# GLM-5.1 FP8 dense-linear local shapes at TP=64 (n_out n_in per line).
# Collected from `Couldn't find tuned sizes` warnings on a v6e server start.
# block_k=128, x_q_dtype=bfloat16 (activation not quantized), w_q_dtype=float8_e4m3fn.
#
# n_out < 256 (e.g. 32, 128) are omitted: the kernel rejects out_block_size < 256
# so no valid candidate exists; those layers stay on the safe-clamped fallback.
256 2048
576 6144
2048 6144
4096 2048
6144 256
6144 2048
