MCPcopy Create free account
hub / github.com/NVIDIA/TensorRT / __init__

Method __init__

demo/HuggingFace/T5/export.py:121–131  ·  view source on GitHub ↗
(self, decoder, lm_head, config, is_trt = False)

Source from the content-addressed store, hash-verified

119 """
120
121 def __init__(self, decoder, lm_head, config, is_trt = False):
122 super().__init__()
123 self.decoder = decoder
124 self.lm_head = lm_head
125 self.config = config
126 # HuggingFace's beam search requires to set self.device. Set it to avoid application crash
127 self.device = torch.device('cuda')
128 # Use hardcoded value to extend compatibility with older HF versions.
129 self.main_input_name = "input_ids"
130 # trt uses cached and precomputed cross attention vs. framework uses the entire kv cache as output. Need to treat them differently.
131 self.is_trt = is_trt
132
133 def prepare_inputs_for_generation(
134 self,

Callers

nothing calls this directly

Calls 2

deviceMethod · 0.80
__init__Method · 0.45

Tested by

no test coverage detected