| I need you to implement Kimi Delta Attention forward (chunk form) for the RTX PRO 6000 (SM120 Blackwell, GDDR7, 1.8 TB/s). The reference implementation is in reference.py and your kernel goes in solution.py. You can make whatever mess you want in this directory — scratch files, intermediate kernels, build artifacts, profiling traces — but the final answer has to be in solution.py with the same Model, get_inputs, and get_init_inputs interface as reference.py. | |
| The op is the chunk-parallel KDA forward from the FLA library: q and k of shape (B, T, H, K) in bf16, v of shape (B, T, H, V) in bf16, g of shape (B, T, H, K) in fp32 (per-channel log-decay with in-chunk cumsum already applied), beta of shape (B, T, H) in bf16, scale a python float, chunk_size 64, no initial state, no final state. You return o of shape (B, T, H, V) in bf16. Correctness tolerance is 0.05 abs/rel — the long recurrence accumulates more error than a single GEMM so the bar's a bit looser than default bf16. The shapes you have to handle are B=2 T=1024 H=8 K=128 V=128 (short-context training step), B=2 T=2048 H=8 K=128 V=128 (the headline shape from the Kimi Linear paper), B=1 T=4096 H=8 K=128 V=128 (long context that stresses the inter-chunk recurrence), and B=1 T=2048 H=4 K=128 V=128 (thin batch decode). | |
| This needs to be a real custom kernel — the whole point of the problem is to write the chunk-parallel attention yourself, not call FLA's existing implementation. Don't import or call fla.ops.kda, fla.ops.chunk_kda, chunk_kda, fused_recurrent_kda, naive_chunk_kda, or naive_recurrent_kda. The intended path is CUTLASS CuTe on SM120 but Triton, CUDA C++ via load_inline, or inline PTX are also fine if you prefer. Anything you're uncertain about, look up PTX docs, clone CUTLASS or FLA or other reference repos, read library headers, and investigate. | |
| Your flywheel is implement, profile (ncu, nsys, torch.profiler — whatever's useful) and time it with benchmark.py, verify correctness by running `python check.py` and reading the output, then iterate. Don't substitute your own one-off correctness snippets for check.py — it iterates over every shape, your spot-check almost certainly won't. If `python check.py` hasn't printed PASS, you're not done. Take as long as you need to actually push the number up. | |