1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
|
skip:
all:
# Difficult to setup accuracy test because .eval() not supported
- Reformer
# Fails deepcopy
- BlenderbotForConditionalGeneration
- GPTNeoForCausalLM
- GPTNeoForSequenceClassification
# Fails with even batch size = 1
- GPTJForCausalLM
- GPTJForQuestionAnswering
device:
cpu: []
control_flow:
- AllenaiLongformerBase
batch_size:
# TODO - Fails even after fake tensors
divisors:
AlbertForMaskedLM: 2
AlbertForQuestionAnswering: 2
AllenaiLongformerBase: 2
BartForCausalLM: 2
BartForConditionalGeneration: 2
BertForMaskedLM: 2
BertForQuestionAnswering: 2
BlenderbotForCausalLM: 8
# BlenderbotForConditionalGeneration : 16
BlenderbotSmallForCausalLM: 4
BlenderbotSmallForConditionalGeneration: 2
CamemBert: 2
DebertaForMaskedLM: 4
DebertaForQuestionAnswering: 2
DebertaV2ForMaskedLM: 4
DebertaV2ForQuestionAnswering: 8
DistilBertForMaskedLM: 2
DistilBertForQuestionAnswering: 2
DistillGPT2: 2
ElectraForCausalLM: 2
ElectraForQuestionAnswering: 2
GPT2ForSequenceClassification: 2
# GPTJForCausalLM : 2
# GPTJForQuestionAnswering : 2
# GPTNeoForCausalLM : 32
# GPTNeoForSequenceClassification : 2
GoogleFnet: 2
LayoutLMForMaskedLM: 2
LayoutLMForSequenceClassification: 2
M2M100ForConditionalGeneration: 4
MBartForCausalLM: 2
MBartForConditionalGeneration: 2
MT5ForConditionalGeneration: 2
MegatronBertForCausalLM: 4
MegatronBertForQuestionAnswering: 2
MobileBertForMaskedLM: 2
MobileBertForQuestionAnswering: 2
OPTForCausalLM: 2
PLBartForCausalLM: 2
PLBartForConditionalGeneration: 2
PegasusForCausalLM: 4
PegasusForConditionalGeneration: 2
RobertaForCausalLM: 2
RobertaForQuestionAnswering: 2
Speech2Text2ForCausalLM: 4
T5ForConditionalGeneration: 2
T5Small: 2
TrOCRForCausalLM: 2
XGLMForCausalLM: 4
XLNetLMHeadModel: 2
YituTechConvBert: 2
tolerance:
higher_training:
- MT5ForConditionalGeneration
# AlbertForQuestionAnswering fails in CI GCP A100 but error does not seem
# harmful.
- AlbertForQuestionAnswering
higher_max_autotune_training:
# DebertaForQuestionAnswering needs higher tolerance in Max-Autotune mode
- DebertaForQuestionAnswering
higher_inference:
- GPT2ForSequenceClassification
- RobertaForQuestionAnswering
higher_inference_cpu:
- LayoutLMForSequenceClassification
- GPT2ForSequenceClassification
cosine: []
accuracy:
skip:
large_models:
# Models too large to have eager, dynamo and fp64_numbers simultaneously
# even for 40 GB machine.
- DebertaV2ForMaskedLM
- BlenderbotForCausalLM
only_inference:
# Fails with dynamo for train mode
- M2M100ForConditionalGeneration
only_fp32:
- GoogleFnet
|
