task: question-answering
Backend: sagemaker-training
Backend args: {'instance_type': 'ml.g4dn.2xlarge', 'supported_instructions': None}
Number of evaluation samples: 1000

Fixed parameters:

model_name_or_path: distilbert-base-uncased-distilled-squad
dataset:
- path: squad
- eval_split: validation
- data_keys: {'question': 'question', 'context': 'context'}
- ref_keys: ['answers']
- calibration_split: train
per_channel: False
calibration:
- method: minmax
- num_calibration_samples: 100
framework: onnxruntime
framework_args:
- opset: 11
- optimization_level: 1
aware_training: False

Benchmarked parameters:

quantization_approach: dynamic, static
operators_to_quantize: ['Add'], ['Add', 'MatMul']
node_exclusion: [], ['layernorm', 'gelu', 'residual', 'gather', 'softmax']

Evaluation

Non-time metrics

quantization_approach	operators_to_quantize	node_exclusion		exact_match (original)	exact_match (optimized)		f1 (original)	f1 (optimized)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	82.300	80.600	\|	87.232	86.097
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	82.300	80.600	\|	87.232	86.097
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	82.300	82.300	\|	87.232	87.232
`dynamic`	`['Add']`	`[]`	\|	82.300	82.300	\|	87.232	87.232
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	82.300	72.900	\|	87.232	79.964
`static`	`['Add', 'MatMul']`	`[]`	\|	82.300	54.500	\|	87.232	64.292
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	82.300	76.900	\|	87.232	83.014
`static`	`['Add']`	`[]`	\|	82.300	59.800	\|	87.232	69.217

Time metrics

Time benchmarks were run for 15 seconds per config.

Below, time metrics for batch size = 1, input length = 32.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	47.87	7.23	\|	20.93	138.40
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	48.10	7.14	\|	20.80	140.13
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	43.83	17.16	\|	22.87	58.33
`dynamic`	`['Add']`	`[]`	\|	34.13	17.02	\|	29.33	58.80
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	35.07	9.21	\|	28.53	108.53
`static`	`['Add', 'MatMul']`	`[]`	\|	48.27	11.62	\|	20.73	86.13
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	34.11	19.23	\|	29.33	52.00
`static`	`['Add']`	`[]`	\|	48.54	21.18	\|	20.67	47.27

Below, time metrics for batch size = 1, input length = 64.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	59.92	12.60	\|	16.73	79.40
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	59.64	13.25	\|	16.80	75.47
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	60.13	29.65	\|	16.67	33.73
`dynamic`	`['Add']`	`[]`	\|	59.62	29.51	\|	16.80	33.93
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	58.94	15.13	\|	17.00	66.13
`static`	`['Add', 'MatMul']`	`[]`	\|	60.49	18.62	\|	16.53	53.73
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	43.32	28.00	\|	23.13	35.73
`static`	`['Add']`	`[]`	\|	44.19	32.72	\|	22.67	30.60

Below, time metrics for batch size = 1, input length = 128.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	73.39	26.56	\|	13.67	37.67
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	57.64	23.42	\|	17.40	42.73
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	64.04	50.14	\|	15.67	20.00
`dynamic`	`['Add']`	`[]`	\|	72.81	57.05	\|	13.80	17.53
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	70.57	27.59	\|	14.20	36.27
`static`	`['Add', 'MatMul']`	`[]`	\|	71.04	37.94	\|	14.13	26.40
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	57.65	57.95	\|	17.40	17.27
`static`	`['Add']`	`[]`	\|	71.66	58.67	\|	14.00	17.07

Below, time metrics for batch size = 4, input length = 32.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	72.11	21.80	\|	13.93	45.93
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	73.15	20.70	\|	13.73	48.33
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	72.05	53.68	\|	13.93	18.67
`dynamic`	`['Add']`	`[]`	\|	55.97	53.60	\|	17.87	18.67
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	70.46	24.88	\|	14.20	40.20
`static`	`['Add', 'MatMul']`	`[]`	\|	56.57	30.90	\|	17.73	32.40
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	62.38	53.64	\|	16.07	18.67
`static`	`['Add']`	`[]`	\|	60.19	67.29	\|	16.67	14.87

Below, time metrics for batch size = 4, input length = 64.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	121.20	40.12	\|	8.27	24.93
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	90.97	41.51	\|	11.00	24.13
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	120.85	106.50	\|	8.33	9.40
`dynamic`	`['Add']`	`[]`	\|	118.58	106.55	\|	8.47	9.40
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	120.57	54.25	\|	8.33	18.47
`static`	`['Add', 'MatMul']`	`[]`	\|	104.93	57.90	\|	9.60	17.33
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	90.85	110.46	\|	11.07	9.07
`static`	`['Add']`	`[]`	\|	120.57	103.62	\|	8.33	9.67

Below, time metrics for batch size = 4, input length = 128.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	172.14	94.78	\|	5.87	10.60
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	220.38	84.18	\|	4.60	11.93
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	221.22	221.37	\|	4.53	4.53
`dynamic`	`['Add']`	`[]`	\|	203.90	175.16	\|	4.93	5.73
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	192.63	113.82	\|	5.20	8.80
`static`	`['Add', 'MatMul']`	`[]`	\|	220.32	122.36	\|	4.60	8.20
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	220.58	207.51	\|	4.60	4.87
`static`	`['Add']`	`[]`	\|	221.94	246.87	\|	4.53	4.07

Below, time metrics for batch size = 8, input length = 32.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	112.67	43.26	\|	8.93	23.13
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	95.78	40.66	\|	10.47	24.60
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	117.38	104.28	\|	8.53	9.60
`dynamic`	`['Add']`	`[]`	\|	89.81	91.00	\|	11.20	11.00
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	89.14	52.09	\|	11.27	19.20
`static`	`['Add', 'MatMul']`	`[]`	\|	92.77	64.21	\|	10.80	15.60
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	119.10	114.43	\|	8.40	8.80
`static`	`['Add']`	`[]`	\|	119.28	127.79	\|	8.40	7.87

Below, time metrics for batch size = 8, input length = 64.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	215.03	78.03	\|	4.67	12.87
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	214.76	87.19	\|	4.67	11.53
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	216.48	162.64	\|	4.67	6.20
`dynamic`	`['Add']`	`[]`	\|	204.29	212.33	\|	4.93	4.73
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	215.47	104.45	\|	4.67	9.60
`static`	`['Add', 'MatMul']`	`[]`	\|	209.66	106.43	\|	4.80	9.40
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	166.13	220.92	\|	6.07	4.53
`static`	`['Add']`	`[]`	\|	214.69	209.01	\|	4.67	4.80

Below, time metrics for batch size = 8, input length = 128.

quantization_approach	operators_to_quantize	node_exclusion		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
`dynamic`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	407.90	151.49	\|	2.47	6.67
`dynamic`	`['Add', 'MatMul']`	`[]`	\|	407.34	154.55	\|	2.47	6.53
`dynamic`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	406.51	394.85	\|	2.47	2.60
`dynamic`	`['Add']`	`[]`	\|	309.53	445.24	\|	3.27	2.27
`static`	`['Add', 'MatMul']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	407.54	224.46	\|	2.47	4.47
`static`	`['Add', 'MatMul']`	`[]`	\|	408.14	236.94	\|	2.47	4.27
`static`	`['Add']`	`['layernorm', 'gelu', 'residual', 'gather', 'softmax']`	\|	309.91	357.87	\|	3.27	2.80
`static`	`['Add']`	`[]`	\|	310.00	406.54	\|	3.27	2.47

Evaluation

Non-time metrics

Time metrics

NSDT 3DConvert

UnrealSynth

DreamTexture.js