benchmark decoding attention kernel with cudnn

[bjmsong]

Motivation

follow this pr， add cudnn

Modifications

Checklist

• Format your code according to the Contributor Guide.
• Add unit tests as outlined in the Contributor Guide.
• Update documentation as needed, including docstrings or example tutorials.

[merrymercy]

Can you share any results?

[bjmsong]

<style> </style>

head_num,batch_size,kv_len	SGLang[triton]	FlashInfer	cuDNN
32,1,64	265.22	270.34	430.08
32,1,128	265.73	271.36	429.57
32,1,256	275.46	285.70	440.32
32,1,512	297.98	299.01	449.54
32,1,1024	316.42	289.79	496.64
32,1,2048	346.11	293.89	522.24
32,1,4096	435.20	326.66	676.86
32,4,64	258.05	270.34	429.06
32,4,128	271.87	270.34	432.13
32,4,256	287.74	288.77	439.30
32,4,512	359.42	295.42	452.61
32,4,1024	428.03	330.75	547.84
32,4,2048	608.26	387.07	943.10
32,4,4096	1005.57	505.34	1775.62
32,16,64	288.77	270.34	442.37
32,16,128	331.78	279.04	447.49
32,16,256	420.86	314.37	565.25
32,16,512	610.30	374.78	949.25
32,16,1024	1003.52	492.54	1730.56
32,16,2048	1823.23	726.02	3349.50
32,16,4096	3307.52	1199.10	6626.30
32,64,64	579.07	311.30	563.20
32,64,128	730.11	367.62	1056.77
32,64,256	1064.96	493.57	1803.78
32,64,512	1806.34	734.21	3406.85
32,64,1024	3356.67	1218.56	6606.85
32,64,2048	6454.27	2183.17	13052.93
32,64,4096	12635.14	3829.76	26053.63
64,1,64	253.95	264.19	423.94
64,1,128	253.95	262.66	417.79
64,1,256	284.67	281.60	429.06
64,1,512	302.08	282.62	453.63
64,1,1024	343.55	288.77	475.14
64,1,2048	404.48	317.44	568.32
64,1,4096	592.90	366.59	979.97
64,4,64	263.68	264.19	430.08
64,4,128	285.70	265.22	429.06
64,4,256	321.54	276.48	459.78
64,4,512	477.18	303.10	555.01
64,4,1024	602.11	351.23	931.84
64,4,2048	992.26	455.68	1730.56
64,4,4096	1806.34	666.11	3369.47
64,16,64	380.93	277.50	464.90
64,16,128	459.78	305.15	592.90
64,16,256	624.64	351.23	980.99
64,16,512	989.18	460.80	1762.82
64,16,1024	1756.16	670.72	3357.70
64,16,2048	3312.64	1099.78	6560.77
64,16,4096	6398.98	1992.70	12994.56
64,64,64	960.51	369.66	973.82
64,64,128	1247.23	489.47	1938.43
64,64,256	1910.78	679.94	3522.05
64,64,512	3371.01	1101.31	6738.43
64,64,1024	6470.66	2007.04	13148.16
64,64,2048	12626.94	3844.10	25998.34
64,64,4096	24993.79	6835.71	47107.07

[zhyncs]

The hopper optimization will be released soon on FlashInfer's latest main branch. cc @yzh119

[zhyncs]

@bjmsong Nice work!!!

[zhyncs]

@bjmsong Is this data execution time, meaning the lower, the better?

[bjmsong]

@bjmsong Is this data execution time, meaning the lower, the better?

yeah, it's in microsecond, the lower, the better.

[zhyncs]

I'm just a bit confused about why cuDNN is even worse than Triton.

[bjmsong]

I will optimize the performance later.

[bjmsong]

• update：only measure the kernel execution time, excluding the data processing part.

<style> </style>

head_num,batch_size,kv_len	cuDNN	SGLang[triton]	FlashInfer
32,1,64	37.88	84.84	28.75
32,1,128	35.45	89.70	28.25
32,1,256	39.78	79.64	33.11
32,1,512	45.10	80.74	33.07
32,1,1024	55.44	82.06	34.19
32,1,2048	71.03	84.60	43.19
32,1,4096	103.06	134.33	77.41
32,4,64	34.28	84.74	28.31
32,4,128	38.45	88.59	29.74
32,4,256	52.45	90.46	33.21
32,4,512	70.65	83.49	43.79
32,4,1024	104.42	105.05	76.36
32,4,2048	166.11	172.79	139.27
32,4,4096	287.11	311.79	256.56
32,16,64	56.99	127.50	29.26
32,16,128	78.53	131.28	39.81
32,16,256	116.72	140.26	73.93
32,16,512	178.32	173.57	131.03
32,16,1024	305.11	294.56	237.77
32,16,2048	556.01	533.53	483.58
32,16,4096	1061.43	1020.02	954.83
32,64,64	165.24	445.24	75.97
32,64,128	229.19	452.60	119.45
32,64,256	366.27	478.52	223.89
32,64,512	618.63	567.01	461.61
32,64,1024	1132.89	1023.42	934.39
32,64,2048	2143.48	1957.95	1914.50
32,64,4096	3379.35	3144.74	2898.98
64,1,64	35.13	85.39	29.26
64,1,128	36.26	78.99	28.35
64,1,256	40.90	80.19	34.23
64,1,512	53.57	79.05	33.35
64,1,1024	68.45	80.47	44.00
64,1,2048	101.84	99.62	76.40
64,1,4096	160.52	167.03	130.98
64,4,64	43.80	81.62	28.39
64,4,128	52.91	82.30	29.14
64,4,256	69.67	86.65	35.52
64,4,512	102.96	105.07	69.38
64,4,1024	170.00	169.51	125.83
64,4,2048	294.75	300.52	234.91
64,4,4096	533.40	565.29	460.78
64,16,64	86.15	232.47	45.80
64,16,128	124.32	237.48	73.89
64,16,256	193.64	251.59	127.38
64,16,512	320.78	305.25	236.49
64,16,1024	575.09	531.26	484.83
64,16,2048	1077.27	994.54	960.10
64,16,4096	2073.99	1915.31	1941.34
64,64,64	281.46	867.22	133.79
64,64,128	413.01	877.68	224.82
64,64,256	679.76	919.92	453.97
64,64,512	1188.34	1082.00	930.13
64,64,1024	2208.71	1989.55	1891.40
64,64,2048	3383.27	3234.63	2957.94
64,64,4096	6510.72	6140.89	5833.15

[ispobock]

1. Did you consider warmup for these kernels before benchmark? Since the Triton kernel is JIT compiled, compilation overhead may be included.
2. It seems you only tested MHA here, while there are many popular models use GQA. Could you also include GQA in that benchmark?

[bjmsong]

Update：Add warmup & GQA

Here're some results:

<style> </style>

(batch_size,kv_len)	cuDNN	Triton	FlashInfer
(1,64)	31.67	78.47	27.47
(1,128)	33.89	76.19	27.52
(1,256)	39.62	76.14	31.54
(1,512)	44.29	78.13	31.89
(1,1024)	54.87	82.72	34.28
(1,2048)	70.26	84.93	42.92
(1,4096)	102.61	133.10	76.31
(4,64)	32.80	79.01	32.76
(4,128)	37.38	80.16	28.25
(4,256)	53.32	77.67	32.96
(4,512)	70.30	81.66	43.75
(4,1024)	103.56	104.41	79.21
(4,2048)	164.40	172.06	144.71
(4,4096)	285.75	314.03	251.61
(16,64)	55.95	125.68	29.11
(16,128)	78.14	131.68	39.18
(16,256)	113.02	140.26	76.06
(16,512)	177.33	170.94	131.64
(16,1024)	305.33	294.01	243.98
(16,2048)	557.51	534.84	471.28
(16,4096)	1055.98	1017.17	976.81
(64,64)	165.15	444.49	75.20
(64,128)	231.59	451.85	122.77
(64,256)	362.76	478.56	225.86
(64,512)	620.90	564.61	462.78
(64,1024)	1126.41	1025.39	944.30
(64,2048)	2146.21	1953.23	1918.42
(64,4096)	3307.69	3091.87	2898.67

<style> </style>

cuDNN	Triton	FlashInfer
31.86	84.42	31.54
34.55	79.58	31.16
36.42	82.85	37.53
40.71	81.03	36.46
45.70	87.36	36.70
51.72	80.57	39.14
61.44	82.22	43.16
34.12	87.26	32.73
34.91	81.41	34.45
40.57	84.48	38.14
45.47	78.70	38.05
56.83	82.04	47.40
73.38	83.88	56.81
107.25	117.13	89.00
37.31	84.59	32.47
39.06	81.56	32.97
55.57	80.72	33.84
73.59	87.54	49.32
108.06	108.73	82.31
171.88	179.98	136.74
296.34	320.65	250.79
70.56	136.65	54.47
91.39	141.97	62.13
128.15	150.03	89.96
195.93	216.71	140.90
322.75	344.60	247.24
581.26	594.42	472.30
1103.89	1083.44	884.71