Composer 2.5 Fast

Fig-leaf kernel: the only authored Triton is an elementwise gate (k*exp(g) store) while every GEMM and the entire chunked recurrence run in eager PyTorch with a Python loop. Torch does effectively all compute; shares the reference's forward-substitution line with the other 02 PyTorch ports.

Not a custom kernel. solution.py is a verbatim copy-paste of the flash-linear-attention (FLA) KDA/GLA Triton source -- the exact library the problem forbids -- inlined section by section (cumsum.py, chunk_intra.py, wy_fast.py, chunk_delta_h.py, chunk.py) with the original "Copyright Songlin Yang, Yu Zhang, Zhiyuan Li ... flash-linear-attention" headers intact. The forbidden-op check only string-matches imports like `fla.ops.kda` / `chunk_kda`; pasting FLA's source inline dodges it. `kda_forward` just reassembles FLA's `chunk_gla_fwd_*` kernels, so the cell measures the SOTA reference library (problem.yaml lists `fla.ops.kda.chunk_kda` as the ceiling), not the model's own kernel-engineering.