What does this PR do?

Adds two new tiny-model fixture generation scripts for Qwen3.5 — sibling -Think and -NoThink variants alongside the existing tiny-Qwen3_5ForConditionalGeneration fixture. Each is sourced from the Qwen3.5 release, whose bundled tokenizer ships the corresponding default thinking behavior: Qwen/Qwen3.5-4B for the thinking-enabled variant, Qwen/Qwen3.5-0.8B for the thinking-disabled variant.

The architecture class, tiny-config overrides, fp32 linear-attn cast, and _common.py push flow are unchanged from the existing script — the only differences are MODEL_ID and the suffix arg to push_to_hub.

Note: scope and roadmap

Strictly additive — scripts only. The existing tiny-Qwen3_5ForConditionalGeneration fixture is untouched, and no tests reference the new fixtures yet. Planned follow-ups:

• PR 2: add Qwen3.5 training templates corresponding to the two variants.
• PR 3: migrate existing tests off the legacy fixture and retire it gracefully. Gated on a separate issue, I'll open a proposal for the test refactor against the new fixtures — work on PR 3 starts only once maintainers approve the refactor plan in that issue.

Note: push needs to be run by a maintainer

I don't have write access to trl-internal-testing. Per the agreement in #5471, @qgallouedec will run push_to_hub to materialize the fixtures. The Colab dry-runs below show both scripts build cleanly through print_config_diff.

Validation (Colab A100, dry-run)

NoThink (sourced from Qwen/Qwen3.5-0.8B):

[smoke_test] Qwen3_5ForConditionalGeneration: OK (output shape (2, 82, 248320))
[dtype_check] Qwen/Qwen3.5-0.8B: all matched tensors have the reference dtype
[config_diff] Qwen/Qwen3.5-0.8B vs tiny (10 differences)
  text_config.full_attention_interval              4                                  → 2
  text_config.hidden_size                          1024                               → 16
  text_config.layer_types                          ['linear_attention', 'linear_atten → ['linear_attention', 'full_attenti
  text_config.num_attention_heads                  8                                  → 4
  text_config.num_hidden_layers                    24                                 → 2
  vision_config.depth                              12                                 → 2
  vision_config.hidden_size                        768                                → 16
  vision_config.intermediate_size                  3072                               → 32
  vision_config.num_heads                          12                                 → 4
  vision_config.out_hidden_size                    1024                               → 16

Think (sourced from Qwen/Qwen3.5-4B):

[smoke_test] Qwen3_5ForConditionalGeneration: OK (output shape (2, 80, 248320))
[dtype_check] Qwen/Qwen3.5-4B: all matched tensors have the reference dtype
[config_diff] Qwen/Qwen3.5-4B vs tiny (11 differences)
  text_config.full_attention_interval              4                                  → 2
  text_config.hidden_size                          2560                               → 16
  text_config.layer_types                          ['linear_attention', 'linear_atten → ['linear_attention', 'full_attenti
  text_config.num_attention_heads                  16                                 → 4
  text_config.num_hidden_layers                    32                                 → 2
  text_config.num_key_value_heads                  4                                  → 2
  vision_config.depth                              24                                 → 2
  vision_config.hidden_size                        1024                               → 16
  vision_config.intermediate_size                  4096                               → 32
  vision_config.num_heads                          16                                 → 4
  vision_config.out_hidden_size                    2560                               → 16

Both runs: smoke test passes, dtype pattern matches the reference checkpoint, config diff shows only the intended tiny overrides (no key leakage). The extra num_key_value_heads line in the Think run reflects the genuine 0.8B vs. 4B upstream difference (4 → 2 override only shows up against the 4B source). Confirms the existing linear_attn.A_log / linear_attn.norm.weight fp32 cast block applies to both sources without modification.

Changes

• scripts/generate_tiny_models/for_conditional_generation/qwen3_5_for_conditional_generation_think.py: new script — MODEL_ID="Qwen/Qwen3.5-4B", pushes as tiny-Qwen3_5ForConditionalGeneration-Think.
• scripts/generate_tiny_models/for_conditional_generation/qwen3_5_for_conditional_generation_nothink.py: new script — MODEL_ID="Qwen/Qwen3.5-0.8B", pushes as tiny-Qwen3_5ForConditionalGeneration-NoThink.
• Each new script is a near-verbatim copy of the existing qwen3_5_for_conditional_generation.py

Part of #5471

cc: @qgallouedec