results.tsv

exp	commit	latency_s	std_s	max_abs_diff	psnr	quality_ok	peak_mem_gb	status	desc
0	06b97d1	14.4666	0.0148	0.0000	126.02	True	53.98	keep	baseline fp32 eager (Wan 2.2 VAE)
1	fb234ed	8.4228	0.0070	0.0708	61.14	True	27.03	keep	bf16 eager
2	a065031	8.2565	0.0080	0.0844	61.29	True	27.03	keep	bf16 + compile decoder
3	f40cb59	8.1545	0.0073	0.0667	61.14	True	24.87	keep	bf16 + channels_last_3d eager
4	4c8888f	8.1047	0.0099	0.0922	61.29	True	27.06	keep	bf16 + compile max-autotune-no-cudagraphs
5	24bf28f	8.1293	0.0249	0.0667	61.14	True	24.87	discard_slower	bf16 + channels_last + cudnn.benchmark
6	91cafd4	8.0205	0.0162	0.0667	61.14	True	24.87	keep	bf16 + channels_last + bf16-native upsample
7	b0eba8b	8.0015	0.0107	0.0667	61.14	True	24.90	keep	full-decode CUDA graph
8	4add1d9	7.4192	0.0015	0.0667	61.14	True	22.91	keep	native-spatial-pad conv (avoid F.pad copy)
9	f15c08c	6.6577	0.0054	0.0771	61.25	True	22.91	keep	compile decoder for elementwise fusion + graph
10	ff30218	6.6431	0.0117	0.0892	61.26	True	22.98	keep	compile max-autotune + graph
11	6161acc	nan	nan	0.1015	61.25	False	nan	discard_quality	drop channels_last (compile+graph manages layout)
12	f8aae3c	6.6519	0.0091	0.0892	61.26	True	22.98	discard_slower	cudnn.benchmark retry (native-pad regime)
13	485204f	6.6415	0.0055	0.0911	61.26	True	22.98	keep	inductor coordinate_descent_tuning
14	efe6802	4.8156	0.0042	0.0593	61.89	True	32.72	keep	break-free decoder (per-module cache, full fusion)
15	8f92e20	5.6533	0.0020	0.0610	61.89	True	36.72	discard_slower	break-free, no channels_last (inductor layout)
16	6fce889	4.7962	0.0028	0.0688	61.89	True	32.66	keep	fuse unpatchify into compiled decoder
17	cb72dd3	4.7982	0.0092	0.0693	61.89	True	32.59	discard_slower	break-free default compile mode
18	832e916	4.7911	0.0027	0.0610	61.89	True	32.66	keep	break-free, no coordinate_descent