https://github.com/triton-lang/triton-ext

https://github.com/flagos-ai/flagtree/tree/triton_v3.5.x/python/triton/experimental/tle

https://github.com/meta-pytorch/tritonbench/pull/498 Enable TileIR for FA

Feature 'add.f32x2' requires .target sm_100 or higher

heuristics_line = heuristics_line.replace("@triton.jit", "@triton_runner.jit")

https://docs.nvidia.com/cuda/cutile-python/data.html#data-types ，但是https://github.com/triton-lang/triton/blob/main/python/tutorials/10-block-scaled-matmul.py

(bsz, seq_len, num_q_heads, head_dim) transpose(1, 2) 成了 [bsz, n_q_head, seq_len, head_dim]