- Remove unused local variable b in attention_backend.py - Remove unused variable rank0_sd in test_broadcast_state_dict.py - Remove unused imports across test files
- Default backend resolves to highest-priority available: flash -> cuda -> torch - attention() falls back per-call for training/fp32/unsupported head_dim - Re-apply index_copy_ for CUDA KV cache writes (index_put_ race mitigation)
- register built-in backends through BaseFactory - derive benchmark choices from registered backends - cover string selection and invalid backend names
- Add attention() functional entry delegating to active backend - GQA/MLA forward calls attention() instead of inline cache/SDPA - CUDA kernels support 2D/3D/4D mask via mask_h_stride field - CudaBackend.fwd_decode builds 2D padding mask for mixed seq_lens - KVCache.max_len precomputed in bind_tasks to avoid GPU sync - batch==1 decode short-circuits mask=None - Split tests into conftest, test_backend, test_backend_equivalence, test_kernel_mask - 440 tests pass, L20 decode 1.44-1.60x speedup vs torch native
ViperEkura