Upload folder using huggingface_hub
Browse files
modeling_internvl_chat.py
CHANGED
@@ -33,6 +33,7 @@ def version_cmp(v1, v2, op='eq'):
|
|
33 |
class InternVLChatModel(PreTrainedModel):
|
34 |
config_class = InternVLChatConfig
|
35 |
main_input_name = 'pixel_values'
|
|
|
36 |
_supports_flash_attn_2 = True
|
37 |
_no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer']
|
38 |
|
@@ -97,7 +98,7 @@ class InternVLChatModel(PreTrainedModel):
|
|
97 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
98 |
|
99 |
image_flags = image_flags.squeeze(-1)
|
100 |
-
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
101 |
|
102 |
vit_embeds = self.extract_feature(pixel_values)
|
103 |
vit_embeds = vit_embeds[image_flags == 1]
|
@@ -230,8 +231,8 @@ class InternVLChatModel(PreTrainedModel):
|
|
230 |
|
231 |
tokenizer.padding_side = 'left'
|
232 |
model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
|
233 |
-
input_ids = model_inputs['input_ids'].
|
234 |
-
attention_mask = model_inputs['attention_mask'].
|
235 |
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
236 |
generation_config['eos_token_id'] = eos_token_id
|
237 |
generation_output = self.generate(
|
@@ -279,8 +280,8 @@ class InternVLChatModel(PreTrainedModel):
|
|
279 |
query = query.replace('<image>', image_tokens, 1)
|
280 |
|
281 |
model_inputs = tokenizer(query, return_tensors='pt')
|
282 |
-
input_ids = model_inputs['input_ids'].
|
283 |
-
attention_mask = model_inputs['attention_mask'].
|
284 |
generation_config['eos_token_id'] = eos_token_id
|
285 |
generation_output = self.generate(
|
286 |
pixel_values=pixel_values,
|
|
|
33 |
class InternVLChatModel(PreTrainedModel):
|
34 |
config_class = InternVLChatConfig
|
35 |
main_input_name = 'pixel_values'
|
36 |
+
base_model_prefix = 'language_model'
|
37 |
_supports_flash_attn_2 = True
|
38 |
_no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer']
|
39 |
|
|
|
98 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
99 |
|
100 |
image_flags = image_flags.squeeze(-1)
|
101 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
|
102 |
|
103 |
vit_embeds = self.extract_feature(pixel_values)
|
104 |
vit_embeds = vit_embeds[image_flags == 1]
|
|
|
231 |
|
232 |
tokenizer.padding_side = 'left'
|
233 |
model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
|
234 |
+
input_ids = model_inputs['input_ids'].to(self.device)
|
235 |
+
attention_mask = model_inputs['attention_mask'].to(self.device)
|
236 |
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
237 |
generation_config['eos_token_id'] = eos_token_id
|
238 |
generation_output = self.generate(
|
|
|
280 |
query = query.replace('<image>', image_tokens, 1)
|
281 |
|
282 |
model_inputs = tokenizer(query, return_tensors='pt')
|
283 |
+
input_ids = model_inputs['input_ids'].to(self.device)
|
284 |
+
attention_mask = model_inputs['attention_mask'].to(self.device)
|
285 |
generation_config['eos_token_id'] = eos_token_id
|
286 |
generation_output = self.generate(
|
287 |
pixel_values=pixel_values,
|