Split tiny-model generation into per-model scripts

Replace the monolithic scripts/generate_tiny_models.py (one 437-line file, a big tuple-driven loop plus a growing if issubclass(...) ladder for VLMs) with a per-model layout:

scripts/generate_tiny_models/
├── _common.py                      # shared helpers: push_to_hub, smoke_test, print_config_diff, ...
├── for_causal_lm/                  # *ForCausalLM + GPT-2 LM head + PEFT + "small" vLLM variants
├── for_sequence_classification/    # reward models
└── for_conditional_generation/     # VLMs + T5 + Bart (encoder-decoder)

One script per tiny model. Each script is fully self-contained. Shared logic (push, smoke test, diff, weight init) stays in _common.py.

Why

• The old file was bloated: 437 lines of tuple-driven loops stacked with one-off special cases for nearly every VLM (Qwen2.5-VL out_hidden_size, Qwen3-VL layer_types deletion, Qwen3.5 linear-attn fp32 cast, Gemma4 in-place mutation, llava-v1.6 dtype hotfix, …).
• To re-push a single model you had to manually comment out all the others.
• It didn't pin the transformers version
• It didn't run any smoke test — a broken config only failed later when CI tried to load the tiny model.

With one file per model, each script reads top-to-bottom in 20–50 lines. Model-specific quirks stay scoped to the model that needs them.

New features added to every script

• Smoke test. Before pushing, each script runs a minimal forward pass on a tiny dummy input. This catches config misspecification at generation time rather than when CI first imports the tiny model.
• Config diff vs reference. print_config_diff(MODEL_ID, model) prints every flat-key difference between the reference Hub config and the tiny model's config before push. Makes it obvious when a shrink kwarg was silently ignored or when an unexpected field drifted.
• Dtype pattern check. check_dtype_pattern(MODEL_ID, model) reads the reference safetensors header via the Hub API (no weight download) and flags any tensor whose dtype diverges from the reference — catches cases like models with mixed-precision weights (e.g. fp32 norms inside a bf16 checkpoint).
• Exact transformers version pin. Each script declares TRANSFORMERS_VERSION = "X.Y.Z" and calls check_transformers_version(...) which raises unless the installed version matches exactly. The pinned value is max(introduction_version, trl_floor=4.56.2). Rationale: transformers is backward-compatible (a checkpoint saved by X loads on any ≥ X) but not forward-compatible; TRL CI runs against the floor, so tiny models must be saved with the oldest version that supports them to avoid config-field drift. Exact match prevents accidental regenerations with a newer transformers from silently breaking min-version CI.
• --create-pr flag. When a tiny model already exists on the Hub, the default is to skip the push. Passing --create-pr opens a single PR against the existing repo instead (all artifacts bundled into one commit via HfApi.create_commit), so updates can be reviewed before landing on main.

How to run

# Regenerate a single tiny model (from repo root)
python -m scripts.generate_tiny_models.for_causal_lm.qwen3_for_causal_lm

# Open a PR against the existing Hub repo instead of skipping
python -m scripts.generate_tiny_models.for_causal_lm.qwen3_for_causal_lm --create-pr

See scripts/generate_tiny_models/README.md for full documentation.

Scope

This PR is refactor-only — the Python logic for each tiny model is preserved exactly. No tiny model on the Hub is regenerated by this PR; the existing Hub repos remain the source of truth for CI.

Follow-up PRs will use the new scripts to regenerate and push individual tiny models where the existing Hub checkpoint drifts from the reference (e.g. non-size config fields defaulting to wrong values, missing upstream-added fields, quantization parity, etc.). Each regeneration is one PR per tiny, with a refs/pr/N override in tests/conftest.py until merged on the Hub.