Support Blackwell GPUs (torch 2.7 + DGL 2.4 cu124)

[jackytamkc]

Summary

• Replace graph.adjacency_matrix().to_dense().shape[0] with graph.num_nodes() at the two pos_weight sites in scniche/trainer/_train.py. Same value, no N×N dense allocation, and avoids the torch-pinned libdgl_sparse_pytorch_<X>.so load that fails on torch 2.7 with the DGL 2.4 cu124 wheel.
• Add a "Blackwell GPUs" subsection to the README documenting the install order for sm_120 hardware (RTX PRO 6000, RTX 50-series). Existing CUDA 11.3 instructions are kept intact — this is purely additive.

Background

Blackwell GPUs (compute capability sm_120) require PyTorch built against CUDA 12.8. No DGL wheel is currently published for cu128, but the DGL 2.4 cu124 wheel's CUDA kernels are ABI-compatible with torch 2.7 and JIT-forward to sm_120. The only blocker for scNiche on this stack is DGLGraph.adjacency_matrix(), which tries to dlopen a torch-version-pinned sparse library (libdgl_sparse_pytorch_2.4.so) and crashes against torch 2.7. Both call sites in scNiche only use the result to read N, so swapping to num_nodes() is functionally equivalent and removes the dependency on the sparse library.

Backward compatibility

• DGLGraph.num_nodes() has been part of the public DGL API since 0.5, so existing users on DGL 1.1.0+cu113 are unaffected (same integer, faster).
• README changes are additive — the existing CUDA 11.3 install path is unchanged.

Test plan

• Install on Ampere/Ada with the existing cu113 instructions; verify Runner.fit and Runner_batch.fit still produce the same pos_weight and embeddings as before.
• Install on Blackwell (sm_120) following the new README section; verify both Runner.fit and Runner_batch.fit train end-to-end without the libdgl_sparse_pytorch load error.
• pip check warning about dgl requires torch<=2.4.0 is expected on the Blackwell stack and documented in the README.

🤖 Generated with Claude Code

[ProDong0512]

Thank you for your comments. We will consider testing it and add it into our next version.