(self, config: ChatGLMConfig, device=None, empty_init=True)
| 643 | |
| 644 | class ChatGLMModel(ChatGLMPreTrainedModel): |
| 645 | def __init__(self, config: ChatGLMConfig, device=None, empty_init=True): |
| 646 | super().__init__(config) |
| 647 | if empty_init: |
| 648 | init_method = skip_init |
| 649 | else: |
| 650 | init_method = default_init |
| 651 | init_kwargs = {} |
| 652 | if device is not None: |
| 653 | init_kwargs["device"] = device |
| 654 | self.embedding = init_method(Embedding, config, **init_kwargs) |
| 655 | |
| 656 | # Rotary positional embeddings |
| 657 | self.seq_length = config.seq_length |
| 658 | rotary_dim = ( |
| 659 | config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels |
| 660 | ) |
| 661 | |
| 662 | self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, rope_ratio=config.rope_ratio, original_impl=config.original_rope, |
| 663 | device=device, dtype=config.torch_dtype) |
| 664 | self.encoder = init_method(GLMTransformer, config, **init_kwargs) |
| 665 | self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False, |
| 666 | dtype=config.torch_dtype, **init_kwargs) |
| 667 | |
| 668 | def get_input_embeddings(self): |
| 669 | return self.embedding.word_embeddings |
nothing calls this directly
no test coverage detected