MCPcopy Create free account
hub / github.com/NVIDIA/FasterTransformer / forward

Method forward

examples/pytorch/longformer/model.py:83–104  ·  view source on GitHub ↗
(self, *args, **kwargs)

Source from the content-addressed store, hash-verified

81 self.hf_plugin_mode = is_plugin
82
83 def forward(self, *args, **kwargs):
84 encoder_in = args[0]
85
86 if self.hf_plugin_mode:
87 # In this mode, assume that HuggingFace's LongformerModel.encoder has been
88 # substituted to this class's instance
89 extended_attention_mask = kwargs['attention_mask']
90 local_attn_mask = torch.zeros_like(extended_attention_mask)
91 local_attn_mask[extended_attention_mask > -10000.] = 1.0
92 global_attn_mask = torch.zeros_like(extended_attention_mask)
93 global_attn_mask[extended_attention_mask > 0.] = 1.0
94 output = self.ft_encoder.forward(encoder_in, local_attn_mask, global_attn_mask, self.all_weight, 0)
95 return LongformerBaseModelOutput(
96 last_hidden_state=output,
97 hidden_states=None,
98 attentions=None,
99 global_attentions=None,
100 )
101 else:
102 local_attn_mask = args[1]
103 global_attn_mask = args[2]
104 return self.ft_encoder.forward(encoder_in, local_attn_mask, global_attn_mask, self.all_weight, 0)

Callers 1

mainFunction · 0.45

Calls

no outgoing calls

Tested by

no test coverage detected