olivame a2cd34dd51 1. einsum('b i d, b j d -> b i j') → torch.bmm(q, k.transpose(-1,-2)) — 直接映射 rocBLAS batched GEMM ...

2. baddbmm 把 scale 融合进 GEMM，少一次 kernel launch
3. 第二个 einsum 同理换torch.bm
每一轮加速1到两秒

2026-02-08 18:54:48 +00:00

..

world_model_interaction_prompts

init

2026-02-08 03:29:15 +00:00

output.log

1. einsum('b i d, b j d -> b i j') → torch.bmm(q, k.transpose(-1,-2)) — 直接映射 rocBLAS batched GEMM

2026-02-08 18:54:48 +00:00

psnr_result.json

所有case的baseline，amd版本的ground truth都上传了

2026-02-08 09:42:14 +00:00

run_world_model_interaction.sh

init

2026-02-08 03:29:15 +00:00