Fix CUDA error when docking with flexible residues

[caic99]

Summary

Fixes #159 — CUDA error 700 (cudaErrorIllegalAddress) when using --flex for flexible receptor residues.

The CUDA kernel completely lacked support for flexible residues, causing crashes through five interrelated bugs:

• Blocking assert: assert(m.num_other_pairs() == 0) in 4 locations in monte_carlo.cu immediately crashed when flex residues created other_pairs (flex-flex interaction pairs)
• Missing other_pairs in CUDA: Flex-flex interaction pairs were never stored in GPU structures or evaluated in the scoring kernel
• Buffer overflow: MAX_NUM_OF_FLEX_TORSION = 1 was too small for real flex residues (which can have 20+ torsions across multiple residues), causing memory corruption when copying torsion data
• Wrong Config grouping: Ligand classification in main.cpp ignored receptor flex atoms/torsions, placing combined models (e.g., 165 atoms) into undersized Config groups (e.g., SmallConfig with max 40 atoms), causing out-of-bounds GPU memory access
• Missing flex conf in results: cuda_to_vina didn't initialize flex conformation data, causing segfault during CPU-side pose refinement when model.set() was called

Changes

File	Changes
kernel.h	Increase MAX_NUM_OF_FLEX_TORSION 1→24 in all Config structs; add other_pairs structs to m_cuda_t and m_cuda_t_<Config>
monte_carlo.cu	Remove 4 blocking asserts; add other_pairs GPU data copying; fix flex torsion copy to flatten all residues
warp_ops.cuh	Evaluate other_pairs in m_eval_deriv_warp; copy other_pairs in m_cuda_init_with_m_cuda_warp
main.cpp	Account for receptor flex atoms/torsions in ligand Config group classification
vina.h	Initialize flex conf from model before pose refinement
CMakeLists.txt	Add sm_100 (Blackwell/B200) to CMAKE_CUDA_ARCHITECTURES

Limitations

This fix enables flex docking to run without crashing and correctly evaluates flex-flex interaction energy (other_pairs). However, flex torsions are not yet optimized during the GPU Monte Carlo search — they remain at their initial values and are refined only during the CPU-side post-processing step. Full flex torsion optimization in the CUDA kernel (flex tree coordinate updates, flex torsion mutation, flex derivatives) would be a follow-up enhancement.

Test plan

• Reproduced original CUDA error 700 with the issue reporter's test files (2am9 receptor with 21 flex residues)
• Verified fix produces docking output (affinity: -3.467 kcal/mol, EXIT_CODE=0)
• Verified normal (non-flex) docking regression: 5798 ligands in 2.6s, identical results

🤖 Generated with Claude Code