MCPcopy Create free account
hub / github.com/THUDM/LongWriter / LlamaPreTrainedModel

Class LlamaPreTrainedModel

train/patch/modeling_llama.py:813–834  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

811 LLAMA_START_DOCSTRING,
812)
813class LlamaPreTrainedModel(PreTrainedModel):
814 config_class = LlamaConfig
815 base_model_prefix = "model"
816 supports_gradient_checkpointing = True
817 _no_split_modules = ["LlamaDecoderLayer"]
818 _skip_keys_device_placement = ["past_key_values"]
819 _supports_flash_attn_2 = True
820 _supports_sdpa = True
821 _supports_cache_class = True
822 _supports_quantized_cache = True
823 _supports_static_cache = True
824
825 def _init_weights(self, module):
826 std = self.config.initializer_range
827 if isinstance(module, nn.Linear):
828 module.weight.data.normal_(mean=0.0, std=std)
829 if module.bias is not None:
830 module.bias.data.zero_()
831 elif isinstance(module, nn.Embedding):
832 module.weight.data.normal_(mean=0.0, std=std)
833 if module.padding_idx is not None:
834 module.weight.data[module.padding_idx].zero_()
835
836
837LLAMA_INPUTS_DOCSTRING = r"""

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected