ganlinyang commited on
Commit
fef85fe
·
verified ·
1 Parent(s): a065255

Upload folder using huggingface_hub

Browse files
1B/20250802/added_tokens.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "</box>": 151673,
3
+ "</img>": 151666,
4
+ "</quad>": 151669,
5
+ "</ref>": 151671,
6
+ "</tool_call>": 151658,
7
+ "<IMG_CONTEXT>": 151667,
8
+ "<box>": 151672,
9
+ "<img>": 151665,
10
+ "<quad>": 151668,
11
+ "<ref>": 151670,
12
+ "<tool_call>": 151657,
13
+ "<|box_end|>": 151649,
14
+ "<|box_start|>": 151648,
15
+ "<|endoftext|>": 151643,
16
+ "<|file_sep|>": 151664,
17
+ "<|fim_middle|>": 151660,
18
+ "<|fim_pad|>": 151662,
19
+ "<|fim_prefix|>": 151659,
20
+ "<|fim_suffix|>": 151661,
21
+ "<|im_end|>": 151645,
22
+ "<|im_start|>": 151644,
23
+ "<|image_pad|>": 151655,
24
+ "<|object_ref_end|>": 151647,
25
+ "<|object_ref_start|>": 151646,
26
+ "<|quad_end|>": 151651,
27
+ "<|quad_start|>": 151650,
28
+ "<|repo_name|>": 151663,
29
+ "<|video_pad|>": 151656,
30
+ "<|vision_end|>": 151653,
31
+ "<|vision_pad|>": 151654,
32
+ "<|vision_start|>": 151652
33
+ }
1B/20250802/all_results.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 1.0,
3
+ "train_loss": 0.8273731040954589,
4
+ "train_runtime": 16224.5195,
5
+ "train_samples": -1,
6
+ "train_samples_per_second": 15.779,
7
+ "train_steps_per_second": 0.123
8
+ }
1B/20250802/config.json ADDED
@@ -0,0 +1,225 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_commit_hash": null,
3
+ "_name_or_path": "/mnt/petrelfs/share_data/wangweiyun/share_internvl/InternVL3-1B",
4
+ "architectures": [
5
+ "InternVLChatModel"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
9
+ "AutoModel": "modeling_internvl_chat.InternVLChatModel",
10
+ "AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel"
11
+ },
12
+ "downsample_ratio": 0.5,
13
+ "dynamic_image_size": true,
14
+ "force_image_size": 448,
15
+ "hidden_size": 896,
16
+ "image_fold": null,
17
+ "llm_config": {
18
+ "_attn_implementation_autoset": true,
19
+ "_name_or_path": "./pretrained/Qwen2.5-32B-Instruct",
20
+ "add_cross_attention": false,
21
+ "architectures": [
22
+ "Qwen2ForCausalLM"
23
+ ],
24
+ "attention_dropout": 0.0,
25
+ "attn_implementation": "flash_attention_2",
26
+ "bad_words_ids": null,
27
+ "begin_suppress_tokens": null,
28
+ "bos_token_id": 151643,
29
+ "chunk_size_feed_forward": 0,
30
+ "cross_attention_hidden_size": null,
31
+ "decoder_start_token_id": null,
32
+ "diversity_penalty": 0.0,
33
+ "do_sample": false,
34
+ "early_stopping": false,
35
+ "encoder_no_repeat_ngram_size": 0,
36
+ "eos_token_id": 151643,
37
+ "exponential_decay_length_penalty": null,
38
+ "finetuning_task": null,
39
+ "forced_bos_token_id": null,
40
+ "forced_eos_token_id": null,
41
+ "hidden_act": "silu",
42
+ "hidden_size": 896,
43
+ "id2label": {
44
+ "0": "LABEL_0",
45
+ "1": "LABEL_1"
46
+ },
47
+ "initializer_range": 0.02,
48
+ "intermediate_size": 4864,
49
+ "is_decoder": false,
50
+ "is_encoder_decoder": false,
51
+ "label2id": {
52
+ "LABEL_0": 0,
53
+ "LABEL_1": 1
54
+ },
55
+ "length_penalty": 1.0,
56
+ "max_length": 20,
57
+ "max_position_embeddings": 32768,
58
+ "max_window_layers": 70,
59
+ "min_length": 0,
60
+ "model_type": "qwen2",
61
+ "moe_config": null,
62
+ "no_repeat_ngram_size": 0,
63
+ "num_attention_heads": 14,
64
+ "num_beam_groups": 1,
65
+ "num_beams": 1,
66
+ "num_hidden_layers": 24,
67
+ "num_key_value_heads": 2,
68
+ "num_return_sequences": 1,
69
+ "output_attentions": false,
70
+ "output_hidden_states": false,
71
+ "output_scores": false,
72
+ "pad_token_id": null,
73
+ "prefix": null,
74
+ "problem_type": null,
75
+ "pruned_heads": {},
76
+ "remove_invalid_values": false,
77
+ "repetition_penalty": 1.0,
78
+ "return_dict": true,
79
+ "return_dict_in_generate": false,
80
+ "rms_norm_eps": 1e-06,
81
+ "rope_scaling": {
82
+ "factor": 2.0,
83
+ "rope_type": "dynamic",
84
+ "type": "dynamic"
85
+ },
86
+ "rope_theta": 1000000.0,
87
+ "sep_token_id": null,
88
+ "sliding_window": null,
89
+ "suppress_tokens": null,
90
+ "task_specific_params": null,
91
+ "temperature": 1.0,
92
+ "tf_legacy_loss": false,
93
+ "tie_encoder_decoder": false,
94
+ "tie_word_embeddings": false,
95
+ "tokenizer_class": null,
96
+ "top_k": 50,
97
+ "top_p": 1.0,
98
+ "torch_dtype": "bfloat16",
99
+ "torchscript": false,
100
+ "transformers_version": "4.37.2",
101
+ "typical_p": 1.0,
102
+ "use_bfloat16": true,
103
+ "use_cache": false,
104
+ "use_sliding_window": false,
105
+ "vocab_size": 151674
106
+ },
107
+ "max_dynamic_patch": 12,
108
+ "min_dynamic_patch": 1,
109
+ "model_type": "internvl_chat",
110
+ "pad2square": false,
111
+ "ps_version": "v2",
112
+ "select_layer": -1,
113
+ "system_message": null,
114
+ "template": "internvl2_5",
115
+ "tie_word_embeddings": false,
116
+ "torch_dtype": "bfloat16",
117
+ "transformers_version": null,
118
+ "use_backbone_lora": 0,
119
+ "use_llm_lora": 0,
120
+ "use_thumbnail": true,
121
+ "vision_config": {
122
+ "_attn_implementation_autoset": true,
123
+ "_name_or_path": "OpenGVLab/InternViT-6B-448px-V1-5",
124
+ "add_cross_attention": false,
125
+ "architectures": [
126
+ "InternVisionModel"
127
+ ],
128
+ "attention_dropout": 0.0,
129
+ "auto_map": {
130
+ "AutoConfig": "configuration_intern_vit.InternVisionConfig",
131
+ "AutoModel": "modeling_intern_vit.InternVisionModel"
132
+ },
133
+ "bad_words_ids": null,
134
+ "begin_suppress_tokens": null,
135
+ "bos_token_id": null,
136
+ "capacity_factor": 1.2,
137
+ "chunk_size_feed_forward": 0,
138
+ "cross_attention_hidden_size": null,
139
+ "decoder_start_token_id": null,
140
+ "diversity_penalty": 0.0,
141
+ "do_sample": false,
142
+ "drop_path_rate": 0.0,
143
+ "dropout": 0.0,
144
+ "early_stopping": false,
145
+ "encoder_no_repeat_ngram_size": 0,
146
+ "eos_token_id": null,
147
+ "eval_capacity_factor": 1.4,
148
+ "exponential_decay_length_penalty": null,
149
+ "finetuning_task": null,
150
+ "forced_bos_token_id": null,
151
+ "forced_eos_token_id": null,
152
+ "hidden_act": "gelu",
153
+ "hidden_size": 1024,
154
+ "id2label": {
155
+ "0": "LABEL_0",
156
+ "1": "LABEL_1"
157
+ },
158
+ "image_size": 448,
159
+ "initializer_factor": 0.1,
160
+ "initializer_range": 1e-10,
161
+ "intermediate_size": 4096,
162
+ "is_decoder": false,
163
+ "is_encoder_decoder": false,
164
+ "label2id": {
165
+ "LABEL_0": 0,
166
+ "LABEL_1": 1
167
+ },
168
+ "laux_allreduce": "all_nodes",
169
+ "layer_norm_eps": 1e-06,
170
+ "length_penalty": 1.0,
171
+ "max_length": 20,
172
+ "min_length": 0,
173
+ "model_type": "intern_vit_6b",
174
+ "moe_coeff_ratio": 0.5,
175
+ "moe_intermediate_size": 768,
176
+ "moe_output_scale": 4.0,
177
+ "no_repeat_ngram_size": 0,
178
+ "noisy_gate_policy": "RSample_before",
179
+ "norm_type": "layer_norm",
180
+ "num_attention_heads": 16,
181
+ "num_beam_groups": 1,
182
+ "num_beams": 1,
183
+ "num_channels": 3,
184
+ "num_experts": 8,
185
+ "num_hidden_layers": 24,
186
+ "num_return_sequences": 1,
187
+ "num_routed_experts": 4,
188
+ "num_shared_experts": 4,
189
+ "output_attentions": false,
190
+ "output_hidden_states": false,
191
+ "output_scores": false,
192
+ "pad_token_id": null,
193
+ "patch_size": 14,
194
+ "prefix": null,
195
+ "problem_type": null,
196
+ "pruned_heads": {},
197
+ "qk_normalization": false,
198
+ "qkv_bias": true,
199
+ "remove_invalid_values": false,
200
+ "repetition_penalty": 1.0,
201
+ "return_dict": true,
202
+ "return_dict_in_generate": false,
203
+ "sep_token_id": null,
204
+ "shared_expert_intermediate_size": 3072,
205
+ "suppress_tokens": null,
206
+ "task_specific_params": null,
207
+ "temperature": 1.0,
208
+ "tf_legacy_loss": false,
209
+ "tie_encoder_decoder": false,
210
+ "tie_word_embeddings": true,
211
+ "tokenizer_class": null,
212
+ "top_k": 50,
213
+ "top_p": 1.0,
214
+ "torch_dtype": "bfloat16",
215
+ "torchscript": false,
216
+ "transformers_version": "4.37.2",
217
+ "typical_p": 1.0,
218
+ "use_bfloat16": true,
219
+ "use_flash_attn": true,
220
+ "use_moe": false,
221
+ "use_residual": true,
222
+ "use_rts": false,
223
+ "use_weighted_residual": false
224
+ }
225
+ }
1B/20250802/configuration_intern_vit.py ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import os
8
+ from typing import Union
9
+
10
+ from transformers.configuration_utils import PretrainedConfig
11
+ from transformers.utils import logging
12
+
13
+ logger = logging.get_logger(__name__)
14
+
15
+
16
+ class InternVisionConfig(PretrainedConfig):
17
+ r"""
18
+ This is the configuration class to store the configuration of a [`InternVisionModel`]. It is used to
19
+ instantiate a vision encoder according to the specified arguments, defining the model architecture.
20
+
21
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
22
+ documentation from [`PretrainedConfig`] for more information.
23
+
24
+ Args:
25
+ num_channels (`int`, *optional*, defaults to 3):
26
+ Number of color channels in the input images (e.g., 3 for RGB).
27
+ patch_size (`int`, *optional*, defaults to 14):
28
+ The size (resolution) of each patch.
29
+ image_size (`int`, *optional*, defaults to 224):
30
+ The size (resolution) of each image.
31
+ qkv_bias (`bool`, *optional*, defaults to `False`):
32
+ Whether to add a bias to the queries and values in the self-attention layers.
33
+ hidden_size (`int`, *optional*, defaults to 3200):
34
+ Dimensionality of the encoder layers and the pooler layer.
35
+ num_attention_heads (`int`, *optional*, defaults to 25):
36
+ Number of attention heads for each attention layer in the Transformer encoder.
37
+ intermediate_size (`int`, *optional*, defaults to 12800):
38
+ Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
39
+ qk_normalization (`bool`, *optional*, defaults to `True`):
40
+ Whether to normalize the queries and keys in the self-attention layers.
41
+ num_hidden_layers (`int`, *optional*, defaults to 48):
42
+ Number of hidden layers in the Transformer encoder.
43
+ use_flash_attn (`bool`, *optional*, defaults to `True`):
44
+ Whether to use flash attention mechanism.
45
+ hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
46
+ The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
47
+ `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported.
48
+ layer_norm_eps (`float`, *optional*, defaults to 1e-6):
49
+ The epsilon used by the layer normalization layers.
50
+ dropout (`float`, *optional*, defaults to 0.0):
51
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
52
+ drop_path_rate (`float`, *optional*, defaults to 0.0):
53
+ Dropout rate for stochastic depth.
54
+ attention_dropout (`float`, *optional*, defaults to 0.0):
55
+ The dropout ratio for the attention probabilities.
56
+ initializer_range (`float`, *optional*, defaults to 0.02):
57
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
58
+ initializer_factor (`float`, *optional*, defaults to 0.1):
59
+ A factor for layer scale.
60
+ """
61
+
62
+ model_type = 'intern_vit_6b'
63
+
64
+ def __init__(
65
+ self,
66
+ num_channels=3,
67
+ patch_size=14,
68
+ image_size=224,
69
+ qkv_bias=False,
70
+ hidden_size=3200,
71
+ num_attention_heads=25,
72
+ intermediate_size=12800,
73
+ qk_normalization=True,
74
+ num_hidden_layers=48,
75
+ use_flash_attn=True,
76
+ hidden_act='gelu',
77
+ norm_type='rms_norm',
78
+ layer_norm_eps=1e-6,
79
+ dropout=0.0,
80
+ drop_path_rate=0.0,
81
+ attention_dropout=0.0,
82
+ initializer_range=0.02,
83
+ initializer_factor=0.1,
84
+ **kwargs,
85
+ ):
86
+ super().__init__(**kwargs)
87
+
88
+ self.hidden_size = hidden_size
89
+ self.intermediate_size = intermediate_size
90
+ self.dropout = dropout
91
+ self.drop_path_rate = drop_path_rate
92
+ self.num_hidden_layers = num_hidden_layers
93
+ self.num_attention_heads = num_attention_heads
94
+ self.num_channels = num_channels
95
+ self.patch_size = patch_size
96
+ self.image_size = image_size
97
+ self.initializer_range = initializer_range
98
+ self.initializer_factor = initializer_factor
99
+ self.attention_dropout = attention_dropout
100
+ self.layer_norm_eps = layer_norm_eps
101
+ self.hidden_act = hidden_act
102
+ self.norm_type = norm_type
103
+ self.qkv_bias = qkv_bias
104
+ self.qk_normalization = qk_normalization
105
+ self.use_flash_attn = use_flash_attn
106
+
107
+ @classmethod
108
+ def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> 'PretrainedConfig':
109
+ config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
110
+
111
+ if 'vision_config' in config_dict:
112
+ config_dict = config_dict['vision_config']
113
+
114
+ if 'model_type' in config_dict and hasattr(cls, 'model_type') and config_dict['model_type'] != cls.model_type:
115
+ logger.warning(
116
+ f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
117
+ f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.'
118
+ )
119
+
120
+ return cls.from_dict(config_dict, **kwargs)
1B/20250802/configuration_internvl_chat.py ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import copy
8
+
9
+ from transformers import AutoConfig, LlamaConfig, Qwen2Config
10
+ from transformers.configuration_utils import PretrainedConfig
11
+ from transformers.utils import logging
12
+
13
+ from .configuration_intern_vit import InternVisionConfig
14
+
15
+ logger = logging.get_logger(__name__)
16
+
17
+
18
+ class InternVLChatConfig(PretrainedConfig):
19
+ model_type = 'internvl_chat'
20
+ is_composition = True
21
+
22
+ def __init__(
23
+ self,
24
+ vision_config=None,
25
+ llm_config=None,
26
+ use_backbone_lora=0,
27
+ use_llm_lora=0,
28
+ select_layer=-1,
29
+ force_image_size=None,
30
+ downsample_ratio=0.5,
31
+ template=None,
32
+ dynamic_image_size=False,
33
+ use_thumbnail=False,
34
+ ps_version='v1',
35
+ min_dynamic_patch=1,
36
+ max_dynamic_patch=6,
37
+ **kwargs):
38
+ super().__init__(**kwargs)
39
+
40
+ if vision_config is None:
41
+ vision_config = {'architectures': ['InternVisionModel']}
42
+ logger.info('vision_config is None. Initializing the InternVisionConfig with default values.')
43
+
44
+ if llm_config is None:
45
+ llm_config = {'architectures': ['Qwen2ForCausalLM']}
46
+ logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).')
47
+
48
+ self.vision_config = InternVisionConfig(**vision_config)
49
+ if llm_config.get('architectures')[0] == 'LlamaForCausalLM':
50
+ self.llm_config = LlamaConfig(**llm_config)
51
+ elif llm_config.get('architectures')[0] == 'Qwen2ForCausalLM':
52
+ self.llm_config = Qwen2Config(**llm_config)
53
+ else:
54
+ raise ValueError('Unsupported architecture: {}'.format(llm_config.get('architectures')[0]))
55
+ self.use_backbone_lora = use_backbone_lora
56
+ self.use_llm_lora = use_llm_lora
57
+ self.select_layer = select_layer
58
+ self.force_image_size = force_image_size
59
+ self.downsample_ratio = downsample_ratio
60
+ self.template = template
61
+ self.dynamic_image_size = dynamic_image_size
62
+ self.use_thumbnail = use_thumbnail
63
+ self.ps_version = ps_version # pixel shuffle version
64
+ self.min_dynamic_patch = min_dynamic_patch
65
+ self.max_dynamic_patch = max_dynamic_patch
66
+ # By default, we use tie_word_embeddings=False for models of all sizes.
67
+ self.tie_word_embeddings = self.llm_config.tie_word_embeddings
68
+
69
+ logger.info(f'vision_select_layer: {self.select_layer}')
70
+ logger.info(f'ps_version: {self.ps_version}')
71
+ logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')
72
+ logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')
73
+
74
+ def to_dict(self):
75
+ """
76
+ Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
77
+
78
+ Returns:
79
+ `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
80
+ """
81
+ output = copy.deepcopy(self.__dict__)
82
+ output['vision_config'] = self.vision_config.to_dict()
83
+ output['llm_config'] = self.llm_config.to_dict()
84
+ output['model_type'] = self.__class__.model_type
85
+ output['use_backbone_lora'] = self.use_backbone_lora
86
+ output['use_llm_lora'] = self.use_llm_lora
87
+ output['select_layer'] = self.select_layer
88
+ output['force_image_size'] = self.force_image_size
89
+ output['downsample_ratio'] = self.downsample_ratio
90
+ output['template'] = self.template
91
+ output['dynamic_image_size'] = self.dynamic_image_size
92
+ output['use_thumbnail'] = self.use_thumbnail
93
+ output['ps_version'] = self.ps_version
94
+ output['min_dynamic_patch'] = self.min_dynamic_patch
95
+ output['max_dynamic_patch'] = self.max_dynamic_patch
96
+
97
+ return output
1B/20250802/conversation.py ADDED
@@ -0,0 +1,391 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Conversation prompt templates.
3
+
4
+ We kindly request that you import fastchat instead of copying this file if you wish to use it.
5
+ If you have changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
6
+
7
+ Modified from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
8
+ """
9
+
10
+ import dataclasses
11
+ from enum import IntEnum, auto
12
+ from typing import Dict, List, Tuple, Union
13
+
14
+
15
+ class SeparatorStyle(IntEnum):
16
+ """Separator styles."""
17
+
18
+ ADD_COLON_SINGLE = auto()
19
+ ADD_COLON_TWO = auto()
20
+ ADD_COLON_SPACE_SINGLE = auto()
21
+ NO_COLON_SINGLE = auto()
22
+ NO_COLON_TWO = auto()
23
+ ADD_NEW_LINE_SINGLE = auto()
24
+ LLAMA2 = auto()
25
+ CHATGLM = auto()
26
+ CHATML = auto()
27
+ CHATINTERN = auto()
28
+ DOLLY = auto()
29
+ RWKV = auto()
30
+ PHOENIX = auto()
31
+ ROBIN = auto()
32
+ FALCON_CHAT = auto()
33
+ CHATGLM3 = auto()
34
+ INTERNVL_ZH = auto()
35
+ MPT = auto()
36
+
37
+
38
+ @dataclasses.dataclass
39
+ class Conversation:
40
+ """A class that manages prompt templates and keeps all conversation history."""
41
+
42
+ # The name of this template
43
+ name: str
44
+ # The template of the system prompt
45
+ system_template: str = '{system_message}'
46
+ # The system message
47
+ system_message: str = ''
48
+ # The names of two roles
49
+ roles: Tuple[str] = ('USER', 'ASSISTANT')
50
+ # All messages. Each item is (role, message).
51
+ messages: List[List[str]] = ()
52
+ # The number of few shot examples
53
+ offset: int = 0
54
+ # The separator style and configurations
55
+ sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
56
+ sep: str = '\n'
57
+ sep2: str = None
58
+ # Stop criteria (the default one is EOS token)
59
+ stop_str: Union[str, List[str]] = None
60
+ # Stops generation if meeting any token in this list
61
+ stop_token_ids: List[int] = None
62
+
63
+ def get_prompt(self) -> str:
64
+ """Get the prompt for generation."""
65
+ system_prompt = self.system_template.format(system_message=self.system_message)
66
+ if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
67
+ ret = system_prompt + self.sep
68
+ for role, message in self.messages:
69
+ if message:
70
+ ret += role + ': ' + message + self.sep
71
+ else:
72
+ ret += role + ':'
73
+ return ret
74
+ elif self.sep_style == SeparatorStyle.ADD_COLON_TWO:
75
+ seps = [self.sep, self.sep2]
76
+ ret = system_prompt + seps[0]
77
+ for i, (role, message) in enumerate(self.messages):
78
+ if message:
79
+ ret += role + ': ' + message + seps[i % 2]
80
+ else:
81
+ ret += role + ':'
82
+ return ret
83
+ elif self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
84
+ ret = system_prompt + self.sep
85
+ for role, message in self.messages:
86
+ if message:
87
+ ret += role + ': ' + message + self.sep
88
+ else:
89
+ ret += role + ': ' # must be end with a space
90
+ return ret
91
+ elif self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
92
+ ret = '' if system_prompt == '' else system_prompt + self.sep
93
+ for role, message in self.messages:
94
+ if message:
95
+ ret += role + '\n' + message + self.sep
96
+ else:
97
+ ret += role + '\n'
98
+ return ret
99
+ elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
100
+ ret = system_prompt
101
+ for role, message in self.messages:
102
+ if message:
103
+ ret += role + message + self.sep
104
+ else:
105
+ ret += role
106
+ return ret
107
+ elif self.sep_style == SeparatorStyle.NO_COLON_TWO:
108
+ seps = [self.sep, self.sep2]
109
+ ret = system_prompt
110
+ for i, (role, message) in enumerate(self.messages):
111
+ if message:
112
+ ret += role + message + seps[i % 2]
113
+ else:
114
+ ret += role
115
+ return ret
116
+ elif self.sep_style == SeparatorStyle.RWKV:
117
+ ret = system_prompt
118
+ for i, (role, message) in enumerate(self.messages):
119
+ if message:
120
+ ret += (
121
+ role
122
+ + ': '
123
+ + message.replace('\r\n', '\n').replace('\n\n', '\n')
124
+ )
125
+ ret += '\n\n'
126
+ else:
127
+ ret += role + ':'
128
+ return ret
129
+ elif self.sep_style == SeparatorStyle.LLAMA2:
130
+ seps = [self.sep, self.sep2]
131
+ if self.system_message:
132
+ ret = system_prompt
133
+ else:
134
+ ret = '[INST] '
135
+ for i, (role, message) in enumerate(self.messages):
136
+ tag = self.roles[i % 2]
137
+ if message:
138
+ if i == 0:
139
+ ret += message + ' '
140
+ else:
141
+ ret += tag + ' ' + message + seps[i % 2]
142
+ else:
143
+ ret += tag
144
+ return ret
145
+ elif self.sep_style == SeparatorStyle.CHATGLM:
146
+ # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
147
+ # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
148
+ round_add_n = 1 if self.name == 'chatglm2' else 0
149
+ if system_prompt:
150
+ ret = system_prompt + self.sep
151
+ else:
152
+ ret = ''
153
+
154
+ for i, (role, message) in enumerate(self.messages):
155
+ if i % 2 == 0:
156
+ ret += f'[Round {i//2 + round_add_n}]{self.sep}'
157
+
158
+ if message:
159
+ ret += f'{role}:{message}{self.sep}'
160
+ else:
161
+ ret += f'{role}:'
162
+ return ret
163
+ elif self.sep_style == SeparatorStyle.CHATML:
164
+ ret = '' if system_prompt == '' else system_prompt + self.sep + '\n'
165
+ for role, message in self.messages:
166
+ if message:
167
+ ret += role + '\n' + message + self.sep + '\n'
168
+ else:
169
+ ret += role + '\n'
170
+ return ret
171
+ elif self.sep_style == SeparatorStyle.CHATGLM3:
172
+ ret = ''
173
+ if self.system_message:
174
+ ret += system_prompt
175
+ for role, message in self.messages:
176
+ if message:
177
+ ret += role + '\n' + ' ' + message
178
+ else:
179
+ ret += role
180
+ return ret
181
+ elif self.sep_style == SeparatorStyle.CHATINTERN:
182
+ # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
183
+ seps = [self.sep, self.sep2]
184
+ ret = system_prompt
185
+ for i, (role, message) in enumerate(self.messages):
186
+ # if i % 2 == 0:
187
+ # ret += "<s>"
188
+ if message:
189
+ ret += role + ':' + message + seps[i % 2] + '\n'
190
+ else:
191
+ ret += role + ':'
192
+ return ret
193
+ elif self.sep_style == SeparatorStyle.DOLLY:
194
+ seps = [self.sep, self.sep2]
195
+ ret = system_prompt
196
+ for i, (role, message) in enumerate(self.messages):
197
+ if message:
198
+ ret += role + ':\n' + message + seps[i % 2]
199
+ if i % 2 == 1:
200
+ ret += '\n\n'
201
+ else:
202
+ ret += role + ':\n'
203
+ return ret
204
+ elif self.sep_style == SeparatorStyle.PHOENIX:
205
+ ret = system_prompt
206
+ for role, message in self.messages:
207
+ if message:
208
+ ret += role + ': ' + '<s>' + message + '</s>'
209
+ else:
210
+ ret += role + ': ' + '<s>'
211
+ return ret
212
+ elif self.sep_style == SeparatorStyle.ROBIN:
213
+ ret = system_prompt + self.sep
214
+ for role, message in self.messages:
215
+ if message:
216
+ ret += role + ':\n' + message + self.sep
217
+ else:
218
+ ret += role + ':\n'
219
+ return ret
220
+ elif self.sep_style == SeparatorStyle.FALCON_CHAT:
221
+ ret = ''
222
+ if self.system_message:
223
+ ret += system_prompt + self.sep
224
+ for role, message in self.messages:
225
+ if message:
226
+ ret += role + ': ' + message + self.sep
227
+ else:
228
+ ret += role + ':'
229
+
230
+ return ret
231
+ elif self.sep_style == SeparatorStyle.INTERNVL_ZH:
232
+ seps = [self.sep, self.sep2]
233
+ ret = self.system_message + seps[0]
234
+ for i, (role, message) in enumerate(self.messages):
235
+ if message:
236
+ ret += role + ': ' + message + seps[i % 2]
237
+ else:
238
+ ret += role + ':'
239
+ return ret
240
+ elif self.sep_style == SeparatorStyle.MPT:
241
+ ret = system_prompt + self.sep
242
+ for role, message in self.messages:
243
+ if message:
244
+ if type(message) is tuple:
245
+ message, _, _ = message
246
+ ret += role + message + self.sep
247
+ else:
248
+ ret += role
249
+ return ret
250
+ else:
251
+ raise ValueError(f'Invalid style: {self.sep_style}')
252
+
253
+ def set_system_message(self, system_message: str):
254
+ """Set the system message."""
255
+ self.system_message = system_message
256
+
257
+ def append_message(self, role: str, message: str):
258
+ """Append a new message."""
259
+ self.messages.append([role, message])
260
+
261
+ def update_last_message(self, message: str):
262
+ """Update the last output.
263
+
264
+ The last message is typically set to be None when constructing the prompt,
265
+ so we need to update it in-place after getting the response from a model.
266
+ """
267
+ self.messages[-1][1] = message
268
+
269
+ def to_gradio_chatbot(self):
270
+ """Convert the conversation to gradio chatbot format."""
271
+ ret = []
272
+ for i, (role, msg) in enumerate(self.messages[self.offset :]):
273
+ if i % 2 == 0:
274
+ ret.append([msg, None])
275
+ else:
276
+ ret[-1][-1] = msg
277
+ return ret
278
+
279
+ def to_openai_api_messages(self):
280
+ """Convert the conversation to OpenAI chat completion format."""
281
+ ret = [{'role': 'system', 'content': self.system_message}]
282
+
283
+ for i, (_, msg) in enumerate(self.messages[self.offset :]):
284
+ if i % 2 == 0:
285
+ ret.append({'role': 'user', 'content': msg})
286
+ else:
287
+ if msg is not None:
288
+ ret.append({'role': 'assistant', 'content': msg})
289
+ return ret
290
+
291
+ def copy(self):
292
+ return Conversation(
293
+ name=self.name,
294
+ system_template=self.system_template,
295
+ system_message=self.system_message,
296
+ roles=self.roles,
297
+ messages=[[x, y] for x, y in self.messages],
298
+ offset=self.offset,
299
+ sep_style=self.sep_style,
300
+ sep=self.sep,
301
+ sep2=self.sep2,
302
+ stop_str=self.stop_str,
303
+ stop_token_ids=self.stop_token_ids,
304
+ )
305
+
306
+ def dict(self):
307
+ return {
308
+ 'template_name': self.name,
309
+ 'system_message': self.system_message,
310
+ 'roles': self.roles,
311
+ 'messages': self.messages,
312
+ 'offset': self.offset,
313
+ }
314
+
315
+
316
+ # A global registry for all conversation templates
317
+ conv_templates: Dict[str, Conversation] = {}
318
+
319
+
320
+ def register_conv_template(template: Conversation, override: bool = False):
321
+ """Register a new conversation template."""
322
+ if not override:
323
+ assert (
324
+ template.name not in conv_templates
325
+ ), f'{template.name} has been registered.'
326
+
327
+ conv_templates[template.name] = template
328
+
329
+
330
+ def get_conv_template(name: str) -> Conversation:
331
+ """Get a conversation template."""
332
+ return conv_templates[name].copy()
333
+
334
+
335
+ # Both Hermes-2 and internlm2-chat are chatml-format conversation templates. The difference
336
+ # is that during training, the preprocessing function for the Hermes-2 template doesn't add
337
+ # <s> at the beginning of the tokenized sequence, while the internlm2-chat template does.
338
+ # Therefore, they are completely equivalent during inference.
339
+ register_conv_template(
340
+ Conversation(
341
+ name='Hermes-2',
342
+ system_template='<|im_start|>system\n{system_message}',
343
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
344
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
345
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
346
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
347
+ sep_style=SeparatorStyle.MPT,
348
+ sep='<|im_end|>',
349
+ stop_str='<|endoftext|>',
350
+ )
351
+ )
352
+
353
+
354
+ register_conv_template(
355
+ Conversation(
356
+ name='internlm2-chat',
357
+ system_template='<|im_start|>system\n{system_message}',
358
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
359
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
360
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
361
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
362
+ sep_style=SeparatorStyle.MPT,
363
+ sep='<|im_end|>',
364
+ )
365
+ )
366
+
367
+
368
+ register_conv_template(
369
+ Conversation(
370
+ name='phi3-chat',
371
+ system_template='<|system|>\n{system_message}',
372
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
373
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
374
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
375
+ roles=('<|user|>\n', '<|assistant|>\n'),
376
+ sep_style=SeparatorStyle.MPT,
377
+ sep='<|end|>',
378
+ )
379
+ )
380
+
381
+
382
+ register_conv_template(
383
+ Conversation(
384
+ name='internvl2_5',
385
+ system_template='<|im_start|>system\n{system_message}',
386
+ system_message='你是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
387
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
388
+ sep_style=SeparatorStyle.MPT,
389
+ sep='<|im_end|>\n',
390
+ )
391
+ )
1B/20250802/generation_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "transformers_version": "4.37.2"
4
+ }
1B/20250802/merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
1B/20250802/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5347b0f129392ea157f7239b9d8b90db644a4eac470da2faee601829695e4908
3
+ size 1876463472
1B/20250802/modeling_intern_vit.py ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ from typing import Optional, Tuple, Union
8
+
9
+ import torch
10
+ import torch.nn.functional as F
11
+ import torch.utils.checkpoint
12
+ from einops import rearrange
13
+ from timm.layers import DropPath
14
+ from torch import nn
15
+ from transformers.activations import ACT2FN
16
+ from transformers.modeling_outputs import (BaseModelOutput,
17
+ BaseModelOutputWithPooling)
18
+ from transformers.modeling_utils import PreTrainedModel
19
+ from transformers.utils import logging
20
+
21
+ from .configuration_intern_vit import InternVisionConfig
22
+
23
+ try:
24
+ from flash_attn.bert_padding import pad_input, unpad_input
25
+ from flash_attn.flash_attn_interface import \
26
+ flash_attn_varlen_qkvpacked_func
27
+ has_flash_attn = True
28
+ except:
29
+ print('FlashAttention2 is not installed.')
30
+ has_flash_attn = False
31
+
32
+ logger = logging.get_logger(__name__)
33
+
34
+
35
+ class FlashAttention(nn.Module):
36
+ """Implement the scaled dot product attention with softmax.
37
+ Arguments
38
+ ---------
39
+ softmax_scale: The temperature to use for the softmax attention.
40
+ (default: 1/sqrt(d_keys) where d_keys is computed at
41
+ runtime)
42
+ attention_dropout: The dropout rate to apply to the attention
43
+ (default: 0.0)
44
+ """
45
+
46
+ def __init__(self, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None):
47
+ super().__init__()
48
+ self.softmax_scale = softmax_scale
49
+ self.dropout_p = attention_dropout
50
+
51
+ def forward(self, qkv, key_padding_mask=None, causal=False, cu_seqlens=None,
52
+ max_s=None, need_weights=False):
53
+ """Implements the multihead softmax attention.
54
+ Arguments
55
+ ---------
56
+ qkv: The tensor containing the query, key, and value. (B, S, 3, H, D) if key_padding_mask is None
57
+ if unpadded: (nnz, 3, h, d)
58
+ key_padding_mask: a bool tensor of shape (B, S)
59
+ """
60
+ assert not need_weights
61
+ assert qkv.dtype in [torch.float16, torch.bfloat16]
62
+ assert qkv.is_cuda
63
+
64
+ if cu_seqlens is None:
65
+ batch_size = qkv.shape[0]
66
+ seqlen = qkv.shape[1]
67
+ if key_padding_mask is None:
68
+ qkv = rearrange(qkv, 'b s ... -> (b s) ...')
69
+ max_s = seqlen
70
+ cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
71
+ device=qkv.device)
72
+ output = flash_attn_varlen_qkvpacked_func(
73
+ qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
74
+ softmax_scale=self.softmax_scale, causal=causal
75
+ )
76
+ output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
77
+ else:
78
+ nheads = qkv.shape[-2]
79
+ x = rearrange(qkv, 'b s three h d -> b s (three h d)')
80
+ x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)
81
+ x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
82
+ output_unpad = flash_attn_varlen_qkvpacked_func(
83
+ x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
84
+ softmax_scale=self.softmax_scale, causal=causal
85
+ )
86
+ output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),
87
+ indices, batch_size, seqlen),
88
+ 'b s (h d) -> b s h d', h=nheads)
89
+ else:
90
+ assert max_s is not None
91
+ output = flash_attn_varlen_qkvpacked_func(
92
+ qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
93
+ softmax_scale=self.softmax_scale, causal=causal
94
+ )
95
+
96
+ return output, None
97
+
98
+
99
+ class InternRMSNorm(nn.Module):
100
+ def __init__(self, hidden_size, eps=1e-6):
101
+ super().__init__()
102
+ self.weight = nn.Parameter(torch.ones(hidden_size))
103
+ self.variance_epsilon = eps
104
+
105
+ def forward(self, hidden_states):
106
+ input_dtype = hidden_states.dtype
107
+ hidden_states = hidden_states.to(torch.float32)
108
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
109
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
110
+ return self.weight * hidden_states.to(input_dtype)
111
+
112
+
113
+ try:
114
+ from apex.normalization import FusedRMSNorm
115
+
116
+ InternRMSNorm = FusedRMSNorm # noqa
117
+
118
+ logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')
119
+ except ImportError:
120
+ # using the normal InternRMSNorm
121
+ pass
122
+ except Exception:
123
+ logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm')
124
+ pass
125
+
126
+
127
+ NORM2FN = {
128
+ 'rms_norm': InternRMSNorm,
129
+ 'layer_norm': nn.LayerNorm,
130
+ }
131
+
132
+
133
+ class InternVisionEmbeddings(nn.Module):
134
+ def __init__(self, config: InternVisionConfig):
135
+ super().__init__()
136
+ self.config = config
137
+ self.embed_dim = config.hidden_size
138
+ self.image_size = config.image_size
139
+ self.patch_size = config.patch_size
140
+
141
+ self.class_embedding = nn.Parameter(
142
+ torch.randn(1, 1, self.embed_dim),
143
+ )
144
+
145
+ self.patch_embedding = nn.Conv2d(
146
+ in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
147
+ )
148
+
149
+ self.num_patches = (self.image_size // self.patch_size) ** 2
150
+ self.num_positions = self.num_patches + 1
151
+
152
+ self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
153
+
154
+ def _get_pos_embed(self, pos_embed, H, W):
155
+ target_dtype = pos_embed.dtype
156
+ pos_embed = pos_embed.float().reshape(
157
+ 1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
158
+ pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False). \
159
+ reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
160
+ return pos_embed
161
+
162
+ def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
163
+ target_dtype = self.patch_embedding.weight.dtype
164
+ patch_embeds = self.patch_embedding(pixel_values) # shape = [*, channel, width, height]
165
+ batch_size, _, height, width = patch_embeds.shape
166
+ patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
167
+ class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
168
+ embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
169
+ position_embedding = torch.cat([
170
+ self.position_embedding[:, :1, :],
171
+ self._get_pos_embed(self.position_embedding[:, 1:, :], height, width)
172
+ ], dim=1)
173
+ embeddings = embeddings + position_embedding.to(target_dtype)
174
+ return embeddings
175
+
176
+
177
+ class InternAttention(nn.Module):
178
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
179
+
180
+ def __init__(self, config: InternVisionConfig):
181
+ super().__init__()
182
+ self.config = config
183
+ self.embed_dim = config.hidden_size
184
+ self.num_heads = config.num_attention_heads
185
+ self.use_flash_attn = config.use_flash_attn and has_flash_attn
186
+ if config.use_flash_attn and not has_flash_attn:
187
+ print('Warning: Flash Attention is not available, use_flash_attn is set to False.')
188
+ self.head_dim = self.embed_dim // self.num_heads
189
+ if self.head_dim * self.num_heads != self.embed_dim:
190
+ raise ValueError(
191
+ f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:'
192
+ f' {self.num_heads}).'
193
+ )
194
+
195
+ self.scale = self.head_dim ** -0.5
196
+ self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias)
197
+ self.attn_drop = nn.Dropout(config.attention_dropout)
198
+ self.proj_drop = nn.Dropout(config.dropout)
199
+
200
+ self.qk_normalization = config.qk_normalization
201
+
202
+ if self.qk_normalization:
203
+ self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
204
+ self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
205
+
206
+ if self.use_flash_attn:
207
+ self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout)
208
+ self.proj = nn.Linear(self.embed_dim, self.embed_dim)
209
+
210
+ def _naive_attn(self, x):
211
+ B, N, C = x.shape
212
+ qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
213
+ q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple)
214
+
215
+ if self.qk_normalization:
216
+ B_, H_, N_, D_ = q.shape
217
+ q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
218
+ k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
219
+
220
+ attn = ((q * self.scale) @ k.transpose(-2, -1))
221
+ attn = attn.softmax(dim=-1)
222
+ attn = self.attn_drop(attn)
223
+
224
+ x = (attn @ v).transpose(1, 2).reshape(B, N, C)
225
+ x = self.proj(x)
226
+ x = self.proj_drop(x)
227
+ return x
228
+
229
+ def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
230
+ qkv = self.qkv(x)
231
+ qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads)
232
+
233
+ if self.qk_normalization:
234
+ q, k, v = qkv.unbind(2)
235
+ q = self.q_norm(q.flatten(-2, -1)).view(q.shape)
236
+ k = self.k_norm(k.flatten(-2, -1)).view(k.shape)
237
+ qkv = torch.stack([q, k, v], dim=2)
238
+
239
+ context, _ = self.inner_attn(
240
+ qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False
241
+ )
242
+ outs = self.proj(rearrange(context, 'b s h d -> b s (h d)'))
243
+ outs = self.proj_drop(outs)
244
+ return outs
245
+
246
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
247
+ x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)
248
+ return x
249
+
250
+
251
+ class InternMLP(nn.Module):
252
+ def __init__(self, config: InternVisionConfig):
253
+ super().__init__()
254
+ self.config = config
255
+ self.act = ACT2FN[config.hidden_act]
256
+ self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
257
+ self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
258
+
259
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
260
+ hidden_states = self.fc1(hidden_states)
261
+ hidden_states = self.act(hidden_states)
262
+ hidden_states = self.fc2(hidden_states)
263
+ return hidden_states
264
+
265
+
266
+ class InternVisionEncoderLayer(nn.Module):
267
+ def __init__(self, config: InternVisionConfig, drop_path_rate: float):
268
+ super().__init__()
269
+ self.embed_dim = config.hidden_size
270
+ self.intermediate_size = config.intermediate_size
271
+ self.norm_type = config.norm_type
272
+
273
+ self.attn = InternAttention(config)
274
+ self.mlp = InternMLP(config)
275
+ self.norm1 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
276
+ self.norm2 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
277
+
278
+ self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
279
+ self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
280
+ self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
281
+ self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
282
+
283
+ def forward(
284
+ self,
285
+ hidden_states: torch.Tensor,
286
+ ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]:
287
+ """
288
+ Args:
289
+ hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
290
+ """
291
+ hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states).to(hidden_states.dtype)) * self.ls1)
292
+
293
+ hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states).to(hidden_states.dtype)) * self.ls2)
294
+
295
+ return hidden_states
296
+
297
+
298
+ class InternVisionEncoder(nn.Module):
299
+ """
300
+ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
301
+ [`InternEncoderLayer`].
302
+
303
+ Args:
304
+ config (`InternConfig`):
305
+ The corresponding vision configuration for the `InternEncoder`.
306
+ """
307
+
308
+ def __init__(self, config: InternVisionConfig):
309
+ super().__init__()
310
+ self.config = config
311
+ # stochastic depth decay rule
312
+ dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]
313
+ self.layers = nn.ModuleList([
314
+ InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
315
+ self.gradient_checkpointing = True
316
+
317
+ def forward(
318
+ self,
319
+ inputs_embeds,
320
+ output_hidden_states: Optional[bool] = None,
321
+ return_dict: Optional[bool] = None,
322
+ ) -> Union[Tuple, BaseModelOutput]:
323
+ r"""
324
+ Args:
325
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
326
+ Embedded representation of the inputs. Should be float, not int tokens.
327
+ output_hidden_states (`bool`, *optional*):
328
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
329
+ for more detail.
330
+ return_dict (`bool`, *optional*):
331
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
332
+ """
333
+ output_hidden_states = (
334
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
335
+ )
336
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
337
+
338
+ encoder_states = () if output_hidden_states else None
339
+ hidden_states = inputs_embeds
340
+
341
+ for idx, encoder_layer in enumerate(self.layers):
342
+ if output_hidden_states:
343
+ encoder_states = encoder_states + (hidden_states,)
344
+ if self.gradient_checkpointing and self.training:
345
+ layer_outputs = torch.utils.checkpoint.checkpoint(
346
+ encoder_layer,
347
+ hidden_states)
348
+ else:
349
+ layer_outputs = encoder_layer(
350
+ hidden_states,
351
+ )
352
+ hidden_states = layer_outputs
353
+
354
+ if output_hidden_states:
355
+ encoder_states = encoder_states + (hidden_states,)
356
+
357
+ if not return_dict:
358
+ return tuple(v for v in [hidden_states, encoder_states] if v is not None)
359
+ return BaseModelOutput(
360
+ last_hidden_state=hidden_states, hidden_states=encoder_states
361
+ )
362
+
363
+
364
+ class InternVisionModel(PreTrainedModel):
365
+ main_input_name = 'pixel_values'
366
+ _supports_flash_attn_2 = True
367
+ supports_gradient_checkpointing = True
368
+ config_class = InternVisionConfig
369
+ _no_split_modules = ['InternVisionEncoderLayer']
370
+
371
+ def __init__(self, config: InternVisionConfig):
372
+ super().__init__(config)
373
+ self.config = config
374
+
375
+ self.embeddings = InternVisionEmbeddings(config)
376
+ self.encoder = InternVisionEncoder(config)
377
+
378
+ def resize_pos_embeddings(self, old_size, new_size, patch_size):
379
+ pos_emb = self.embeddings.position_embedding
380
+ _, num_positions, embed_dim = pos_emb.shape
381
+ cls_emb = pos_emb[:, :1, :]
382
+ pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)
383
+ pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False)
384
+ pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1)
385
+ pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
386
+ self.embeddings.position_embedding = nn.Parameter(pos_emb)
387
+ self.embeddings.image_size = new_size
388
+ logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size))
389
+
390
+ def get_input_embeddings(self):
391
+ return self.embeddings
392
+
393
+ def forward(
394
+ self,
395
+ pixel_values: Optional[torch.FloatTensor] = None,
396
+ output_hidden_states: Optional[bool] = None,
397
+ return_dict: Optional[bool] = None,
398
+ pixel_embeds: Optional[torch.FloatTensor] = None,
399
+ ) -> Union[Tuple, BaseModelOutputWithPooling]:
400
+ output_hidden_states = (
401
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
402
+ )
403
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
404
+
405
+ if pixel_values is None and pixel_embeds is None:
406
+ raise ValueError('You have to specify pixel_values or pixel_embeds')
407
+
408
+ if pixel_embeds is not None:
409
+ hidden_states = pixel_embeds
410
+ else:
411
+ if len(pixel_values.shape) == 4:
412
+ hidden_states = self.embeddings(pixel_values)
413
+ else:
414
+ raise ValueError(f'wrong pixel_values size: {pixel_values.shape}')
415
+ encoder_outputs = self.encoder(
416
+ inputs_embeds=hidden_states,
417
+ output_hidden_states=output_hidden_states,
418
+ return_dict=return_dict,
419
+ )
420
+ last_hidden_state = encoder_outputs.last_hidden_state
421
+ pooled_output = last_hidden_state[:, 0, :]
422
+
423
+ if not return_dict:
424
+ return (last_hidden_state, pooled_output) + encoder_outputs[1:]
425
+
426
+ return BaseModelOutputWithPooling(
427
+ last_hidden_state=last_hidden_state,
428
+ pooler_output=pooled_output,
429
+ hidden_states=encoder_outputs.hidden_states,
430
+ attentions=encoder_outputs.attentions,
431
+ )
1B/20250802/modeling_internvl_chat.py ADDED
@@ -0,0 +1,359 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import warnings
8
+ from typing import List, Optional, Tuple, Union
9
+
10
+ import torch.utils.checkpoint
11
+ import transformers
12
+ from torch import nn
13
+ from torch.nn import CrossEntropyLoss
14
+ from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
15
+ Qwen2ForCausalLM)
16
+ from transformers.modeling_outputs import CausalLMOutputWithPast
17
+ from transformers.modeling_utils import PreTrainedModel
18
+ from transformers.utils import ModelOutput, logging
19
+
20
+ from .configuration_internvl_chat import InternVLChatConfig
21
+ from .conversation import get_conv_template
22
+ from .modeling_intern_vit import InternVisionModel, has_flash_attn
23
+
24
+ logger = logging.get_logger(__name__)
25
+
26
+
27
+ def version_cmp(v1, v2, op='eq'):
28
+ import operator
29
+
30
+ from packaging import version
31
+ op_func = getattr(operator, op)
32
+ return op_func(version.parse(v1), version.parse(v2))
33
+
34
+
35
+ class InternVLChatModel(PreTrainedModel):
36
+ config_class = InternVLChatConfig
37
+ main_input_name = 'pixel_values'
38
+ base_model_prefix = 'language_model'
39
+ _supports_flash_attn_2 = True
40
+ supports_gradient_checkpointing = True
41
+ _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'Qwen2DecoderLayer']
42
+
43
+ def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True):
44
+ super().__init__(config)
45
+
46
+ assert version_cmp(transformers.__version__, '4.37.0', 'ge')
47
+ image_size = config.force_image_size or config.vision_config.image_size
48
+ patch_size = config.vision_config.patch_size
49
+ self.patch_size = patch_size
50
+ self.select_layer = config.select_layer
51
+ self.template = config.template
52
+ self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
53
+ self.downsample_ratio = config.downsample_ratio
54
+ self.ps_version = config.ps_version
55
+ use_flash_attn = use_flash_attn if has_flash_attn else False
56
+ config.vision_config.use_flash_attn = True if use_flash_attn else False
57
+ config.llm_config._attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'
58
+
59
+ logger.info(f'num_image_token: {self.num_image_token}')
60
+ logger.info(f'ps_version: {self.ps_version}')
61
+ if vision_model is not None:
62
+ self.vision_model = vision_model
63
+ else:
64
+ self.vision_model = InternVisionModel(config.vision_config)
65
+ if language_model is not None:
66
+ self.language_model = language_model
67
+ else:
68
+ if config.llm_config.architectures[0] == 'LlamaForCausalLM':
69
+ self.language_model = LlamaForCausalLM(config.llm_config)
70
+ elif config.llm_config.architectures[0] == 'Qwen2ForCausalLM':
71
+ self.language_model = Qwen2ForCausalLM(config.llm_config)
72
+ else:
73
+ raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')
74
+
75
+ vit_hidden_size = config.vision_config.hidden_size
76
+ llm_hidden_size = config.llm_config.hidden_size
77
+
78
+ self.mlp1 = nn.Sequential(
79
+ nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
80
+ nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
81
+ nn.GELU(),
82
+ nn.Linear(llm_hidden_size, llm_hidden_size)
83
+ )
84
+
85
+ self.img_context_token_id = None
86
+ self.conv_template = get_conv_template(self.template)
87
+ self.system_message = self.conv_template.system_message
88
+
89
+ def forward(
90
+ self,
91
+ pixel_values: torch.FloatTensor,
92
+ input_ids: torch.LongTensor = None,
93
+ attention_mask: Optional[torch.Tensor] = None,
94
+ position_ids: Optional[torch.LongTensor] = None,
95
+ image_flags: Optional[torch.LongTensor] = None,
96
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
97
+ labels: Optional[torch.LongTensor] = None,
98
+ use_cache: Optional[bool] = None,
99
+ output_attentions: Optional[bool] = None,
100
+ output_hidden_states: Optional[bool] = None,
101
+ return_dict: Optional[bool] = None,
102
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
103
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
104
+
105
+ image_flags = image_flags.squeeze(-1)
106
+ input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
107
+
108
+ vit_embeds = self.extract_feature(pixel_values)
109
+ vit_embeds = vit_embeds[image_flags == 1]
110
+ vit_batch_size = pixel_values.shape[0]
111
+
112
+ B, N, C = input_embeds.shape
113
+ input_embeds = input_embeds.reshape(B * N, C)
114
+
115
+ if torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
116
+ print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')
117
+
118
+ input_ids = input_ids.reshape(B * N)
119
+ selected = (input_ids == self.img_context_token_id)
120
+ try:
121
+ input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
122
+ except Exception as e:
123
+ vit_embeds = vit_embeds.reshape(-1, C)
124
+ print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
125
+ f'vit_embeds.shape={vit_embeds.shape}')
126
+ n_token = min(selected.sum(), vit_embeds.size(0))
127
+ input_embeds[selected][:n_token] = input_embeds[selected][:n_token] * 0.0 + vit_embeds[:n_token]
128
+
129
+ input_embeds = input_embeds.reshape(B, N, C)
130
+
131
+ outputs = self.language_model(
132
+ inputs_embeds=input_embeds,
133
+ attention_mask=attention_mask,
134
+ position_ids=position_ids,
135
+ past_key_values=past_key_values,
136
+ use_cache=use_cache,
137
+ output_attentions=output_attentions,
138
+ output_hidden_states=output_hidden_states,
139
+ return_dict=return_dict,
140
+ )
141
+ logits = outputs.logits
142
+
143
+ loss = None
144
+ if labels is not None:
145
+ # Shift so that tokens < n predict n
146
+ shift_logits = logits[..., :-1, :].contiguous()
147
+ shift_labels = labels[..., 1:].contiguous()
148
+ # Flatten the tokens
149
+ loss_fct = CrossEntropyLoss()
150
+ shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
151
+ shift_labels = shift_labels.view(-1)
152
+ # Enable model parallelism
153
+ shift_labels = shift_labels.to(shift_logits.device)
154
+ loss = loss_fct(shift_logits, shift_labels)
155
+
156
+ if not return_dict:
157
+ output = (logits,) + outputs[1:]
158
+ return (loss,) + output if loss is not None else output
159
+
160
+ return CausalLMOutputWithPast(
161
+ loss=loss,
162
+ logits=logits,
163
+ past_key_values=outputs.past_key_values,
164
+ hidden_states=outputs.hidden_states,
165
+ attentions=outputs.attentions,
166
+ )
167
+
168
+ def pixel_shuffle(self, x, scale_factor=0.5):
169
+ n, w, h, c = x.size()
170
+ # N, W, H, C --> N, W, H * scale, C // scale
171
+ x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
172
+ # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
173
+ x = x.permute(0, 2, 1, 3).contiguous()
174
+ # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
175
+ x = x.view(n, int(h * scale_factor), int(w * scale_factor),
176
+ int(c / (scale_factor * scale_factor)))
177
+ if self.ps_version == 'v1':
178
+ warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
179
+ 'which results in a transposed image.')
180
+ else:
181
+ x = x.permute(0, 2, 1, 3).contiguous()
182
+ return x
183
+
184
+ def extract_feature(self, pixel_values):
185
+ if self.select_layer == -1:
186
+ vit_embeds = self.vision_model(
187
+ pixel_values=pixel_values,
188
+ output_hidden_states=False,
189
+ return_dict=True).last_hidden_state
190
+ else:
191
+ vit_embeds = self.vision_model(
192
+ pixel_values=pixel_values,
193
+ output_hidden_states=True,
194
+ return_dict=True).hidden_states[self.select_layer]
195
+ vit_embeds = vit_embeds[:, 1:, :]
196
+
197
+ h = w = int(vit_embeds.shape[1] ** 0.5)
198
+ vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
199
+ vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
200
+ vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
201
+ vit_embeds = self.mlp1(vit_embeds)
202
+ return vit_embeds
203
+
204
+ def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None,
205
+ history=None, return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
206
+ IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False, image_counts=None):
207
+ if history is not None or return_history:
208
+ print('Now multi-turn chat is not supported in batch_chat.')
209
+ raise NotImplementedError
210
+
211
+ if image_counts is not None:
212
+ num_patches_list = image_counts
213
+ print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.')
214
+
215
+ img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
216
+ self.img_context_token_id = img_context_token_id
217
+
218
+ if verbose and pixel_values is not None:
219
+ image_bs = pixel_values.shape[0]
220
+ print(f'dynamic ViT batch size: {image_bs}')
221
+
222
+ queries = []
223
+ for idx, num_patches in enumerate(num_patches_list):
224
+ question = questions[idx]
225
+ if pixel_values is not None and '<image>' not in question:
226
+ question = '<image>\n' + question
227
+ template = get_conv_template(self.template)
228
+ template.system_message = self.system_message
229
+ template.append_message(template.roles[0], question)
230
+ template.append_message(template.roles[1], None)
231
+ query = template.get_prompt()
232
+
233
+ image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
234
+ query = query.replace('<image>', image_tokens, 1)
235
+ queries.append(query)
236
+
237
+ tokenizer.padding_side = 'left'
238
+ model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
239
+ input_ids = model_inputs['input_ids'].to(self.device)
240
+ attention_mask = model_inputs['attention_mask'].to(self.device)
241
+ eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
242
+ generation_config['eos_token_id'] = eos_token_id
243
+ generation_output = self.generate(
244
+ pixel_values=pixel_values,
245
+ input_ids=input_ids,
246
+ attention_mask=attention_mask,
247
+ **generation_config
248
+ )
249
+ responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
250
+ responses = [response.split(template.sep.strip())[0].strip() for response in responses]
251
+ return responses
252
+
253
+ def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
254
+ num_patches_list=None, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
255
+ verbose=False):
256
+
257
+ if history is None and pixel_values is not None and '<image>' not in question:
258
+ question = '<image>\n' + question
259
+
260
+ if num_patches_list is None:
261
+ num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
262
+ assert pixel_values is None or len(pixel_values) == sum(num_patches_list)
263
+
264
+ img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
265
+ self.img_context_token_id = img_context_token_id
266
+
267
+ template = get_conv_template(self.template)
268
+ template.system_message = self.system_message
269
+ eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
270
+
271
+ history = [] if history is None else history
272
+ for (old_question, old_answer) in history:
273
+ template.append_message(template.roles[0], old_question)
274
+ template.append_message(template.roles[1], old_answer)
275
+ template.append_message(template.roles[0], question)
276
+ template.append_message(template.roles[1], None)
277
+ query = template.get_prompt()
278
+
279
+ if verbose and pixel_values is not None:
280
+ image_bs = pixel_values.shape[0]
281
+ print(f'dynamic ViT batch size: {image_bs}')
282
+
283
+ for num_patches in num_patches_list:
284
+ image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
285
+ query = query.replace('<image>', image_tokens, 1)
286
+
287
+ model_inputs = tokenizer(query, return_tensors='pt')
288
+ input_ids = model_inputs['input_ids'].to(self.device)
289
+ attention_mask = model_inputs['attention_mask'].to(self.device)
290
+ generation_config['eos_token_id'] = eos_token_id
291
+ generation_output = self.generate(
292
+ pixel_values=pixel_values,
293
+ input_ids=input_ids,
294
+ attention_mask=attention_mask,
295
+ **generation_config
296
+ )
297
+ response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
298
+ response = response.split(template.sep.strip())[0].strip()
299
+ history.append((question, response))
300
+ if return_history:
301
+ return response, history
302
+ else:
303
+ query_to_print = query.replace(IMG_CONTEXT_TOKEN, '')
304
+ query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
305
+ if verbose:
306
+ print(query_to_print, response)
307
+ return response
308
+
309
+ @torch.no_grad()
310
+ def generate(
311
+ self,
312
+ pixel_values: Optional[torch.FloatTensor] = None,
313
+ input_ids: Optional[torch.FloatTensor] = None,
314
+ attention_mask: Optional[torch.LongTensor] = None,
315
+ visual_features: Optional[torch.FloatTensor] = None,
316
+ generation_config: Optional[GenerationConfig] = None,
317
+ output_hidden_states: Optional[bool] = None,
318
+ **generate_kwargs,
319
+ ) -> torch.LongTensor:
320
+
321
+ assert self.img_context_token_id is not None
322
+ if pixel_values is not None:
323
+ if visual_features is not None:
324
+ vit_embeds = visual_features
325
+ else:
326
+ vit_embeds = self.extract_feature(pixel_values)
327
+ input_embeds = self.language_model.get_input_embeddings()(input_ids)
328
+ B, N, C = input_embeds.shape
329
+ input_embeds = input_embeds.reshape(B * N, C)
330
+
331
+ input_ids = input_ids.reshape(B * N)
332
+ selected = (input_ids == self.img_context_token_id)
333
+ assert selected.sum() != 0
334
+ input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
335
+
336
+ input_embeds = input_embeds.reshape(B, N, C)
337
+ else:
338
+ input_embeds = self.language_model.get_input_embeddings()(input_ids)
339
+
340
+ outputs = self.language_model.generate(
341
+ inputs_embeds=input_embeds,
342
+ attention_mask=attention_mask,
343
+ generation_config=generation_config,
344
+ output_hidden_states=output_hidden_states,
345
+ use_cache=True,
346
+ **generate_kwargs,
347
+ )
348
+
349
+ return outputs
350
+
351
+ @property
352
+ def lm_head(self):
353
+ return self.language_model.get_output_embeddings()
354
+
355
+ def get_input_embeddings(self):
356
+ return self.language_model.get_input_embeddings()
357
+
358
+ def get_output_embeddings(self):
359
+ return self.language_model.get_output_embeddings()
1B/20250802/special_tokens_map.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|im_start|>",
4
+ "<|im_end|>",
5
+ "<|object_ref_start|>",
6
+ "<|object_ref_end|>",
7
+ "<|box_start|>",
8
+ "<|box_end|>",
9
+ "<|quad_start|>",
10
+ "<|quad_end|>",
11
+ "<|vision_start|>",
12
+ "<|vision_end|>",
13
+ "<|vision_pad|>",
14
+ "<|image_pad|>",
15
+ "<|video_pad|>"
16
+ ],
17
+ "eos_token": {
18
+ "content": "<|im_end|>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ },
24
+ "pad_token": {
25
+ "content": "<|endoftext|>",
26
+ "lstrip": false,
27
+ "normalized": false,
28
+ "rstrip": false,
29
+ "single_word": false
30
+ }
31
+ }
1B/20250802/tokenizer_config.json ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "151643": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "151644": {
15
+ "content": "<|im_start|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "151645": {
23
+ "content": "<|im_end|>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": true
29
+ },
30
+ "151646": {
31
+ "content": "<|object_ref_start|>",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": true
37
+ },
38
+ "151647": {
39
+ "content": "<|object_ref_end|>",
40
+ "lstrip": false,
41
+ "normalized": false,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": true
45
+ },
46
+ "151648": {
47
+ "content": "<|box_start|>",
48
+ "lstrip": false,
49
+ "normalized": false,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": true
53
+ },
54
+ "151649": {
55
+ "content": "<|box_end|>",
56
+ "lstrip": false,
57
+ "normalized": false,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": true
61
+ },
62
+ "151650": {
63
+ "content": "<|quad_start|>",
64
+ "lstrip": false,
65
+ "normalized": false,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": true
69
+ },
70
+ "151651": {
71
+ "content": "<|quad_end|>",
72
+ "lstrip": false,
73
+ "normalized": false,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": true
77
+ },
78
+ "151652": {
79
+ "content": "<|vision_start|>",
80
+ "lstrip": false,
81
+ "normalized": false,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": true
85
+ },
86
+ "151653": {
87
+ "content": "<|vision_end|>",
88
+ "lstrip": false,
89
+ "normalized": false,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": true
93
+ },
94
+ "151654": {
95
+ "content": "<|vision_pad|>",
96
+ "lstrip": false,
97
+ "normalized": false,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": true
101
+ },
102
+ "151655": {
103
+ "content": "<|image_pad|>",
104
+ "lstrip": false,
105
+ "normalized": false,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": true
109
+ },
110
+ "151656": {
111
+ "content": "<|video_pad|>",
112
+ "lstrip": false,
113
+ "normalized": false,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": true
117
+ },
118
+ "151657": {
119
+ "content": "<tool_call>",
120
+ "lstrip": false,
121
+ "normalized": false,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "151658": {
127
+ "content": "</tool_call>",
128
+ "lstrip": false,
129
+ "normalized": false,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "151659": {
135
+ "content": "<|fim_prefix|>",
136
+ "lstrip": false,
137
+ "normalized": false,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "151660": {
143
+ "content": "<|fim_middle|>",
144
+ "lstrip": false,
145
+ "normalized": false,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "151661": {
151
+ "content": "<|fim_suffix|>",
152
+ "lstrip": false,
153
+ "normalized": false,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "151662": {
159
+ "content": "<|fim_pad|>",
160
+ "lstrip": false,
161
+ "normalized": false,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "151663": {
167
+ "content": "<|repo_name|>",
168
+ "lstrip": false,
169
+ "normalized": false,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "151664": {
175
+ "content": "<|file_sep|>",
176
+ "lstrip": false,
177
+ "normalized": false,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "151665": {
183
+ "content": "<img>",
184
+ "lstrip": false,
185
+ "normalized": false,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": true
189
+ },
190
+ "151666": {
191
+ "content": "</img>",
192
+ "lstrip": false,
193
+ "normalized": false,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": true
197
+ },
198
+ "151667": {
199
+ "content": "<IMG_CONTEXT>",
200
+ "lstrip": false,
201
+ "normalized": false,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": true
205
+ },
206
+ "151668": {
207
+ "content": "<quad>",
208
+ "lstrip": false,
209
+ "normalized": false,
210
+ "rstrip": false,
211
+ "single_word": false,
212
+ "special": true
213
+ },
214
+ "151669": {
215
+ "content": "</quad>",
216
+ "lstrip": false,
217
+ "normalized": false,
218
+ "rstrip": false,
219
+ "single_word": false,
220
+ "special": true
221
+ },
222
+ "151670": {
223
+ "content": "<ref>",
224
+ "lstrip": false,
225
+ "normalized": false,
226
+ "rstrip": false,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "151671": {
231
+ "content": "</ref>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": false,
235
+ "single_word": false,
236
+ "special": true
237
+ },
238
+ "151672": {
239
+ "content": "<box>",
240
+ "lstrip": false,
241
+ "normalized": false,
242
+ "rstrip": false,
243
+ "single_word": false,
244
+ "special": true
245
+ },
246
+ "151673": {
247
+ "content": "</box>",
248
+ "lstrip": false,
249
+ "normalized": false,
250
+ "rstrip": false,
251
+ "single_word": false,
252
+ "special": true
253
+ }
254
+ },
255
+ "additional_special_tokens": [
256
+ "<|im_start|>",
257
+ "<|im_end|>",
258
+ "<|object_ref_start|>",
259
+ "<|object_ref_end|>",
260
+ "<|box_start|>",
261
+ "<|box_end|>",
262
+ "<|quad_start|>",
263
+ "<|quad_end|>",
264
+ "<|vision_start|>",
265
+ "<|vision_end|>",
266
+ "<|vision_pad|>",
267
+ "<|image_pad|>",
268
+ "<|video_pad|>"
269
+ ],
270
+ "bos_token": null,
271
+ "chat_template": "{%- if tools %}\n {{- '<|im_start|>system\\n' }}\n {%- if messages[0]['role'] == 'system' %}\n {{- messages[0]['content'] }}\n {%- else %}\n {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n {%- endif %}\n {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n {%- for tool in tools %}\n {{- \"\\n\" }}\n {{- tool | tojson }}\n {%- endfor %}\n {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n {%- if messages[0]['role'] == 'system' %}\n {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n {%- else %}\n {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n {%- elif message.role == \"assistant\" %}\n {{- '<|im_start|>' + message.role }}\n {%- if message.content %}\n {{- '\\n' + message.content }}\n {%- endif %}\n {%- for tool_call in message.tool_calls %}\n {%- if tool_call.function is defined %}\n {%- set tool_call = tool_call.function %}\n {%- endif %}\n {{- '\\n<tool_call>\\n{\"name\": \"' }}\n {{- tool_call.name }}\n {{- '\", \"arguments\": ' }}\n {{- tool_call.arguments | tojson }}\n {{- '}\\n</tool_call>' }}\n {%- endfor %}\n {{- '<|im_end|>\\n' }}\n {%- elif message.role == \"tool\" %}\n {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n {{- '<|im_start|>user' }}\n {%- endif %}\n {{- '\\n<tool_response>\\n' }}\n {{- message.content }}\n {{- '\\n</tool_response>' }}\n {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n {{- '<|im_end|>\\n' }}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
272
+ "clean_up_tokenization_spaces": false,
273
+ "eos_token": "<|im_end|>",
274
+ "errors": "replace",
275
+ "extra_special_tokens": {},
276
+ "model_max_length": 16384,
277
+ "pad_token": "<|endoftext|>",
278
+ "split_special_tokens": false,
279
+ "tokenizer_class": "Qwen2Tokenizer",
280
+ "unk_token": null
281
+ }
1B/20250802/train_results.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 1.0,
3
+ "train_loss": 0.8273731040954589,
4
+ "train_runtime": 16224.5195,
5
+ "train_samples": -1,
6
+ "train_samples_per_second": 15.779,
7
+ "train_steps_per_second": 0.123
8
+ }
1B/20250802/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
1B/20250802/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9af8976154f5a4a72a8e923703a48d66e9f740dd088df651b269201f804d7096
3
+ size 6776
1B/20250802/training_log.txt ADDED
The diff for this file is too large to render. See raw diff
 
1B/20250802/vocab.json ADDED
The diff for this file is too large to render. See raw diff
 
1B/20250802/zero_to_fp32.py ADDED
@@ -0,0 +1,604 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+
3
+ # Copyright (c) Microsoft Corporation.
4
+ # SPDX-License-Identifier: Apache-2.0
5
+
6
+ # DeepSpeed Team
7
+
8
+ # This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
9
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
10
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
11
+ # application.
12
+ #
13
+ # example: python zero_to_fp32.py . pytorch_model.bin
14
+
15
+ import argparse
16
+ import torch
17
+ import glob
18
+ import math
19
+ import os
20
+ import re
21
+ from collections import OrderedDict
22
+ from dataclasses import dataclass
23
+
24
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
25
+ # DeepSpeed data structures it has to be available in the current python environment.
26
+ from deepspeed.utils import logger
27
+ from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
28
+ FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
29
+ FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
30
+
31
+
32
+ @dataclass
33
+ class zero_model_state:
34
+ buffers: dict()
35
+ param_shapes: dict()
36
+ shared_params: list
37
+ ds_version: int
38
+ frozen_param_shapes: dict()
39
+ frozen_param_fragments: dict()
40
+
41
+
42
+ debug = 0
43
+
44
+ # load to cpu
45
+ device = torch.device('cpu')
46
+
47
+
48
+ def atoi(text):
49
+ return int(text) if text.isdigit() else text
50
+
51
+
52
+ def natural_keys(text):
53
+ '''
54
+ alist.sort(key=natural_keys) sorts in human order
55
+ http://nedbatchelder.com/blog/200712/human_sorting.html
56
+ (See Toothy's implementation in the comments)
57
+ '''
58
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
59
+
60
+
61
+ def get_model_state_file(checkpoint_dir, zero_stage):
62
+ if not os.path.isdir(checkpoint_dir):
63
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
64
+
65
+ # there should be only one file
66
+ if zero_stage <= 2:
67
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
68
+ elif zero_stage == 3:
69
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
70
+
71
+ if not os.path.exists(file):
72
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
73
+
74
+ return file
75
+
76
+
77
+ def get_checkpoint_files(checkpoint_dir, glob_pattern):
78
+ # XXX: need to test that this simple glob rule works for multi-node setup too
79
+ ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
80
+
81
+ if len(ckpt_files) == 0:
82
+ raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
83
+
84
+ return ckpt_files
85
+
86
+
87
+ def get_optim_files(checkpoint_dir):
88
+ return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
89
+
90
+
91
+ def get_model_state_files(checkpoint_dir):
92
+ return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
93
+
94
+
95
+ def parse_model_states(files):
96
+ zero_model_states = []
97
+ for file in files:
98
+ state_dict = torch.load(file, map_location=device)
99
+
100
+ if BUFFER_NAMES not in state_dict:
101
+ raise ValueError(f"{file} is not a model state checkpoint")
102
+ buffer_names = state_dict[BUFFER_NAMES]
103
+ if debug:
104
+ print("Found buffers:", buffer_names)
105
+
106
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
107
+ buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
108
+ param_shapes = state_dict[PARAM_SHAPES]
109
+
110
+ # collect parameters that are included in param_shapes
111
+ param_names = []
112
+ for s in param_shapes:
113
+ for name in s.keys():
114
+ param_names.append(name)
115
+
116
+ # update with frozen parameters
117
+ frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
118
+ if frozen_param_shapes is not None:
119
+ if debug:
120
+ print(f"Found frozen_param_shapes: {frozen_param_shapes}")
121
+ param_names += list(frozen_param_shapes.keys())
122
+
123
+ # handle shared params
124
+ shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
125
+
126
+ ds_version = state_dict.get(DS_VERSION, None)
127
+
128
+ frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
129
+
130
+ z_model_state = zero_model_state(buffers=buffers,
131
+ param_shapes=param_shapes,
132
+ shared_params=shared_params,
133
+ ds_version=ds_version,
134
+ frozen_param_shapes=frozen_param_shapes,
135
+ frozen_param_fragments=frozen_param_fragments)
136
+ zero_model_states.append(z_model_state)
137
+
138
+ return zero_model_states
139
+
140
+
141
+ def parse_optim_states(files, ds_checkpoint_dir):
142
+
143
+ total_files = len(files)
144
+ state_dicts = []
145
+ for f in files:
146
+ state_dict = torch.load(f, map_location=device)
147
+ # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
148
+ # and also handle the case where it was already removed by another helper script
149
+ state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
150
+ state_dicts.append(state_dict)
151
+
152
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
153
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
154
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
155
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
156
+
157
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
158
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
159
+ # use the max of the partition_count to get the dp world_size.
160
+
161
+ if type(world_size) is list:
162
+ world_size = max(world_size)
163
+
164
+ if world_size != total_files:
165
+ raise ValueError(
166
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
167
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
168
+ )
169
+
170
+ # the groups are named differently in each stage
171
+ if zero_stage <= 2:
172
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
173
+ elif zero_stage == 3:
174
+ fp32_groups_key = FP32_FLAT_GROUPS
175
+ else:
176
+ raise ValueError(f"unknown zero stage {zero_stage}")
177
+
178
+ if zero_stage <= 2:
179
+ fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
180
+ elif zero_stage == 3:
181
+ # if there is more than one param group, there will be multiple flattened tensors - one
182
+ # flattened tensor per group - for simplicity merge them into a single tensor
183
+ #
184
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
185
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
186
+
187
+ fp32_flat_groups = [
188
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
189
+ ]
190
+
191
+ return zero_stage, world_size, fp32_flat_groups
192
+
193
+
194
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
195
+ """
196
+ Returns fp32 state_dict reconstructed from ds checkpoint
197
+
198
+ Args:
199
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
200
+
201
+ """
202
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
203
+
204
+ optim_files = get_optim_files(ds_checkpoint_dir)
205
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
206
+ print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
207
+
208
+ model_files = get_model_state_files(ds_checkpoint_dir)
209
+
210
+ zero_model_states = parse_model_states(model_files)
211
+ print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
212
+
213
+ if zero_stage <= 2:
214
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
215
+ exclude_frozen_parameters)
216
+ elif zero_stage == 3:
217
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
218
+ exclude_frozen_parameters)
219
+
220
+
221
+ def _zero2_merge_frozen_params(state_dict, zero_model_states):
222
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
223
+ return
224
+
225
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
226
+ frozen_param_fragments = zero_model_states[0].frozen_param_fragments
227
+
228
+ if debug:
229
+ num_elem = sum(s.numel() for s in frozen_param_shapes.values())
230
+ print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
231
+
232
+ wanted_params = len(frozen_param_shapes)
233
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
234
+ avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
235
+ print(f'Frozen params: Have {avail_numel} numels to process.')
236
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
237
+
238
+ total_params = 0
239
+ total_numel = 0
240
+ for name, shape in frozen_param_shapes.items():
241
+ total_params += 1
242
+ unpartitioned_numel = shape.numel()
243
+ total_numel += unpartitioned_numel
244
+
245
+ state_dict[name] = frozen_param_fragments[name]
246
+
247
+ if debug:
248
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
249
+
250
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
251
+
252
+
253
+ def _has_callable(obj, fn):
254
+ attr = getattr(obj, fn, None)
255
+ return callable(attr)
256
+
257
+
258
+ def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
259
+ param_shapes = zero_model_states[0].param_shapes
260
+
261
+ # Reconstruction protocol:
262
+ #
263
+ # XXX: document this
264
+
265
+ if debug:
266
+ for i in range(world_size):
267
+ for j in range(len(fp32_flat_groups[0])):
268
+ print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
269
+
270
+ # XXX: memory usage doubles here (zero2)
271
+ num_param_groups = len(fp32_flat_groups[0])
272
+ merged_single_partition_of_fp32_groups = []
273
+ for i in range(num_param_groups):
274
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
275
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
276
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
277
+ avail_numel = sum(
278
+ [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
279
+
280
+ if debug:
281
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
282
+ wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
283
+ # not asserting if there is a mismatch due to possible padding
284
+ print(f"Have {avail_numel} numels to process.")
285
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
286
+
287
+ # params
288
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
289
+ # out-of-core computing solution
290
+ total_numel = 0
291
+ total_params = 0
292
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
293
+ offset = 0
294
+ avail_numel = full_single_fp32_vector.numel()
295
+ for name, shape in shapes.items():
296
+
297
+ unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
298
+ total_numel += unpartitioned_numel
299
+ total_params += 1
300
+
301
+ if debug:
302
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
303
+ state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
304
+ offset += unpartitioned_numel
305
+
306
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
307
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
308
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
309
+ # live optimizer object, so we are checking that the numbers are within the right range
310
+ align_to = 2 * world_size
311
+
312
+ def zero2_align(x):
313
+ return align_to * math.ceil(x / align_to)
314
+
315
+ if debug:
316
+ print(f"original offset={offset}, avail_numel={avail_numel}")
317
+
318
+ offset = zero2_align(offset)
319
+ avail_numel = zero2_align(avail_numel)
320
+
321
+ if debug:
322
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
323
+
324
+ # Sanity check
325
+ if offset != avail_numel:
326
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
327
+
328
+ print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
329
+
330
+
331
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
332
+ exclude_frozen_parameters):
333
+ state_dict = OrderedDict()
334
+
335
+ # buffers
336
+ buffers = zero_model_states[0].buffers
337
+ state_dict.update(buffers)
338
+ if debug:
339
+ print(f"added {len(buffers)} buffers")
340
+
341
+ if not exclude_frozen_parameters:
342
+ _zero2_merge_frozen_params(state_dict, zero_model_states)
343
+
344
+ _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
345
+
346
+ # recover shared parameters
347
+ for pair in zero_model_states[0].shared_params:
348
+ if pair[1] in state_dict:
349
+ state_dict[pair[0]] = state_dict[pair[1]]
350
+
351
+ return state_dict
352
+
353
+
354
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
355
+ remainder = unpartitioned_numel % world_size
356
+ padding_numel = (world_size - remainder) if remainder else 0
357
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
358
+ return partitioned_numel, padding_numel
359
+
360
+
361
+ def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
362
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
363
+ return
364
+
365
+ if debug:
366
+ for i in range(world_size):
367
+ num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
368
+ print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
369
+
370
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
371
+ wanted_params = len(frozen_param_shapes)
372
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
373
+ avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
374
+ print(f'Frozen params: Have {avail_numel} numels to process.')
375
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
376
+
377
+ total_params = 0
378
+ total_numel = 0
379
+ for name, shape in zero_model_states[0].frozen_param_shapes.items():
380
+ total_params += 1
381
+ unpartitioned_numel = shape.numel()
382
+ total_numel += unpartitioned_numel
383
+
384
+ param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
385
+ state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
386
+
387
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
388
+
389
+ if debug:
390
+ print(
391
+ f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
392
+ )
393
+
394
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
395
+
396
+
397
+ def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
398
+ param_shapes = zero_model_states[0].param_shapes
399
+ avail_numel = fp32_flat_groups[0].numel() * world_size
400
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
401
+ # param, re-consolidating each param, while dealing with padding if any
402
+
403
+ # merge list of dicts, preserving order
404
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
405
+
406
+ if debug:
407
+ for i in range(world_size):
408
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
409
+
410
+ wanted_params = len(param_shapes)
411
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
412
+ # not asserting if there is a mismatch due to possible padding
413
+ avail_numel = fp32_flat_groups[0].numel() * world_size
414
+ print(f"Trainable params: Have {avail_numel} numels to process.")
415
+ print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
416
+
417
+ # params
418
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
419
+ # out-of-core computing solution
420
+ offset = 0
421
+ total_numel = 0
422
+ total_params = 0
423
+ for name, shape in param_shapes.items():
424
+
425
+ unpartitioned_numel = shape.numel()
426
+ total_numel += unpartitioned_numel
427
+ total_params += 1
428
+
429
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
430
+
431
+ if debug:
432
+ print(
433
+ f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
434
+ )
435
+
436
+ # XXX: memory usage doubles here
437
+ state_dict[name] = torch.cat(
438
+ tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
439
+ 0).narrow(0, 0, unpartitioned_numel).view(shape)
440
+ offset += partitioned_numel
441
+
442
+ offset *= world_size
443
+
444
+ # Sanity check
445
+ if offset != avail_numel:
446
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
447
+
448
+ print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
449
+
450
+
451
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
452
+ exclude_frozen_parameters):
453
+ state_dict = OrderedDict()
454
+
455
+ # buffers
456
+ buffers = zero_model_states[0].buffers
457
+ state_dict.update(buffers)
458
+ if debug:
459
+ print(f"added {len(buffers)} buffers")
460
+
461
+ if not exclude_frozen_parameters:
462
+ _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
463
+
464
+ _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
465
+
466
+ # recover shared parameters
467
+ for pair in zero_model_states[0].shared_params:
468
+ if pair[1] in state_dict:
469
+ state_dict[pair[0]] = state_dict[pair[1]]
470
+
471
+ return state_dict
472
+
473
+
474
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
475
+ """
476
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
477
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
478
+ via a model hub.
479
+
480
+ Args:
481
+ - ``checkpoint_dir``: path to the desired checkpoint folder
482
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
483
+ - ``exclude_frozen_parameters``: exclude frozen parameters
484
+
485
+ Returns:
486
+ - pytorch ``state_dict``
487
+
488
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
489
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
490
+ the checkpoint.
491
+
492
+ A typical usage might be ::
493
+
494
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
495
+ # do the training and checkpoint saving
496
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
497
+ model = model.cpu() # move to cpu
498
+ model.load_state_dict(state_dict)
499
+ # submit to model hub or save the model to share with others
500
+
501
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
502
+ application. i.e. you will need to re-initialize the deepspeed engine, since
503
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
504
+
505
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
506
+
507
+ """
508
+ if tag is None:
509
+ latest_path = os.path.join(checkpoint_dir, 'latest')
510
+ if os.path.isfile(latest_path):
511
+ with open(latest_path, 'r') as fd:
512
+ tag = fd.read().strip()
513
+ else:
514
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
515
+
516
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
517
+
518
+ if not os.path.isdir(ds_checkpoint_dir):
519
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
520
+
521
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
522
+
523
+
524
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
525
+ """
526
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
527
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
528
+
529
+ Args:
530
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
531
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
532
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
533
+ - ``exclude_frozen_parameters``: exclude frozen parameters
534
+ """
535
+
536
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
537
+ print(f"Saving fp32 state dict to {output_file}")
538
+ torch.save(state_dict, output_file)
539
+
540
+
541
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
542
+ """
543
+ 1. Put the provided model to cpu
544
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
545
+ 3. Load it into the provided model
546
+
547
+ Args:
548
+ - ``model``: the model object to update
549
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
550
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
551
+
552
+ Returns:
553
+ - ``model`: modified model
554
+
555
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
556
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
557
+ conveniently placed for you in the checkpoint folder.
558
+
559
+ A typical usage might be ::
560
+
561
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
562
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
563
+ # submit to model hub or save the model to share with others
564
+
565
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
566
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
567
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
568
+
569
+ """
570
+ logger.info(f"Extracting fp32 weights")
571
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
572
+
573
+ logger.info(f"Overwriting model with fp32 weights")
574
+ model = model.cpu()
575
+ model.load_state_dict(state_dict, strict=False)
576
+
577
+ return model
578
+
579
+
580
+ if __name__ == "__main__":
581
+
582
+ parser = argparse.ArgumentParser()
583
+ parser.add_argument("checkpoint_dir",
584
+ type=str,
585
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
586
+ parser.add_argument(
587
+ "output_file",
588
+ type=str,
589
+ help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
590
+ parser.add_argument("-t",
591
+ "--tag",
592
+ type=str,
593
+ default=None,
594
+ help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
595
+ parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
596
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
597
+ args = parser.parse_args()
598
+
599
+ debug = args.debug
600
+
601
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
602
+ args.output_file,
603
+ tag=args.tag,
604
+ exclude_frozen_parameters=args.exclude_frozen_parameters)