ybelkada commited on
Commit
26072b4
1 Parent(s): 83c91c4

Update modeling_chatglm.py

Browse files
Files changed (1) hide show
  1. modeling_chatglm.py +90 -29
modeling_chatglm.py CHANGED
@@ -157,7 +157,7 @@ class RotaryEmbedding(nn.Module):
157
  )
158
 
159
 
160
- @torch.jit.script
161
  def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
162
  # x: [sq, b, np, hn]
163
  sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
@@ -223,8 +223,7 @@ class CoreAttention(torch.nn.Module):
223
  if pytorch_major_version >= 2:
224
  query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
225
  if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
226
- context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
227
- is_causal=True)
228
  else:
229
  if attention_mask is not None:
230
  attention_mask = ~attention_mask
@@ -237,7 +236,7 @@ class CoreAttention(torch.nn.Module):
237
  # Raw attention scores
238
 
239
  # [b, np, sq, sk]
240
- output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
241
 
242
  # [sq, b, np, hn] -> [sq, b * np, hn]
243
  query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
@@ -312,7 +311,6 @@ class CoreAttention(torch.nn.Module):
312
 
313
  class SelfAttention(torch.nn.Module):
314
  """Parallel self-attention layer abstract class.
315
-
316
  Self-attention layer takes input with size [s, b, h]
317
  and returns output of the same size.
318
  """
@@ -448,7 +446,6 @@ class SelfAttention(torch.nn.Module):
448
 
449
  return output, kv_cache
450
 
451
-
452
  def _config_to_kwargs(args):
453
  common_kwargs = {
454
  "dtype": args.torch_dtype,
@@ -504,7 +501,6 @@ class MLP(torch.nn.Module):
504
 
505
  class GLMBlock(torch.nn.Module):
506
  """A single transformer layer.
507
-
508
  Transformer layer takes input with size [s, b, h] and returns an
509
  output of the same size.
510
  """
@@ -597,7 +593,7 @@ class GLMTransformer(torch.nn.Module):
597
  if self.post_layer_norm:
598
  LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
599
  # Final layer norm before output.
600
- self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
601
  dtype=config.torch_dtype)
602
 
603
  self.gradient_checkpointing = False
@@ -653,7 +649,7 @@ class GLMTransformer(torch.nn.Module):
653
 
654
  # Final layer norm.
655
  if self.post_layer_norm:
656
- hidden_states = self.final_layernorm(hidden_states)
657
 
658
  return hidden_states, presents, all_hidden_states, all_self_attentions
659
 
@@ -740,7 +736,14 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
740
  init_kwargs = {}
741
  if device is not None:
742
  init_kwargs["device"] = device
743
- self.embedding = init_method(Embedding, config, **init_kwargs)
 
 
 
 
 
 
 
744
  self.num_layers = config.num_layers
745
  self.multi_query_group_num = config.multi_query_group_num
746
  self.kv_channels = config.kv_channels
@@ -753,9 +756,21 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
753
 
754
  self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device,
755
  dtype=config.torch_dtype)
756
- self.encoder = init_method(GLMTransformer, config, **init_kwargs)
757
- self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
758
- dtype=config.torch_dtype, **init_kwargs)
 
 
 
 
 
 
 
 
 
 
 
 
759
  self.pre_seq_len = config.pre_seq_len
760
  self.prefix_projection = config.prefix_projection
761
  if self.pre_seq_len is not None:
@@ -765,6 +780,8 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
765
  self.prefix_encoder = PrefixEncoder(config)
766
  self.dropout = torch.nn.Dropout(0.1)
767
 
 
 
768
  def get_input_embeddings(self):
769
  return self.embedding.word_embeddings
770
 
@@ -804,7 +821,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
804
  batch_size, seq_length = input_ids.shape
805
 
806
  if inputs_embeds is None:
807
- inputs_embeds = self.embedding(input_ids)
808
 
809
  if self.pre_seq_len is not None:
810
  if past_key_values is None:
@@ -827,10 +844,54 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
827
  rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
828
 
829
  # Run encoder.
830
- hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
831
- inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
832
- kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
833
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
834
 
835
  if not return_dict:
836
  return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
@@ -844,7 +905,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
844
 
845
  def quantize(self, weight_bit_width: int):
846
  from .quantization import quantize
847
- quantize(self.encoder, weight_bit_width)
848
  return self
849
 
850
 
@@ -853,7 +914,8 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
853
  super().__init__(config)
854
 
855
  self.max_sequence_length = config.max_length
856
- self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
 
857
  self.config = config
858
  self.quantized = False
859
 
@@ -934,7 +996,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
934
  use_cache = use_cache if use_cache is not None else self.config.use_cache
935
  return_dict = return_dict if return_dict is not None else self.config.use_return_dict
936
 
937
- transformer_outputs = self.transformer(
938
  input_ids=input_ids,
939
  position_ids=position_ids,
940
  attention_mask=attention_mask,
@@ -948,8 +1010,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
948
  hidden_states = transformer_outputs[0]
949
  if return_last_logit:
950
  hidden_states = hidden_states[-1:]
951
- lm_logits = self.transformer.output_layer(hidden_states)
952
- lm_logits = lm_logits.transpose(0, 1).contiguous()
953
 
954
  loss = None
955
  if labels is not None:
@@ -1062,8 +1123,8 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1062
  inputs = inputs.to(self.device)
1063
  if past_key_values is not None:
1064
  past_length = past_key_values[0][0].shape[0]
1065
- if self.transformer.pre_seq_len is not None:
1066
- past_length -= self.transformer.pre_seq_len
1067
  inputs.position_ids += past_length
1068
  attention_mask = inputs.attention_mask
1069
  attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
@@ -1205,7 +1266,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1205
 
1206
  self.config.quantization_bit = bits
1207
 
1208
- self.transformer.encoder = quantize(self.transformer.encoder, bits, empty_init=empty_init, device=device,
1209
  **kwargs)
1210
  return self
1211
 
@@ -1215,7 +1276,7 @@ class ChatGLMForSequenceClassification(ChatGLMPreTrainedModel):
1215
  super().__init__(config)
1216
 
1217
  self.num_labels = config.num_labels
1218
- self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
1219
 
1220
  self.classifier_head = nn.Linear(config.hidden_size, config.num_labels, bias=True, dtype=torch.half)
1221
  if config.classifier_dropout is not None:
@@ -1242,7 +1303,7 @@ class ChatGLMForSequenceClassification(ChatGLMPreTrainedModel):
1242
  ) -> Union[Tuple[torch.Tensor, ...], SequenceClassifierOutputWithPast]:
1243
  return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1244
 
1245
- transformer_outputs = self.transformer(
1246
  input_ids=input_ids,
1247
  position_ids=position_ids,
1248
  attention_mask=attention_mask,
@@ -1293,4 +1354,4 @@ class ChatGLMForSequenceClassification(ChatGLMPreTrainedModel):
1293
  past_key_values=transformer_outputs.past_key_values,
1294
  hidden_states=transformer_outputs.hidden_states,
1295
  attentions=transformer_outputs.attentions,
1296
- )
 
157
  )
158
 
159
 
160
+ # @torch.jit.script
161
  def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
162
  # x: [sq, b, np, hn]
163
  sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
 
223
  if pytorch_major_version >= 2:
224
  query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
225
  if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
226
+ context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,is_causal=True)
 
227
  else:
228
  if attention_mask is not None:
229
  attention_mask = ~attention_mask
 
236
  # Raw attention scores
237
 
238
  # [b, np, sq, sk]
239
+ output_size = (query_layer.size(0), query_layer.size(2), query_layer.size(1), key_layer.size(0))
240
 
241
  # [sq, b, np, hn] -> [sq, b * np, hn]
242
  query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
 
311
 
312
  class SelfAttention(torch.nn.Module):
313
  """Parallel self-attention layer abstract class.
 
314
  Self-attention layer takes input with size [s, b, h]
315
  and returns output of the same size.
316
  """
 
446
 
447
  return output, kv_cache
448
 
 
449
  def _config_to_kwargs(args):
450
  common_kwargs = {
451
  "dtype": args.torch_dtype,
 
501
 
502
  class GLMBlock(torch.nn.Module):
503
  """A single transformer layer.
 
504
  Transformer layer takes input with size [s, b, h] and returns an
505
  output of the same size.
506
  """
 
593
  if self.post_layer_norm:
594
  LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
595
  # Final layer norm before output.
596
+ self.norm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
597
  dtype=config.torch_dtype)
598
 
599
  self.gradient_checkpointing = False
 
649
 
650
  # Final layer norm.
651
  if self.post_layer_norm:
652
+ hidden_states = self.norm(hidden_states)
653
 
654
  return hidden_states, presents, all_hidden_states, all_self_attentions
655
 
 
736
  init_kwargs = {}
737
  if device is not None:
738
  init_kwargs["device"] = device
739
+
740
+ self.embed_tokens = nn.Embedding(
741
+ config.padded_vocab_size,
742
+ config.hidden_size,
743
+ dtype=config.torch_dtype,
744
+ device=device
745
+ )
746
+
747
  self.num_layers = config.num_layers
748
  self.multi_query_group_num = config.multi_query_group_num
749
  self.kv_channels = config.kv_channels
 
756
 
757
  self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device,
758
  dtype=config.torch_dtype)
759
+
760
+ # Transformer layers.
761
+ def build_layer(layer_number):
762
+ return GLMBlock(config, layer_number, device=device)
763
+
764
+ self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])
765
+ self.num_layers = config.num_layers
766
+ self.post_layer_norm = config.post_layer_norm
767
+
768
+ if self.post_layer_norm:
769
+ LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
770
+ # Final layer norm before output.
771
+ self.norm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
772
+ dtype=config.torch_dtype)
773
+
774
  self.pre_seq_len = config.pre_seq_len
775
  self.prefix_projection = config.prefix_projection
776
  if self.pre_seq_len is not None:
 
780
  self.prefix_encoder = PrefixEncoder(config)
781
  self.dropout = torch.nn.Dropout(0.1)
782
 
783
+ self.gradient_checkpointing = False
784
+
785
  def get_input_embeddings(self):
786
  return self.embedding.word_embeddings
787
 
 
821
  batch_size, seq_length = input_ids.shape
822
 
823
  if inputs_embeds is None:
824
+ inputs_embeds = self.embed_tokens(input_ids)
825
 
826
  if self.pre_seq_len is not None:
827
  if past_key_values is None:
 
844
  rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
845
 
846
  # Run encoder.
847
+ if not past_key_values:
848
+ past_key_values = [None for _ in range(self.num_layers)]
849
+ presents = () if use_cache else None
850
+ if self.gradient_checkpointing and self.training:
851
+ if use_cache:
852
+ logger.warning_once(
853
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
854
+ )
855
+ use_cache = False
856
+
857
+ all_self_attentions = None
858
+ all_hidden_states = () if output_hidden_states else None
859
+
860
+ hidden_states = inputs_embeds
861
+ # To comply with former chat-glm format that expects (seqlen, bs, hd)
862
+ hidden_states = hidden_states.permute(1, 0, 2)
863
+
864
+ for index, layer in enumerate(self.layers):
865
+ if output_hidden_states:
866
+ all_hidden_states = all_hidden_states + (hidden_states,)
867
+
868
+ if self.gradient_checkpointing and self.training:
869
+ layer_ret = torch.utils.checkpoint.checkpoint(
870
+ layer,
871
+ hidden_states,
872
+ full_attention_mask,
873
+ rotary_pos_emb,
874
+ past_key_values[index],
875
+ use_cache
876
+ )
877
+ else:
878
+ layer_ret = layer(
879
+ hidden_states,
880
+ full_attention_mask,
881
+ rotary_pos_emb,
882
+ kv_cache=past_key_values[index],
883
+ use_cache=use_cache
884
+ )
885
+ hidden_states, kv_cache = layer_ret
886
+ if use_cache:
887
+ presents = presents + (kv_cache,)
888
+
889
+ if output_hidden_states:
890
+ all_hidden_states = all_hidden_states + (hidden_states,)
891
+
892
+ # Final layer norm.
893
+ if self.post_layer_norm:
894
+ hidden_states = self.norm(hidden_states)
895
 
896
  if not return_dict:
897
  return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
 
905
 
906
  def quantize(self, weight_bit_width: int):
907
  from .quantization import quantize
908
+ quantize(self, weight_bit_width)
909
  return self
910
 
911
 
 
914
  super().__init__(config)
915
 
916
  self.max_sequence_length = config.max_length
917
+ self.model = ChatGLMModel(config, empty_init=empty_init, device=device)
918
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
919
  self.config = config
920
  self.quantized = False
921
 
 
996
  use_cache = use_cache if use_cache is not None else self.config.use_cache
997
  return_dict = return_dict if return_dict is not None else self.config.use_return_dict
998
 
999
+ transformer_outputs = self.model(
1000
  input_ids=input_ids,
1001
  position_ids=position_ids,
1002
  attention_mask=attention_mask,
 
1010
  hidden_states = transformer_outputs[0]
1011
  if return_last_logit:
1012
  hidden_states = hidden_states[-1:]
1013
+ lm_logits = self.lm_head(hidden_states)
 
1014
 
1015
  loss = None
1016
  if labels is not None:
 
1123
  inputs = inputs.to(self.device)
1124
  if past_key_values is not None:
1125
  past_length = past_key_values[0][0].shape[0]
1126
+ if self.model.pre_seq_len is not None:
1127
+ past_length -= self.model.pre_seq_len
1128
  inputs.position_ids += past_length
1129
  attention_mask = inputs.attention_mask
1130
  attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
 
1266
 
1267
  self.config.quantization_bit = bits
1268
 
1269
+ self.model = quantize(self.model, bits, empty_init=empty_init, device=device,
1270
  **kwargs)
1271
  return self
1272
 
 
1276
  super().__init__(config)
1277
 
1278
  self.num_labels = config.num_labels
1279
+ self.model = ChatGLMModel(config, empty_init=empty_init, device=device)
1280
 
1281
  self.classifier_head = nn.Linear(config.hidden_size, config.num_labels, bias=True, dtype=torch.half)
1282
  if config.classifier_dropout is not None:
 
1303
  ) -> Union[Tuple[torch.Tensor, ...], SequenceClassifierOutputWithPast]:
1304
  return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1305
 
1306
+ transformer_outputs = self.model(
1307
  input_ids=input_ids,
1308
  position_ids=position_ids,
1309
  attention_mask=attention_mask,
 
1354
  past_key_values=transformer_outputs.past_key_values,
1355
  hidden_states=transformer_outputs.hidden_states,
1356
  attentions=transformer_outputs.attentions,
1357
+ )