[khanhnd61-vr]

Ports allenai/MolmoAct2-LIBERO as a hand-written arch, following GR00T / π0.5. GGUF: vrfai/molmoact2-libero-gguf.

Changes

• src/models/molmoact2.cpp: SigLIP tower + 2×2 attention pooling, Qwen3-style LM, flow-matching expert that cross-attends to each LM layer's K/V. Reuses EncStack and Qwen3LM.
• src/modules/qwen3_lm.*: optional Qwen3KV* out-param on block() to return each layer's K/V. Off by default; GR00T N1.7 and VLA-JEPA outputs are unchanged.
• scripts/convert_molmoact2_to_gguf.py: fp32 checkpoint → bf16 GGUF.
• Client: eval/client/molmoact2_prompt.py builds the prompt (task, binned state, image slots). Used by vla_cpp_client.py, tokenize_prompt.py and run_libero.sh -m molmoact2.
• README: MolmoAct2 replaces ACT in the support matrix and acknowledgements.

Verification

Max |Δ| of the action chunk vs fp32 PyTorch:

	bf16 (default)	--weight-dtype f16 --flash-attn
CPU	1.7e-3
CUDA (RTX 3060)	1.4e-3 (446 ms)	9.2e-4 (328 ms)
Metal (M4)	5.6e-4

• f32 weights vs the bf16-rounded reference: 7.7e-6.
• LIBERO smoke (task 0 of each suite, 3 episodes): 12/12.
• ctest passes.