feat: add mini Qwen3-VL model for local e2e RL testing

[samsja]

Summary

• Adds scripts/mini_vlm.py — creates a tiny ~44M param Qwen3-VL model with random weights for local e2e RL testing, following the same pattern as scripts/mini_moe.py
• Updates is_vlm_model() to also detect VLM models from local paths by reading config.json model_type, so mini/custom checkpoints work without matching the Qwen/Qwen3-VL* naming convention

Mini model specs

Component	Config
Text	hidden=256, intermediate=512, layers=4, heads=4, kv_heads=2, head_dim=64
Vision	hidden=128, intermediate=256, depth=4, heads=4, out_hidden=256
Total	~44M parameters

Usage

# Create and verify
uv run python scripts/mini_vlm.py --output-dir ./mini-qwen3-vl

# Verify only
uv run python scripts/mini_vlm.py --output-dir ./mini-qwen3-vl --verify-only

Test plan

• Script creates model, saves with tokenizer+processor
• Verification passes: text-only and multimodal forward passes produce correct logit shapes
• is_vlm_model() detects local path models via config.json model_type
• Existing tests unaffected

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches VLM detection used in model loading/orchestration; misclassification or unexpected config.json contents could route runs down VLM-specific loading/freezing paths. The rest is an additive local testing script.

Overview
Adds scripts/mini_vlm.py to create a small random-weight Qwen3-VL checkpoint, copy a compatible tokenizer/processor, and sanity-check both text-only and multimodal forward passes on CPU.

Updates is_vlm_model() to recognize local/custom checkpoints by reading config.json and checking model_type (in addition to the existing Qwen/Qwen3-VL* name pattern), enabling VLM-specific codepaths to work with mini models stored on disk.

^{Written by Cursor Bugbot for commit 951b442. This will update automatically on new commits. Configure here.}