehensive test coverage.

^{You can customize Macroscope's approvability policy. Learn more.}

[hubert-marek]

@cursoragent fix the ruff and style here - make comments concise

[cursor]

You need to increase your spend limit or enable usage-based billing to run background agents. Go to Cursor

Dismissing prior approval to re-evaluate f146935