[Proposal] Add Falcon-H1 parallel hybrid adapter (FalconH1ForCausalLM)

[jlarson4]

Proposal

Add a TransformerBridge adapter for FalconH1ForCausalLM (TII Falcon-H1), a hybrid that runs attention and Mamba-2 in parallel.

Motivation

In each Falcon-H1 block, attention heads and Mamba-2 heads process the same input side by side and their outputs are concatenated. Because both paths see the same token, researchers can ablate one and compare, which is a clean way to study what attention adds over a state-space layer. The family ships very small checkpoints (down to a 90M and a 0.5B), so it is cheap to verify.

This runs hand-in-hand with #1402 in extending our capabilities to research Mamba-base state-space models.

Gap scan (2026-06-18): 14 models, ~171K downloads.

Pitch

Expose the attention and Mamba-2 sub-paths in each block as separate hookable streams.

• Claude Code users can scaffold with /add-model-support tiiuae/Falcon-H1-0.5B-Base.
• Register at the four sites listed in contributing.md.
• Verify smallest-first: tiiuae/Falcon-H1-Tiny-90M-Instruct, then tiiuae/Falcon-H1-0.5B-Base.

Additional context

• Paper: https://arxiv.org/abs/2507.22448
• Checkpoint: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
• Found via the hf_scraper architecture-gaps pass (2026-06-18).

Checklist

• I have checked that there is no similar issue in the repo (required)

[udapy]

Hi @jlarson4,
I would like to take this issue.

[jlarson4]

Certainly! You are now assigned