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