Text Generation
Transformers
Safetensors
jetmoe
alignment-handbook
generated_from_trainer
conversational
Inference Endpoints
zguo0525 commited on
Commit
501a9da
1 Parent(s): c31d099

Upload JetMoEForCausalLM

Browse files
README.md ADDED
@@ -0,0 +1,199 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags: []
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+ This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
config.json ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_attn_implementation_internal": "flash_attention_2",
3
+ "_commit_hash": null,
4
+ "_name_or_path": "jetmoe/jetmoe-8b-sft",
5
+ "activation_function": "silu",
6
+ "add_cross_attention": false,
7
+ "architectures": [
8
+ "JetMoEForCausalLM"
9
+ ],
10
+ "auto_map": {
11
+ "AutoConfig": "configuration_jetmoe.JetMoEConfig",
12
+ "AutoModelForCausalLM": "modeling_jetmoe.JetMoEForCausalLM"
13
+ },
14
+ "aux_loss_coef": 0.01,
15
+ "bad_words_ids": null,
16
+ "begin_suppress_tokens": null,
17
+ "bias": true,
18
+ "bos_token_id": 1,
19
+ "chunk_size_feed_forward": 0,
20
+ "cross_attention_hidden_size": null,
21
+ "decoder_start_token_id": null,
22
+ "diversity_penalty": 0.0,
23
+ "do_sample": false,
24
+ "early_stopping": false,
25
+ "encoder_no_repeat_ngram_size": 0,
26
+ "eos_token_id": 2,
27
+ "exponential_decay_length_penalty": null,
28
+ "ffn_hidden_size": 5632,
29
+ "finetuning_task": null,
30
+ "forced_bos_token_id": null,
31
+ "forced_eos_token_id": null,
32
+ "glu": true,
33
+ "id2label": {
34
+ "0": "LABEL_0",
35
+ "1": "LABEL_1"
36
+ },
37
+ "initializer_range": 0.01,
38
+ "is_decoder": false,
39
+ "is_encoder_decoder": false,
40
+ "kv_channels": 128,
41
+ "label2id": {
42
+ "LABEL_0": 0,
43
+ "LABEL_1": 1
44
+ },
45
+ "layer_norm_epsilon": 1e-05,
46
+ "length_penalty": 1.0,
47
+ "max_length": 20,
48
+ "min_length": 0,
49
+ "model_type": "jetmoe",
50
+ "moe_num_experts": 8,
51
+ "moe_top_k": 2,
52
+ "n_embd": 2048,
53
+ "n_head": 16,
54
+ "n_layer": 24,
55
+ "n_positions": 4096,
56
+ "no_repeat_ngram_size": 0,
57
+ "num_beam_groups": 1,
58
+ "num_beams": 1,
59
+ "num_key_value_heads": 8,
60
+ "num_layers": 24,
61
+ "num_return_sequences": 1,
62
+ "output_attentions": false,
63
+ "output_hidden_states": false,
64
+ "output_scores": false,
65
+ "pad_token_id": null,
66
+ "prefix": null,
67
+ "problem_type": null,
68
+ "pruned_heads": {},
69
+ "remove_invalid_values": false,
70
+ "repetition_penalty": 1.0,
71
+ "return_dict": true,
72
+ "return_dict_in_generate": false,
73
+ "rms_norm_eps": 1e-05,
74
+ "rope_theta": 10000.0,
75
+ "rotary_percent": 1.0,
76
+ "sep_token_id": null,
77
+ "suppress_tokens": null,
78
+ "task_specific_params": null,
79
+ "temperature": 1.0,
80
+ "tf_legacy_loss": false,
81
+ "tie_encoder_decoder": false,
82
+ "tie_word_embeddings": true,
83
+ "tokenizer_class": null,
84
+ "top_k": 50,
85
+ "top_p": 1.0,
86
+ "torchscript": false,
87
+ "transformers_version": null,
88
+ "typical_p": 1.0,
89
+ "use_bfloat16": false,
90
+ "use_cache": true,
91
+ "vocab_size": 32000
92
+ }
configuration_jetmoe.py ADDED
@@ -0,0 +1,269 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ JetMoE model configuration"""
2
+ from collections import OrderedDict
3
+ from typing import Any, List, Mapping, Optional
4
+
5
+ from transformers import PreTrainedTokenizer, TensorType, is_torch_available
6
+ from transformers.configuration_utils import PretrainedConfig
7
+ from transformers.onnx import OnnxConfigWithPast, PatchingSpec
8
+ from transformers.utils import logging
9
+ import torch.nn.init as init
10
+ import json
11
+
12
+ logger = logging.get_logger(__name__)
13
+
14
+
15
+ class JetMoEConfig(PretrainedConfig):
16
+ r"""
17
+ This is the configuration class to store the configuration of a [`JetMoEModel`]. It is used to instantiate a
18
+ JetMoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
19
+ with the defaults will yield a similar configuration to that of the JetMoE
20
+ [jetmoe-small](https://huggingface.co/jetmoe-small) architecture. Configuration objects
21
+ inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from
22
+ [`PretrainedConfig`] for more information.
23
+
24
+ Args:
25
+ vocab_size (`int`, *optional*, defaults to 50400):
26
+ Vocabulary size of the JetMoE model. Defines the number of different tokens that can be represented by the
27
+ `inputs_ids` passed when calling [`JetMoEModel`].
28
+ n_positions (`int`, *optional*, defaults to 2048):
29
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
30
+ just in case (e.g., 512 or 1024 or 2048).
31
+ n_embd (`int`, *optional*, defaults to 4096):
32
+ Dimensionality of the embeddings and hidden states.
33
+ n_layer (`int`, *optional*, defaults to 28):
34
+ Number of hidden layers in the Transformer encoder.
35
+ n_head (`int`, *optional*, defaults to 16):
36
+ Number of attention heads for each attention layer in the Transformer encoder.
37
+ rotary_dim (`int`, *optional*, defaults to 64):
38
+ Number of dimensions in the embedding that Rotary Position Embedding is applied to.
39
+ n_inner (`int`, *optional*, defaults to None):
40
+ Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
41
+ activation_function (`str`, *optional*, defaults to `"gelu_new"`):
42
+ Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
43
+ resid_pdrop (`float`, *optional*, defaults to 0.1):
44
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
45
+ embd_pdrop (`int`, *optional*, defaults to 0.1):
46
+ The dropout ratio for the embeddings.
47
+ attn_pdrop (`float`, *optional*, defaults to 0.1):
48
+ The dropout ratio for the attention.
49
+ layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
50
+ The epsilon to use in the layer normalization layers.
51
+ initializer_range (`float`, *optional*, defaults to 0.02):
52
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
53
+ use_cache (`bool`, *optional*, defaults to `True`):
54
+ Whether or not the model should return the last key/values attentions (not used by all models).
55
+
56
+ Example:
57
+
58
+ ```python
59
+ >>> from transformers import JetMoEConfig, JetMoEModel
60
+
61
+ >>> # Initializing a JetMoE 6B configuration
62
+ >>> configuration = JetMoEConfig()
63
+
64
+ >>> # Initializing a model (with random weights) from the configuration
65
+ >>> model = JetMoEModel(configuration)
66
+
67
+ >>> # Accessing the model configuration
68
+ >>> configuration = model.config
69
+ ```"""
70
+ model_type = "jetmoe"
71
+ attribute_map = {
72
+ "max_position_embeddings": "n_positions",
73
+ "hidden_size": "n_embd",
74
+ "num_attention_heads": "n_head",
75
+ "num_hidden_layers": "num_layers",
76
+ }
77
+
78
+ def __init__(
79
+ self,
80
+ vocab_size=50295,
81
+ hidden_size=1024,
82
+ num_layers=24,
83
+ num_attention_heads=16,
84
+ kv_channels = 128,
85
+ ffn_hidden_size=2048,
86
+ max_position_embeddings=4096,
87
+ rotary_percent=1.0,
88
+ activation_function="silu",
89
+ glu=True,
90
+ moe_num_experts=8,
91
+ moe_top_k=2,
92
+ use_cache=True,
93
+ bos_token_id=1,
94
+ eos_token_id=2,
95
+ tie_word_embeddings=True,
96
+ bias=True,
97
+ rope_theta=10000.0,
98
+ rms_norm_eps=1e-6,
99
+ initializer_range=0.01,
100
+ **kwargs,
101
+ ):
102
+ self.vocab_size = vocab_size
103
+ self.hidden_size = hidden_size
104
+ self.num_layers = num_layers
105
+ self.num_attention_heads = num_attention_heads
106
+ self.kv_channels = kv_channels
107
+ self.ffn_hidden_size = ffn_hidden_size
108
+ self.max_position_embeddings = max_position_embeddings
109
+ self.rotary_percent = rotary_percent
110
+ self.activation_function = activation_function
111
+ self.glu = glu
112
+ self.moe_num_experts = moe_num_experts
113
+ self.moe_top_k = moe_top_k
114
+ self.use_cache = use_cache
115
+ self.initializer_range = initializer_range
116
+
117
+ self.bos_token_id = bos_token_id
118
+ self.eos_token_id = eos_token_id
119
+
120
+ self.init_method = init.xavier_uniform_
121
+ self.output_layer_init_method = init.xavier_uniform_
122
+ self.bias = bias
123
+ self.rope_theta = rope_theta
124
+ self.rms_norm_eps = rms_norm_eps
125
+
126
+ super().__init__(
127
+ bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
128
+ )
129
+
130
+ def to_dict(self):
131
+ """Returns a dictionary representation of the config, excluding non-serializable attributes."""
132
+ return {k: v for k, v in self.__dict__.items() if k not in ['init_method', 'output_layer_init_method', 'torch_dtype', '_pre_quantization_dtype', 'quantization_config']}
133
+
134
+ def to_json_string(self, use_diff=False):
135
+ """Serializes this instance to a JSON string, excluding non-serializable attributes.
136
+
137
+ Args:
138
+ use_diff (bool): Whether to use differences with the default config. This argument is
139
+ accepted for compatibility with the transformers library but is not
140
+ used in this custom implementation.
141
+ """
142
+ config_dict = self.to_dict() # Assuming you have a to_dict method as shown earlier
143
+ return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
144
+
145
+ class JetMoEOnnxConfig(OnnxConfigWithPast):
146
+ def __init__(
147
+ self,
148
+ config: PretrainedConfig,
149
+ task: str = "default",
150
+ patching_specs: List[PatchingSpec] = None,
151
+ use_past: bool = False,
152
+ ):
153
+ """
154
+ Initialize the JetMoEOnnxConfig.
155
+
156
+ Args:
157
+ config (PretrainedConfig): Pretrained model configuration.
158
+ task (str): Task description.
159
+ patching_specs (List[PatchingSpec]): List of patching specifications.
160
+ use_past (bool): Whether to use past tokens in the configuration.
161
+ """
162
+ super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)
163
+ if not getattr(self._config, "pad_token_id", None):
164
+ # TODO: how to do that better?
165
+ self._config.pad_token_id = 0
166
+
167
+ @property
168
+ def inputs(self) -> Mapping[str, Mapping[int, str]]:
169
+ """
170
+ Define the input mappings.
171
+
172
+ Returns:
173
+ Mapping[str, Mapping[int, str]]: Input mappings.
174
+ """
175
+ common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})
176
+ if self.use_past:
177
+ self.fill_with_past_key_values_(common_inputs, direction="inputs")
178
+ common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"}
179
+ else:
180
+ common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
181
+
182
+ return common_inputs
183
+
184
+ @property
185
+ def num_layers(self) -> int:
186
+ """
187
+ Get the number of layers.
188
+
189
+ Returns:
190
+ int: Number of layers.
191
+ """
192
+ return self._config.n_layer
193
+
194
+ @property
195
+ def num_attention_heads(self) -> int:
196
+ """
197
+ Get the number of attention heads.
198
+
199
+ Returns:
200
+ int: Number of attention heads.
201
+ """
202
+ return self._config.n_head
203
+
204
+ def generate_dummy_inputs(
205
+ self,
206
+ tokenizer: PreTrainedTokenizer,
207
+ batch_size: int = -1,
208
+ seq_length: int = -1,
209
+ is_pair: bool = False,
210
+ framework: Optional[TensorType] = None,
211
+ ) -> Mapping[str, Any]:
212
+ """
213
+ Generate dummy inputs for testing.
214
+
215
+ Args:
216
+ tokenizer (PreTrainedTokenizer): Pretrained tokenizer.
217
+ batch_size (int): Batch size.
218
+ seq_length (int): Sequence length.
219
+ is_pair (bool): Whether the input is a pair.
220
+ framework (Optional[TensorType]): Tensor framework.
221
+
222
+ Returns:
223
+ Mapping[str, Any]: Dummy inputs.
224
+ """
225
+ common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(
226
+ tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
227
+ )
228
+
229
+ # We need to order the input in the way they appears in the forward()
230
+ ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
231
+
232
+ # Need to add the past_keys
233
+ if self.use_past:
234
+ if not is_torch_available():
235
+ raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
236
+ else:
237
+ import torch
238
+
239
+ batch, seqlen = common_inputs["input_ids"].shape
240
+ # Not using the same length for past_key_values
241
+ past_key_values_length = seqlen + 2
242
+ past_shape = (
243
+ batch,
244
+ self.num_attention_heads,
245
+ past_key_values_length,
246
+ self._config.hidden_size // self.num_attention_heads,
247
+ )
248
+ ordered_inputs["past_key_values"] = [
249
+ (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)
250
+ ]
251
+
252
+ ordered_inputs["attention_mask"] = common_inputs["attention_mask"]
253
+ if self.use_past:
254
+ mask_dtype = ordered_inputs["attention_mask"].dtype
255
+ ordered_inputs["attention_mask"] = torch.cat(
256
+ [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
257
+ )
258
+
259
+ return ordered_inputs
260
+
261
+ @property
262
+ def default_onnx_opset(self) -> int:
263
+ """
264
+ Get the default ONNX opset version.
265
+
266
+ Returns:
267
+ int: Default ONNX opset version.
268
+ """
269
+ return 13
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 2,
6
+ "transformers_version": "4.39.0"
7
+ }
model-00001-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11c177c7be6b45d895ba27dd8340858ed8a4ef01f449fdf7e2adbe80d472b71a
3
+ size 4879573992
model-00002-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:afd54f0385bdc42259b54e8c950a9cb5fda9481f99f8aeb0e46cb1374b49e654
3
+ size 4933084288
model-00003-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4a9b7138c7919aa89d997a7b6887e34ca23fc76b3023ce60d029b77e363c77b
3
+ size 4933084344
model-00004-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5c68a3bf7e68e8e9a8035f6e06e8cc0e81b72e1252a675213c740b6d50c0b143
3
+ size 2298765576
model.safetensors.index.json ADDED
@@ -0,0 +1,273 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 17044475904
4
+ },
5
+ "weight_map": {
6
+ "model.embed_tokens.weight": "model-00001-of-00004.safetensors",
7
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
8
+ "model.layers.0.mlp.bias": "model-00001-of-00004.safetensors",
9
+ "model.layers.0.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
10
+ "model.layers.0.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
11
+ "model.layers.0.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
12
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
13
+ "model.layers.0.self_attention.experts.bias": "model-00001-of-00004.safetensors",
14
+ "model.layers.0.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
15
+ "model.layers.0.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
16
+ "model.layers.0.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
17
+ "model.layers.0.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
18
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
19
+ "model.layers.1.mlp.bias": "model-00001-of-00004.safetensors",
20
+ "model.layers.1.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
21
+ "model.layers.1.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
22
+ "model.layers.1.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
23
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
24
+ "model.layers.1.self_attention.experts.bias": "model-00001-of-00004.safetensors",
25
+ "model.layers.1.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
26
+ "model.layers.1.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
27
+ "model.layers.1.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
28
+ "model.layers.1.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
29
+ "model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
30
+ "model.layers.10.mlp.bias": "model-00002-of-00004.safetensors",
31
+ "model.layers.10.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
32
+ "model.layers.10.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
33
+ "model.layers.10.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
34
+ "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
35
+ "model.layers.10.self_attention.experts.bias": "model-00002-of-00004.safetensors",
36
+ "model.layers.10.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
37
+ "model.layers.10.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
38
+ "model.layers.10.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
39
+ "model.layers.10.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
40
+ "model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
41
+ "model.layers.11.mlp.bias": "model-00002-of-00004.safetensors",
42
+ "model.layers.11.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
43
+ "model.layers.11.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
44
+ "model.layers.11.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
45
+ "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
46
+ "model.layers.11.self_attention.experts.bias": "model-00002-of-00004.safetensors",
47
+ "model.layers.11.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
48
+ "model.layers.11.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
49
+ "model.layers.11.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
50
+ "model.layers.11.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
51
+ "model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
52
+ "model.layers.12.mlp.bias": "model-00002-of-00004.safetensors",
53
+ "model.layers.12.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
54
+ "model.layers.12.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
55
+ "model.layers.12.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
56
+ "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
57
+ "model.layers.12.self_attention.experts.bias": "model-00002-of-00004.safetensors",
58
+ "model.layers.12.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
59
+ "model.layers.12.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
60
+ "model.layers.12.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
61
+ "model.layers.12.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
62
+ "model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
63
+ "model.layers.13.mlp.bias": "model-00002-of-00004.safetensors",
64
+ "model.layers.13.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
65
+ "model.layers.13.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
66
+ "model.layers.13.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
67
+ "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
68
+ "model.layers.13.self_attention.experts.bias": "model-00002-of-00004.safetensors",
69
+ "model.layers.13.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
70
+ "model.layers.13.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
71
+ "model.layers.13.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
72
+ "model.layers.13.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
73
+ "model.layers.14.input_layernorm.weight": "model-00003-of-00004.safetensors",
74
+ "model.layers.14.mlp.bias": "model-00003-of-00004.safetensors",
75
+ "model.layers.14.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
76
+ "model.layers.14.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
77
+ "model.layers.14.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
78
+ "model.layers.14.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
79
+ "model.layers.14.self_attention.experts.bias": "model-00003-of-00004.safetensors",
80
+ "model.layers.14.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
81
+ "model.layers.14.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
82
+ "model.layers.14.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
83
+ "model.layers.14.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
84
+ "model.layers.15.input_layernorm.weight": "model-00003-of-00004.safetensors",
85
+ "model.layers.15.mlp.bias": "model-00003-of-00004.safetensors",
86
+ "model.layers.15.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
87
+ "model.layers.15.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
88
+ "model.layers.15.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
89
+ "model.layers.15.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
90
+ "model.layers.15.self_attention.experts.bias": "model-00003-of-00004.safetensors",
91
+ "model.layers.15.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
92
+ "model.layers.15.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
93
+ "model.layers.15.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
94
+ "model.layers.15.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
95
+ "model.layers.16.input_layernorm.weight": "model-00003-of-00004.safetensors",
96
+ "model.layers.16.mlp.bias": "model-00003-of-00004.safetensors",
97
+ "model.layers.16.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
98
+ "model.layers.16.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
99
+ "model.layers.16.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
100
+ "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
101
+ "model.layers.16.self_attention.experts.bias": "model-00003-of-00004.safetensors",
102
+ "model.layers.16.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
103
+ "model.layers.16.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
104
+ "model.layers.16.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
105
+ "model.layers.16.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
106
+ "model.layers.17.input_layernorm.weight": "model-00003-of-00004.safetensors",
107
+ "model.layers.17.mlp.bias": "model-00003-of-00004.safetensors",
108
+ "model.layers.17.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
109
+ "model.layers.17.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
110
+ "model.layers.17.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
111
+ "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
112
+ "model.layers.17.self_attention.experts.bias": "model-00003-of-00004.safetensors",
113
+ "model.layers.17.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
114
+ "model.layers.17.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
115
+ "model.layers.17.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
116
+ "model.layers.17.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
117
+ "model.layers.18.input_layernorm.weight": "model-00003-of-00004.safetensors",
118
+ "model.layers.18.mlp.bias": "model-00003-of-00004.safetensors",
119
+ "model.layers.18.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
120
+ "model.layers.18.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
121
+ "model.layers.18.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
122
+ "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
123
+ "model.layers.18.self_attention.experts.bias": "model-00003-of-00004.safetensors",
124
+ "model.layers.18.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
125
+ "model.layers.18.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
126
+ "model.layers.18.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
127
+ "model.layers.18.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
128
+ "model.layers.19.input_layernorm.weight": "model-00003-of-00004.safetensors",
129
+ "model.layers.19.mlp.bias": "model-00003-of-00004.safetensors",
130
+ "model.layers.19.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
131
+ "model.layers.19.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
132
+ "model.layers.19.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
133
+ "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
134
+ "model.layers.19.self_attention.experts.bias": "model-00003-of-00004.safetensors",
135
+ "model.layers.19.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
136
+ "model.layers.19.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
137
+ "model.layers.19.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
138
+ "model.layers.19.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
139
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
140
+ "model.layers.2.mlp.bias": "model-00001-of-00004.safetensors",
141
+ "model.layers.2.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
142
+ "model.layers.2.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
143
+ "model.layers.2.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
144
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
145
+ "model.layers.2.self_attention.experts.bias": "model-00001-of-00004.safetensors",
146
+ "model.layers.2.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
147
+ "model.layers.2.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
148
+ "model.layers.2.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
149
+ "model.layers.2.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
150
+ "model.layers.20.input_layernorm.weight": "model-00003-of-00004.safetensors",
151
+ "model.layers.20.mlp.bias": "model-00003-of-00004.safetensors",
152
+ "model.layers.20.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
153
+ "model.layers.20.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
154
+ "model.layers.20.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
155
+ "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
156
+ "model.layers.20.self_attention.experts.bias": "model-00003-of-00004.safetensors",
157
+ "model.layers.20.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
158
+ "model.layers.20.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
159
+ "model.layers.20.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
160
+ "model.layers.20.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
161
+ "model.layers.21.input_layernorm.weight": "model-00004-of-00004.safetensors",
162
+ "model.layers.21.mlp.bias": "model-00004-of-00004.safetensors",
163
+ "model.layers.21.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
164
+ "model.layers.21.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
165
+ "model.layers.21.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
166
+ "model.layers.21.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
167
+ "model.layers.21.self_attention.experts.bias": "model-00004-of-00004.safetensors",
168
+ "model.layers.21.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
169
+ "model.layers.21.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
170
+ "model.layers.21.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
171
+ "model.layers.21.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
172
+ "model.layers.22.input_layernorm.weight": "model-00004-of-00004.safetensors",
173
+ "model.layers.22.mlp.bias": "model-00004-of-00004.safetensors",
174
+ "model.layers.22.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
175
+ "model.layers.22.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
176
+ "model.layers.22.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
177
+ "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
178
+ "model.layers.22.self_attention.experts.bias": "model-00004-of-00004.safetensors",
179
+ "model.layers.22.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
180
+ "model.layers.22.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
181
+ "model.layers.22.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
182
+ "model.layers.22.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
183
+ "model.layers.23.input_layernorm.weight": "model-00004-of-00004.safetensors",
184
+ "model.layers.23.mlp.bias": "model-00004-of-00004.safetensors",
185
+ "model.layers.23.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
186
+ "model.layers.23.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
187
+ "model.layers.23.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
188
+ "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
189
+ "model.layers.23.self_attention.experts.bias": "model-00004-of-00004.safetensors",
190
+ "model.layers.23.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
191
+ "model.layers.23.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
192
+ "model.layers.23.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
193
+ "model.layers.23.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
194
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
195
+ "model.layers.3.mlp.bias": "model-00001-of-00004.safetensors",
196
+ "model.layers.3.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
197
+ "model.layers.3.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
198
+ "model.layers.3.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
199
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
200
+ "model.layers.3.self_attention.experts.bias": "model-00001-of-00004.safetensors",
201
+ "model.layers.3.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
202
+ "model.layers.3.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
203
+ "model.layers.3.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
204
+ "model.layers.3.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
205
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
206
+ "model.layers.4.mlp.bias": "model-00001-of-00004.safetensors",
207
+ "model.layers.4.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
208
+ "model.layers.4.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
209
+ "model.layers.4.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
210
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
211
+ "model.layers.4.self_attention.experts.bias": "model-00001-of-00004.safetensors",
212
+ "model.layers.4.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
213
+ "model.layers.4.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
214
+ "model.layers.4.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
215
+ "model.layers.4.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
216
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00004.safetensors",
217
+ "model.layers.5.mlp.bias": "model-00001-of-00004.safetensors",
218
+ "model.layers.5.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
219
+ "model.layers.5.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
220
+ "model.layers.5.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
221
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
222
+ "model.layers.5.self_attention.experts.bias": "model-00001-of-00004.safetensors",
223
+ "model.layers.5.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
224
+ "model.layers.5.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
225
+ "model.layers.5.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
226
+ "model.layers.5.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
227
+ "model.layers.6.input_layernorm.weight": "model-00001-of-00004.safetensors",
228
+ "model.layers.6.mlp.bias": "model-00001-of-00004.safetensors",
229
+ "model.layers.6.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
230
+ "model.layers.6.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
231
+ "model.layers.6.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
232
+ "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
233
+ "model.layers.6.self_attention.experts.bias": "model-00001-of-00004.safetensors",
234
+ "model.layers.6.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
235
+ "model.layers.6.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
236
+ "model.layers.6.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
237
+ "model.layers.6.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
238
+ "model.layers.7.input_layernorm.weight": "model-00002-of-00004.safetensors",
239
+ "model.layers.7.mlp.bias": "model-00002-of-00004.safetensors",
240
+ "model.layers.7.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
241
+ "model.layers.7.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
242
+ "model.layers.7.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
243
+ "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
244
+ "model.layers.7.self_attention.experts.bias": "model-00002-of-00004.safetensors",
245
+ "model.layers.7.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
246
+ "model.layers.7.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
247
+ "model.layers.7.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
248
+ "model.layers.7.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
249
+ "model.layers.8.input_layernorm.weight": "model-00002-of-00004.safetensors",
250
+ "model.layers.8.mlp.bias": "model-00002-of-00004.safetensors",
251
+ "model.layers.8.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
252
+ "model.layers.8.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
253
+ "model.layers.8.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
254
+ "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
255
+ "model.layers.8.self_attention.experts.bias": "model-00002-of-00004.safetensors",
256
+ "model.layers.8.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
257
+ "model.layers.8.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
258
+ "model.layers.8.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
259
+ "model.layers.8.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
260
+ "model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
261
+ "model.layers.9.mlp.bias": "model-00002-of-00004.safetensors",
262
+ "model.layers.9.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
263
+ "model.layers.9.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
264
+ "model.layers.9.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
265
+ "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
266
+ "model.layers.9.self_attention.experts.bias": "model-00002-of-00004.safetensors",
267
+ "model.layers.9.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
268
+ "model.layers.9.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
269
+ "model.layers.9.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
270
+ "model.layers.9.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
271
+ "model.norm.weight": "model-00004-of-00004.safetensors"
272
+ }
273
+ }
modeling_jetmoe.py ADDED
@@ -0,0 +1,1399 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ PyTorch JetMoE model."""
2
+
3
+ from typing import List, Optional, Tuple, Union
4
+ import warnings, math
5
+
6
+ import torch
7
+ import torch.utils.checkpoint
8
+ from torch import nn
9
+ from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
10
+ from torch.nn import functional as F
11
+
12
+ import megablocks
13
+ from transformers.modeling_outputs import (
14
+ BaseModelOutputWithPast,
15
+ CausalLMOutputWithPast,
16
+ SequenceClassifierOutputWithPast,
17
+ dataclass
18
+ )
19
+ from transformers.modeling_utils import PreTrainedModel
20
+ from transformers.utils import (
21
+ add_start_docstrings,
22
+ add_start_docstrings_to_model_forward,
23
+ is_flash_attn_2_available,
24
+ is_flash_attn_greater_or_equal_2_10,
25
+ replace_return_docstrings,
26
+ logging
27
+ )
28
+ from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
29
+ from transformers.cache_utils import Cache, DynamicCache
30
+ from .configuration_jetmoe import JetMoEConfig
31
+ from jetmoe_model.utils import moe
32
+
33
+ if is_flash_attn_2_available():
34
+ from flash_attn import flash_attn_func, flash_attn_varlen_func
35
+ from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
36
+
37
+ logger = logging.get_logger(__name__)
38
+
39
+ _CHECKPOINT_FOR_DOC = "jetmoe"
40
+ _CONFIG_FOR_DOC = "JetMoEConfig"
41
+
42
+
43
+ @dataclass
44
+ class JetMoEBaseModelOutputWithPast(BaseModelOutputWithPast):
45
+ """
46
+ Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
47
+
48
+ Args:
49
+ last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
50
+ Sequence of hidden-states at the output of the last layer of the model.
51
+
52
+ If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
53
+ hidden_size)` is output.
54
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
55
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
56
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
57
+ `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
58
+ encoder_sequence_length, embed_size_per_head)`.
59
+
60
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
61
+ `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
62
+ input) to speed up sequential decoding.
63
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
64
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
65
+ one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
66
+
67
+ Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
68
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
69
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
70
+ sequence_length)`.
71
+
72
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
73
+ heads.
74
+ """
75
+
76
+ last_hidden_state: torch.FloatTensor = None
77
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
78
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
79
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
80
+ aux_loss: Optional[torch.FloatTensor] = None
81
+
82
+
83
+ @dataclass
84
+ class JetMoECausalLMOutputWithPast(CausalLMOutputWithPast):
85
+ """
86
+ Base class for causal language model (or autoregressive) outputs.
87
+
88
+ Args:
89
+ loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
90
+ Language modeling loss (for next-token prediction).
91
+ logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
92
+ Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
93
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
94
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
95
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
96
+
97
+ Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
98
+ `past_key_values` input) to speed up sequential decoding.
99
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
100
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
101
+ one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
102
+
103
+ Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
104
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
105
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
106
+ sequence_length)`.
107
+
108
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
109
+ heads.
110
+ """
111
+
112
+ loss: Optional[torch.FloatTensor] = None
113
+ logits: torch.FloatTensor = None
114
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
115
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
116
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
117
+ aux_loss: Optional[torch.FloatTensor] = None
118
+
119
+
120
+ @dataclass
121
+ class JetMoESequenceClassifierOutputWithPast(SequenceClassifierOutputWithPast):
122
+ """
123
+ Base class for outputs of sentence classification models.
124
+
125
+ Args:
126
+ loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
127
+ Classification (or regression if config.num_labels==1) loss.
128
+ logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
129
+ Classification (or regression if config.num_labels==1) scores (before SoftMax).
130
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
131
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
132
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
133
+
134
+ Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
135
+ `past_key_values` input) to speed up sequential decoding.
136
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
137
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
138
+ one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
139
+
140
+ Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
141
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
142
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
143
+ sequence_length)`.
144
+
145
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
146
+ heads.
147
+ """
148
+
149
+ loss: Optional[torch.FloatTensor] = None
150
+ logits: torch.FloatTensor = None
151
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
152
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
153
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
154
+ aux_loss: Optional[torch.FloatTensor] = None
155
+
156
+
157
+ # Copied from transformers.models.llama.modeling_llama._get_unpad_data
158
+ def _get_unpad_data(attention_mask):
159
+ seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
160
+ indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
161
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
162
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
163
+ return (
164
+ indices,
165
+ cu_seqlens,
166
+ max_seqlen_in_batch,
167
+ )
168
+
169
+ class JetMoERMSNorm(nn.Module):
170
+ def __init__(self, hidden_size, eps=1e-6):
171
+ """
172
+ JetMoERMSNorm module
173
+ """
174
+ super().__init__()
175
+ self.weight = nn.Parameter(torch.ones(hidden_size))
176
+ self.variance_epsilon = eps
177
+
178
+ def forward(self, hidden_states):
179
+ input_dtype = hidden_states.dtype
180
+ hidden_states = hidden_states.to(torch.float32)
181
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
182
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
183
+ return self.weight * hidden_states.to(input_dtype)
184
+
185
+
186
+ # copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding
187
+ class JetMoERotaryEmbedding(nn.Module):
188
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
189
+ super().__init__()
190
+
191
+ self.dim = dim
192
+ self.max_position_embeddings = max_position_embeddings
193
+ self.base = base
194
+ inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
195
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
196
+
197
+ # Build here to make `torch.jit.trace` work.
198
+ self._set_cos_sin_cache(
199
+ seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
200
+ )
201
+
202
+ def _set_cos_sin_cache(self, seq_len, device, dtype):
203
+ self.max_seq_len_cached = seq_len
204
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
205
+
206
+ freqs = torch.outer(t, self.inv_freq)
207
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
208
+ emb = torch.cat((freqs, freqs), dim=-1)
209
+ self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
210
+ self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
211
+
212
+ def forward(self, x, seq_len=None):
213
+ # x: [bs, num_attention_heads, seq_len, head_size]
214
+ if seq_len > self.max_seq_len_cached:
215
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
216
+
217
+ return (
218
+ self.cos_cached[:seq_len].to(dtype=x.dtype),
219
+ self.sin_cached[:seq_len].to(dtype=x.dtype),
220
+ )
221
+
222
+
223
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
224
+ def rotate_half(x):
225
+ """Rotates half the hidden dims of the input."""
226
+ x1 = x[..., : x.shape[-1] // 2]
227
+ x2 = x[..., x.shape[-1] // 2 :]
228
+ return torch.cat((-x2, x1), dim=-1)
229
+
230
+
231
+ # copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
232
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=2):
233
+ """Applies Rotary Position Embedding to the query and key tensors.
234
+
235
+ Args:
236
+ q (`torch.Tensor`): The query tensor.
237
+ k (`torch.Tensor`): The key tensor.
238
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
239
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
240
+ position_ids (`torch.Tensor`):
241
+ The position indices of the tokens corresponding to the query and key tensors. For example, this can be
242
+ used to pass offsetted position ids when working with a KV-cache.
243
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
244
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
245
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
246
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
247
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
248
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
249
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
250
+ Returns:
251
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
252
+ """
253
+ cos = cos[position_ids].unsqueeze(unsqueeze_dim)
254
+ sin = sin[position_ids].unsqueeze(unsqueeze_dim)
255
+ q_embed = (q * cos) + (rotate_half(q) * sin)
256
+ k_embed = (k * cos) + (rotate_half(k) * sin)
257
+ return q_embed, k_embed
258
+
259
+
260
+ class JetMoEAttention(nn.Module):
261
+ """
262
+ Multi-headed attention from 'Attention Is All You Need' paper.
263
+ """
264
+
265
+ def __init__(self, config: JetMoEConfig, layer_idx: Optional[int] = None):
266
+ """
267
+ Initialize the JetMoEAttention module.
268
+
269
+ Args:
270
+ config: Configuration object with model hyperparameters.
271
+ """
272
+ super().__init__()
273
+ self.config = config
274
+ self.layer_idx = layer_idx
275
+ self.is_causal = True
276
+ if layer_idx is None:
277
+ logger.warning_once(
278
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
279
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
280
+ "when creating this class."
281
+ )
282
+
283
+ self.top_k = config.moe_top_k
284
+
285
+ self.kv_projection_size = config.kv_channels * config.num_attention_heads
286
+ self.num_key_value_heads = config.num_attention_heads
287
+ self.num_heads = self.num_key_value_heads * self.top_k
288
+ self.hidden_size_per_attention_head = config.kv_channels
289
+
290
+ self.experts = moe.MoE(
291
+ input_size=config.hidden_size,
292
+ hidden_size=self.kv_projection_size,
293
+ num_experts=config.moe_num_experts,
294
+ top_k=config.moe_top_k,
295
+ glu=False
296
+ )
297
+
298
+ self.kv_proj = torch.nn.Linear(
299
+ config.hidden_size, self.kv_projection_size * 2, bias=False
300
+ )
301
+
302
+ self.rotary_emb = JetMoERotaryEmbedding(
303
+ config.kv_channels,
304
+ max_position_embeddings=config.max_position_embeddings,
305
+ base=config.rope_theta,
306
+ )
307
+
308
+ # def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
309
+ # return tensor.view(bsz, seq_len, self.num_attention_heads, self.hidden_size_per_attention_head).transpose(1, 2).contiguous()
310
+
311
+ def forward(
312
+ self,
313
+ hidden_states: torch.Tensor,
314
+ attention_mask: Optional[torch.Tensor] = None,
315
+ position_ids: Optional[torch.LongTensor] = None,
316
+ past_key_value: Optional[Cache] = None,
317
+ output_attentions: bool = False,
318
+ use_cache: bool = False,
319
+ **kwargs,
320
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
321
+ if "padding_mask" in kwargs:
322
+ warnings.warn(
323
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
324
+ )
325
+ bsz, q_len, _ = hidden_states.size()
326
+
327
+ query_states, aux_loss = self.experts.map(hidden_states)
328
+ key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)
329
+
330
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).transpose(1, 2)
331
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
332
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
333
+
334
+ kv_seq_len = key_states.shape[2]
335
+ if past_key_value is not None:
336
+ if self.layer_idx is None:
337
+ raise ValueError(
338
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
339
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
340
+ "with a layer index."
341
+ )
342
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
343
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
344
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, unsqueeze_dim=1)
345
+
346
+ if past_key_value is not None:
347
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
348
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
349
+
350
+ # repeat k/v heads if n_kv_heads < n_heads
351
+ key_states = key_states.repeat(1, self.top_k, 1, 1)
352
+ value_states = value_states.repeat(1, self.top_k, 1, 1)
353
+
354
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.hidden_size_per_attention_head)
355
+
356
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
357
+ raise ValueError(
358
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
359
+ f" {attn_weights.size()}"
360
+ )
361
+
362
+ if attention_mask is not None:
363
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
364
+ raise ValueError(
365
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
366
+ )
367
+
368
+ attn_weights = attn_weights + attention_mask
369
+
370
+ # upcast attention to fp32
371
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
372
+ # attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
373
+ attn_output = torch.matmul(attn_weights, value_states)
374
+
375
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.hidden_size_per_attention_head):
376
+ raise ValueError(
377
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.hidden_size_per_attention_head)}, but is"
378
+ f" {attn_output.size()}"
379
+ )
380
+
381
+ attn_output = attn_output.transpose(1, 2).contiguous()
382
+ attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size)
383
+
384
+ attn_output = self.experts.reduce(attn_output)
385
+ attn_output = attn_output.view(bsz, q_len, -1)
386
+
387
+ if not output_attentions:
388
+ attn_weights = None
389
+
390
+ return attn_output, attn_weights, past_key_value, aux_loss
391
+
392
+
393
+ # copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->JetMoE
394
+ class JetMoESdpaAttention(JetMoEAttention):
395
+ """
396
+ JetMoE attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
397
+ `JetMoEAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
398
+ SDPA API.
399
+ """
400
+
401
+ # Adapted from JetMoEAttention.forward
402
+ def forward(
403
+ self,
404
+ hidden_states: torch.Tensor,
405
+ attention_mask: Optional[torch.Tensor] = None,
406
+ position_ids: Optional[torch.LongTensor] = None,
407
+ past_key_value: Optional[Cache] = None,
408
+ output_attentions: bool = False,
409
+ use_cache: bool = False,
410
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
411
+ if output_attentions:
412
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
413
+ logger.warning_once(
414
+ "JetMoEModel is using JetMoESdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
415
+ 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
416
+ )
417
+ return super().forward(
418
+ hidden_states=hidden_states,
419
+ attention_mask=attention_mask,
420
+ position_ids=position_ids,
421
+ past_key_value=past_key_value,
422
+ output_attentions=output_attentions,
423
+ use_cache=use_cache,
424
+ )
425
+
426
+ bsz, q_len, _ = hidden_states.size()
427
+
428
+ query_states, aux_loss = self.experts.map(hidden_states)
429
+ key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)
430
+
431
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).transpose(1, 2)
432
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
433
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
434
+
435
+ kv_seq_len = key_states.shape[2]
436
+ if past_key_value is not None:
437
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
438
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
439
+
440
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, unsqueeze_dim=1)
441
+
442
+ if past_key_value is not None:
443
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
444
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
445
+
446
+ key_states = key_states.repeat(1, self.top_k, 1, 1)
447
+ value_states = value_states.repeat(1, self.top_k, 1, 1)
448
+
449
+ if attention_mask is not None:
450
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
451
+ raise ValueError(
452
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
453
+ )
454
+
455
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
456
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
457
+ if query_states.device.type == "cuda" and attention_mask is not None:
458
+ query_states = query_states.contiguous()
459
+ key_states = key_states.contiguous()
460
+ value_states = value_states.contiguous()
461
+
462
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
463
+ query_states,
464
+ key_states,
465
+ value_states,
466
+ attn_mask=attention_mask,
467
+ dropout_p=0.0,
468
+ # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
469
+ is_causal=self.is_causal and attention_mask is None and q_len > 1,
470
+ )
471
+
472
+ attn_output = attn_output.transpose(1, 2).contiguous()
473
+ attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size)
474
+
475
+ attn_output = self.experts.reduce(attn_output)
476
+ attn_output = attn_output.view(bsz, q_len, -1)
477
+
478
+ return attn_output, None, past_key_value, aux_loss
479
+
480
+
481
+ class JetMoEFlashAttention2(JetMoEAttention):
482
+ def __init__(self, *args, **kwargs):
483
+ super().__init__(*args, **kwargs)
484
+
485
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
486
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
487
+ # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
488
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
489
+
490
+ def forward(
491
+ self,
492
+ hidden_states: Optional[torch.FloatTensor],
493
+ attention_mask: Optional[torch.FloatTensor] = None,
494
+ position_ids: Optional[torch.LongTensor] = None,
495
+ past_key_value: Optional[Cache] = None,
496
+ use_cache: Optional[bool] = False,
497
+ output_attentions: Optional[bool] = False,
498
+ **kwargs,
499
+ ) -> Union[
500
+ Tuple[torch.Tensor, Tuple[torch.Tensor]],
501
+ Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
502
+ ]:
503
+ """
504
+ Forward pass of the JetMoEAttention module.
505
+
506
+ Args:
507
+ hidden_states (Optional[torch.FloatTensor]): Input hidden states.
508
+ attention_mask (Optional[torch.FloatTensor]): Attention mask.
509
+ layer_past (Optional[Tuple[torch.Tensor]]): Past layer state.
510
+ use_cache (Optional[bool]): Whether to use cached states.
511
+ output_attentions (Optional[bool]): Whether to output attention weights.
512
+
513
+ Returns:
514
+ Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Optional[Tuple[...]]]: Tuple containing outputs.
515
+ """
516
+ #assert attention_mask is None, "attention_mask is not supported"
517
+ assert output_attentions is False, "output_attentions is not supported"
518
+
519
+ B, T, C = hidden_states.size() # batch size, sequence length, embedding dimensionality (hidden_size)
520
+
521
+ # calculate query, key, values
522
+ query_layer, aux_loss = self.experts.map(hidden_states)
523
+ key_layer, value_layer = self.kv_proj(hidden_states).chunk(2, dim=-1)
524
+
525
+ query_layer = query_layer.view(B, T, self.num_heads, self.hidden_size_per_attention_head) # (B, T, k * nh, hs)
526
+ key_layer = key_layer.view(B, T, self.num_key_value_heads, self.hidden_size_per_attention_head) # (B, T, nh, hs)
527
+ value_layer = value_layer.view(B, T, self.num_key_value_heads, self.hidden_size_per_attention_head) # (B, T, nh, hs)
528
+
529
+ kv_seq_len = key_layer.shape[1]
530
+ if past_key_value is not None:
531
+ if self.layer_idx is None:
532
+ raise ValueError(
533
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
534
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
535
+ "with a layer index."
536
+ )
537
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
538
+ cos, sin = self.rotary_emb(value_layer, seq_len=kv_seq_len)
539
+ query_layer, key_layer = apply_rotary_pos_emb(query_layer, key_layer, cos, sin, position_ids)
540
+
541
+ # query_layer = query_layer.contiguous()
542
+ # expand the key_layer and value_layer [sk, b, ng, hn] -> [sk, b, np, hn]
543
+ key_layer = key_layer.repeat(1, 1, self.top_k, 1)
544
+ value_layer = value_layer.repeat(1, 1, self.top_k, 1)
545
+
546
+ if past_key_value is not None:
547
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
548
+ # print(self.layer_idx, key_layer.size())
549
+ key_layer = key_layer.transpose(1, 2)
550
+ value_layer = value_layer.transpose(1, 2)
551
+ key_layer, value_layer = past_key_value.update(key_layer, value_layer, self.layer_idx, cache_kwargs)
552
+ key_layer = key_layer.transpose(1, 2)
553
+ value_layer = value_layer.transpose(1, 2)
554
+
555
+ context_layer = self._flash_attention_forward(
556
+ query_layer,
557
+ key_layer,
558
+ value_layer,
559
+ attention_mask,
560
+ T,
561
+ )
562
+
563
+ # output projection
564
+ y = self.experts.reduce(context_layer.reshape(T, B, self.top_k, self.kv_projection_size))
565
+ y = y.view(B, T, C) # re-assemble all head outputs side by side
566
+
567
+ if not output_attentions:
568
+ attn_weights = None
569
+
570
+ return y, attn_weights, past_key_value, aux_loss
571
+
572
+ def _flash_attention_forward(
573
+ self,
574
+ query_states,
575
+ key_states,
576
+ value_states,
577
+ attention_mask,
578
+ query_length,
579
+ dropout=0.0,
580
+ softmax_scale=None,
581
+ ):
582
+ """
583
+ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
584
+ first unpad the input, then computes the attention scores and pad the final attention scores.
585
+
586
+ Args:
587
+ query_states (`torch.Tensor`):
588
+ Input query states to be passed to Flash Attention API
589
+ key_states (`torch.Tensor`):
590
+ Input key states to be passed to Flash Attention API
591
+ value_states (`torch.Tensor`):
592
+ Input value states to be passed to Flash Attention API
593
+ attention_mask (`torch.Tensor`):
594
+ The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
595
+ position of padding tokens and 1 for the position of non-padding tokens.
596
+ dropout (`float`):
597
+ Attention dropout
598
+ softmax_scale (`float`, *optional*):
599
+ The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
600
+ """
601
+ if not self._flash_attn_uses_top_left_mask:
602
+ causal = self.is_causal
603
+ else:
604
+ # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
605
+ causal = self.is_causal and query_length != 1
606
+
607
+ # Contains at least one padding token in the sequence
608
+ if attention_mask is not None:
609
+ batch_size = query_states.shape[0]
610
+ query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
611
+ query_states, key_states, value_states, attention_mask, query_length
612
+ )
613
+
614
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
615
+ max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
616
+
617
+ attn_output_unpad = flash_attn_varlen_func(
618
+ query_states,
619
+ key_states,
620
+ value_states,
621
+ cu_seqlens_q=cu_seqlens_q,
622
+ cu_seqlens_k=cu_seqlens_k,
623
+ max_seqlen_q=max_seqlen_in_batch_q,
624
+ max_seqlen_k=max_seqlen_in_batch_k,
625
+ dropout_p=dropout,
626
+ softmax_scale=softmax_scale,
627
+ causal=causal,
628
+ )
629
+
630
+ attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
631
+ else:
632
+ attn_output = flash_attn_func(
633
+ query_states,
634
+ key_states,
635
+ value_states,
636
+ dropout,
637
+ softmax_scale=softmax_scale,
638
+ causal=causal
639
+ )
640
+
641
+ return attn_output
642
+
643
+
644
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
645
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
646
+ batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
647
+
648
+ key_layer = index_first_axis(
649
+ key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
650
+ )
651
+ value_layer = index_first_axis(
652
+ value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
653
+ )
654
+ if query_length == kv_seq_len:
655
+ query_layer = index_first_axis(
656
+ query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
657
+ )
658
+ cu_seqlens_q = cu_seqlens_k
659
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
660
+ indices_q = indices_k
661
+ elif query_length == 1:
662
+ max_seqlen_in_batch_q = 1
663
+ cu_seqlens_q = torch.arange(
664
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
665
+ ) # There is a memcpy here, that is very bad.
666
+ indices_q = cu_seqlens_q[:-1]
667
+ query_layer = query_layer.squeeze(1)
668
+ else:
669
+ # The -q_len: slice assumes left padding.
670
+ attention_mask = attention_mask[:, -query_length:]
671
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
672
+
673
+ return (
674
+ query_layer,
675
+ key_layer,
676
+ value_layer,
677
+ indices_q,
678
+ (cu_seqlens_q, cu_seqlens_k),
679
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
680
+ )
681
+
682
+
683
+ JETMOE_ATTENTION_CLASSES = {
684
+ "eager": JetMoEAttention,
685
+ "flash_attention_2": JetMoEFlashAttention2,
686
+ "sdpa": JetMoESdpaAttention,
687
+ }
688
+
689
+
690
+ class JetMoEBlock(nn.Module):
691
+ def __init__(self, config: JetMoEConfig, layer_idx: Optional[int] = None):
692
+ """
693
+ Initialize the JetMoEBlock module.
694
+
695
+ Args:
696
+ config: Configuration object with model hyperparameters.
697
+ """
698
+ super().__init__()
699
+ self.input_layernorm = JetMoERMSNorm(config.hidden_size)
700
+ #self.self_attention = JetMoEAttention(config, layer_idx)
701
+ self.self_attention = JETMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
702
+ self.post_attention_layernorm = JetMoERMSNorm(config.hidden_size)
703
+
704
+ moe_args = megablocks.layers.arguments.from_megatron(config)
705
+ moe_args.activation_fn = F.silu
706
+ moe_args.return_bias = False
707
+ # self.mlp = megablocks.layers.dmoe.dMoE(moe_args)
708
+ self.mlp = moe.MoE(
709
+ input_size=config.hidden_size,
710
+ hidden_size=config.ffn_hidden_size,
711
+ num_experts=config.moe_num_experts,
712
+ activation=F.silu,
713
+ top_k=config.moe_top_k,
714
+ glu=config.glu
715
+ )
716
+
717
+ def forward(
718
+ self,
719
+ hidden_states: Optional[torch.FloatTensor],
720
+ position_ids: Optional[torch.LongTensor] = None,
721
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
722
+ attention_mask: Optional[torch.FloatTensor] = None,
723
+ output_attentions: Optional[bool] = False,
724
+ use_cache: Optional[bool] = False,
725
+ **kwargs,
726
+ ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
727
+ """
728
+ Forward pass of the JetMoEBlock module.
729
+
730
+ Args:
731
+ hidden_states (Optional[torch.FloatTensor]): Input hidden states.
732
+ layer_past (Optional[Tuple[torch.Tensor]]): Past layer state.
733
+ attention_mask (Optional[torch.FloatTensor]): Attention mask.
734
+ head_mask (Optional[torch.FloatTensor]): Head mask.
735
+ use_cache (Optional[bool]): Whether to use cached states.
736
+ output_attentions (Optional[bool]): Whether to output attention weights.
737
+
738
+ Returns:
739
+ Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
740
+ Tuple containing outputs or optional attention weights.
741
+ """
742
+ # Self Attention
743
+ attn_output, self_attn_weights, present_key_value, att_aux_loss = self.self_attention(
744
+ hidden_states=self.input_layernorm(hidden_states),
745
+ attention_mask=attention_mask,
746
+ position_ids=position_ids,
747
+ past_key_value=past_key_value,
748
+ output_attentions=output_attentions,
749
+ use_cache=use_cache,
750
+ )
751
+
752
+ hidden_states = hidden_states + attn_output
753
+ x_mlp, mlp_aux_loss = self.mlp(self.post_attention_layernorm(hidden_states))
754
+ hidden_states = hidden_states + x_mlp
755
+
756
+ outputs = (hidden_states,)
757
+
758
+ if output_attentions:
759
+ outputs += (self_attn_weights,)
760
+
761
+ if use_cache:
762
+ outputs += (present_key_value,)
763
+
764
+ outputs += (att_aux_loss + mlp_aux_loss,)
765
+
766
+ return outputs
767
+
768
+
769
+
770
+ class JetMoEPreTrainedModel(PreTrainedModel):
771
+ """
772
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
773
+ models.
774
+ """
775
+
776
+ config_class = JetMoEConfig
777
+ base_model_prefix = "transformer"
778
+ supports_gradient_checkpointing = True
779
+ _no_split_modules = ["JetMoEBlock"]
780
+ _skip_keys_device_placement = "past_key_values"
781
+ _supports_flash_attn_2 = True
782
+ _supports_sdpa = True
783
+ _supports_cache_class = True
784
+
785
+ def __init__(self, *inputs, **kwargs):
786
+ """
787
+ Initialize the JetMoEPreTrainedModel.
788
+
789
+ Args:
790
+ *inputs: Variable length input arguments.
791
+ **kwargs: Keyword arguments.
792
+ """
793
+ super().__init__(*inputs, **kwargs)
794
+
795
+ self.gradient_checkpointing = False
796
+
797
+ def _init_weights(self, module):
798
+ """Initialize the weights."""
799
+ if isinstance(module, (nn.Linear,)):
800
+ # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
801
+ # cf https://github.com/pytorch/pytorch/pull/5617
802
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
803
+ if module.bias is not None:
804
+ module.bias.data.zero_()
805
+ elif isinstance(module, nn.Embedding):
806
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
807
+ if module.padding_idx is not None:
808
+ module.weight.data[module.padding_idx].zero_()
809
+ elif isinstance(module, nn.LayerNorm):
810
+ module.bias.data.zero_()
811
+ module.weight.data.fill_(1.0)
812
+
813
+ # def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs={}):
814
+ # for module in self.modules():
815
+ # if hasattr(module, "gradient_checkpointing"):
816
+ # self._set_gradient_checkpointing(
817
+ # module, True, gradient_checkpointing_kwargs
818
+ # )
819
+
820
+ # def gradient_checkpointing_disable(self):
821
+ # for module in self.modules():
822
+ # if hasattr(module, "gradient_checkpointing"):
823
+ # self._set_gradient_checkpointing(
824
+ # module, False
825
+ # )
826
+
827
+ # def _set_gradient_checkpointing(
828
+ # self,
829
+ # module,
830
+ # value=False,
831
+ # gradient_checkpointing_kwargs={"use_reentrant": False},
832
+ # ):
833
+ # """
834
+ # Set gradient checkpointing for the JetMoEModel.
835
+
836
+ # Args:
837
+ # module: The module for which gradient checkpointing is set.
838
+ # value (bool): Whether to enable gradient checkpointing.
839
+ # """
840
+ # self._gradient_checkpointing_func = checkpoint
841
+ # self.gradient_checkpointing = True
842
+ # if isinstance(module, JetMoEModel):
843
+ # module.gradient_checkpointing = value
844
+ # module.gradient_checkpointing_kwargs = gradient_checkpointing_kwargs
845
+ # module._gradient_checkpointing_func = checkpoint
846
+
847
+ MODULEFORMER_START_DOCSTRING = r"""
848
+ This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
849
+ it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
850
+ behavior.
851
+
852
+ Parameters:
853
+ config ([`JetMoEConfig`]): Model configuration class with all the parameters of the model.
854
+ Initializing with a config file does not load the weights associated with the model, only the
855
+ configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
856
+ """
857
+
858
+ MODULEFORMER_INPUTS_DOCSTRING = r"""
859
+ Args:
860
+ input_ids (`torch.LongTensor` of shape `({0})`):
861
+ Indices of input sequence tokens in the vocabulary.
862
+
863
+ Indices can be obtained using [`AutoProcenizer`]. See [`PreTrainedTokenizer.encode`] and
864
+ [`PreTrainedTokenizer.__call__`] for details.
865
+
866
+ [What are input IDs?](../glossary#input-ids)
867
+ attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
868
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
869
+
870
+ - 1 for tokens that are **not masked**,
871
+ - 0 for tokens that are **masked**.
872
+
873
+ [What are attention masks?](../glossary#attention-mask)
874
+ token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
875
+ Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
876
+ 1]`:
877
+
878
+ - 0 corresponds to a *sentence A* token,
879
+ - 1 corresponds to a *sentence B* token.
880
+
881
+ [What are token type IDs?](../glossary#token-type-ids)
882
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
883
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
884
+ config.n_positions - 1]`.
885
+
886
+ [What are position IDs?](../glossary#position-ids)
887
+ head_mask (`torch.FloatTensor` of shape `(num_attention_heads,)` or `(n_layer, num_attention_heads)`, *optional*):
888
+ Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
889
+
890
+ - 1 indicates the head is **not masked**,
891
+ - 0 indicates the head is **masked**.
892
+
893
+ inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_dim)`, *optional*):
894
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
895
+ is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
896
+ model's internal embedding lookup matrix.
897
+ output_attentions (`bool`, *optional*):
898
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
899
+ tensors for more detail.
900
+ output_hidden_states (`bool`, *optional*):
901
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
902
+ more detail.
903
+ return_dict (`bool`, *optional*):
904
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
905
+ """
906
+
907
+
908
+ @add_start_docstrings(
909
+ "The bare JetMoE Model outputting raw hidden-states without any specific head on top.",
910
+ MODULEFORMER_START_DOCSTRING,
911
+ )
912
+ class JetMoEModel(JetMoEPreTrainedModel):
913
+ """
914
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`JetMoEBlock`]
915
+
916
+ Args:
917
+ config: JetMoEConfig
918
+ """
919
+
920
+ def __init__(self, config: JetMoEConfig):
921
+ super().__init__(config)
922
+ self.padding_idx = config.pad_token_id
923
+ self.vocab_size = config.vocab_size
924
+
925
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
926
+ self.layers = nn.ModuleList(
927
+ [JetMoEBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
928
+ )
929
+ self._attn_implementation = config._attn_implementation
930
+ self.norm = JetMoERMSNorm(config.hidden_size, eps=config.rms_norm_eps)
931
+
932
+ self.gradient_checkpointing = False
933
+ # Initialize weights and apply final processing
934
+ self.post_init()
935
+
936
+ def get_input_embeddings(self):
937
+ return self.embed_tokens
938
+
939
+ def set_input_embeddings(self, value):
940
+ self.embed_tokens = value
941
+
942
+ @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
943
+ def forward(
944
+ self,
945
+ input_ids: torch.LongTensor = None,
946
+ attention_mask: Optional[torch.Tensor] = None,
947
+ position_ids: Optional[torch.LongTensor] = None,
948
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
949
+ inputs_embeds: Optional[torch.FloatTensor] = None,
950
+ use_cache: Optional[bool] = None,
951
+ output_attentions: Optional[bool] = None,
952
+ output_hidden_states: Optional[bool] = None,
953
+ return_dict: Optional[bool] = None,
954
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
955
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
956
+ output_hidden_states = (
957
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
958
+ )
959
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
960
+
961
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
962
+
963
+ # retrieve input_ids and inputs_embeds
964
+ if input_ids is not None and inputs_embeds is not None:
965
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
966
+ elif input_ids is not None:
967
+ batch_size, seq_length = input_ids.shape
968
+ elif inputs_embeds is not None:
969
+ batch_size, seq_length, _ = inputs_embeds.shape
970
+ else:
971
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
972
+
973
+ if self.gradient_checkpointing and self.training:
974
+ if use_cache:
975
+ logger.warning_once(
976
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
977
+ )
978
+ use_cache = False
979
+
980
+ past_key_values_length = 0
981
+
982
+ if use_cache:
983
+ use_legacy_cache = not isinstance(past_key_values, Cache)
984
+ if use_legacy_cache:
985
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
986
+ past_key_values_length = past_key_values.get_usable_length(seq_length)
987
+
988
+ if position_ids is None:
989
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
990
+ position_ids = torch.arange(
991
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
992
+ )
993
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
994
+ else:
995
+ position_ids = position_ids.view(-1, seq_length).long()
996
+
997
+ if inputs_embeds is None:
998
+ inputs_embeds = self.embed_tokens(input_ids)
999
+
1000
+ if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
1001
+ is_padding_right = attention_mask[:, -1].sum().item() != batch_size
1002
+ if is_padding_right:
1003
+ raise ValueError(
1004
+ "You are attempting to perform batched generation with padding_side='right'"
1005
+ " this may lead to unexpected behaviour for Flash Attention version of JetMoE. Make sure to "
1006
+ " call `tokenizer.padding_side = 'left'` before tokenizing the input. "
1007
+ )
1008
+
1009
+ if self._attn_implementation == "flash_attention_2":
1010
+ # 2d mask is passed through the layers
1011
+ attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
1012
+ elif self._attn_implementation == "sdpa" and not output_attentions:
1013
+ # output_attentions=True can not be supported when using SDPA, and we fall back on
1014
+ # the manual implementation that requires a 4D causal mask in all cases.
1015
+ attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
1016
+ attention_mask,
1017
+ (batch_size, seq_length),
1018
+ inputs_embeds,
1019
+ past_key_values_length,
1020
+ )
1021
+ else:
1022
+ # 4d mask is passed through the layers
1023
+ attention_mask = _prepare_4d_causal_attention_mask(
1024
+ attention_mask,
1025
+ (batch_size, seq_length),
1026
+ inputs_embeds,
1027
+ past_key_values_length,
1028
+ )
1029
+
1030
+ hidden_states = inputs_embeds
1031
+
1032
+ # decoder layers
1033
+ all_hidden_states = () if output_hidden_states else None
1034
+ all_self_attns = () if output_attentions else None
1035
+ next_decoder_cache = None
1036
+
1037
+ aux_loss = 0
1038
+ for decoder_layer in self.layers:
1039
+ if output_hidden_states:
1040
+ all_hidden_states += (hidden_states,)
1041
+
1042
+ # hidden_states: Optional[torch.FloatTensor],
1043
+ # position_ids: Optional[torch.LongTensor] = None,
1044
+ # past_key_value: Optional[Tuple[torch.Tensor]] = None,
1045
+ # attention_mask: Optional[torch.FloatTensor] = None,
1046
+ # output_attentions: Optional[bool] = False,
1047
+ # use_cache: Optional[bool] = False,
1048
+
1049
+ if self.gradient_checkpointing and self.training:
1050
+ layer_outputs = self._gradient_checkpointing_func(
1051
+ #decoder_layer.__call__,
1052
+ decoder_layer,
1053
+ hidden_states,
1054
+ position_ids,
1055
+ past_key_values,
1056
+ attention_mask,
1057
+ output_attentions,
1058
+ use_cache,
1059
+ use_reentrant=False,
1060
+ )
1061
+ else:
1062
+ layer_outputs = decoder_layer(
1063
+ hidden_states,
1064
+ attention_mask=attention_mask,
1065
+ position_ids=position_ids,
1066
+ past_key_value=past_key_values,
1067
+ output_attentions=output_attentions,
1068
+ use_cache=use_cache,
1069
+ )
1070
+
1071
+ hidden_states = layer_outputs[0]
1072
+
1073
+ if use_cache:
1074
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
1075
+
1076
+ if output_attentions:
1077
+ all_self_attns += (layer_outputs[1],)
1078
+
1079
+ aux_loss += layer_outputs[-1]
1080
+
1081
+ hidden_states = self.norm(hidden_states)
1082
+
1083
+ # add hidden states from the last decoder layer
1084
+ if output_hidden_states:
1085
+ all_hidden_states += (hidden_states,)
1086
+
1087
+ next_cache = None
1088
+ if use_cache:
1089
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
1090
+
1091
+ if not return_dict:
1092
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
1093
+ return JetMoEBaseModelOutputWithPast(
1094
+ last_hidden_state=hidden_states,
1095
+ past_key_values=next_cache,
1096
+ hidden_states=all_hidden_states,
1097
+ attentions=all_self_attns,
1098
+ aux_loss=aux_loss,
1099
+ )
1100
+
1101
+
1102
+ class JetMoEForCausalLM(JetMoEPreTrainedModel):
1103
+ _tied_weights_keys = ["lm_head.weight"]
1104
+
1105
+ def __init__(self, config):
1106
+ super().__init__(config)
1107
+ self.model = JetMoEModel(config)
1108
+ self.vocab_size = config.vocab_size
1109
+ self.aux_loss_coef = getattr(config, 'aux_loss_coef', 0.01)
1110
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1111
+
1112
+ # Initialize weights and apply final processing
1113
+ self.post_init()
1114
+
1115
+ def get_input_embeddings(self):
1116
+ return self.model.embed_tokens
1117
+
1118
+ def set_input_embeddings(self, value):
1119
+ self.model.embed_tokens = value
1120
+
1121
+ def get_output_embeddings(self):
1122
+ return self.lm_head
1123
+
1124
+ def set_output_embeddings(self, new_embeddings):
1125
+ self.lm_head = new_embeddings
1126
+
1127
+ def set_decoder(self, decoder):
1128
+ self.model = decoder
1129
+
1130
+ def get_decoder(self):
1131
+ return self.model
1132
+
1133
+ @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
1134
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1135
+ def forward(
1136
+ self,
1137
+ input_ids: torch.LongTensor = None,
1138
+ attention_mask: Optional[torch.Tensor] = None,
1139
+ position_ids: Optional[torch.LongTensor] = None,
1140
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1141
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1142
+ labels: Optional[torch.LongTensor] = None,
1143
+ use_cache: Optional[bool] = None,
1144
+ output_attentions: Optional[bool] = None,
1145
+ output_hidden_states: Optional[bool] = None,
1146
+ return_dict: Optional[bool] = None,
1147
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1148
+ r"""
1149
+ Args:
1150
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1151
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1152
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1153
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1154
+
1155
+ Returns:
1156
+ """
1157
+
1158
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1159
+ output_hidden_states = (
1160
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1161
+ )
1162
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1163
+
1164
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1165
+ outputs = self.model(
1166
+ input_ids=input_ids,
1167
+ attention_mask=attention_mask,
1168
+ position_ids=position_ids,
1169
+ past_key_values=past_key_values,
1170
+ inputs_embeds=inputs_embeds,
1171
+ use_cache=use_cache,
1172
+ output_attentions=output_attentions,
1173
+ output_hidden_states=output_hidden_states,
1174
+ return_dict=return_dict,
1175
+ )
1176
+
1177
+ hidden_states = outputs[0]
1178
+ logits = self.lm_head(hidden_states)
1179
+ logits = logits.float()
1180
+
1181
+ loss = None
1182
+ if labels is not None:
1183
+ # Shift so that tokens < n predict n
1184
+ shift_logits = logits[..., :-1, :].contiguous()
1185
+ shift_labels = labels[..., 1:].contiguous()
1186
+ # Flatten the tokens
1187
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1188
+ shift_labels = shift_labels.view(-1)
1189
+ # Ensure tensors are on the same device
1190
+ shift_labels = shift_labels.to(shift_logits.device)
1191
+ loss_fct = CrossEntropyLoss()
1192
+ loss = loss_fct(shift_logits, shift_labels)
1193
+
1194
+ if not return_dict:
1195
+ output = (logits,) + outputs[1:]
1196
+ return (loss,) + output if loss is not None else output
1197
+
1198
+ if labels is not None and self.model.training:
1199
+ loss += self.aux_loss_coef * outputs.aux_loss.to(loss.device)
1200
+
1201
+ return JetMoECausalLMOutputWithPast(
1202
+ loss=loss,
1203
+ logits=logits,
1204
+ past_key_values=outputs.past_key_values,
1205
+ hidden_states=outputs.hidden_states,
1206
+ attentions=outputs.attentions,
1207
+ aux_loss=outputs.aux_loss,
1208
+ )
1209
+
1210
+ def prepare_inputs_for_generation(
1211
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
1212
+ ):
1213
+ # Omit tokens covered by past_key_values
1214
+ if past_key_values is not None:
1215
+ if isinstance(past_key_values, Cache):
1216
+ cache_length = past_key_values.get_seq_length()
1217
+ past_length = past_key_values.seen_tokens
1218
+ max_cache_length = past_key_values.get_max_length()
1219
+ else:
1220
+ cache_length = past_length = past_key_values[0][0].shape[2]
1221
+ max_cache_length = None
1222
+
1223
+ # Keep only the unprocessed tokens:
1224
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
1225
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
1226
+ # input)
1227
+ if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
1228
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
1229
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
1230
+ # input_ids based on the past_length.
1231
+ elif past_length < input_ids.shape[1]:
1232
+ input_ids = input_ids[:, past_length:]
1233
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
1234
+
1235
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
1236
+ if (
1237
+ max_cache_length is not None
1238
+ and attention_mask is not None
1239
+ and cache_length + input_ids.shape[1] > max_cache_length
1240
+ ):
1241
+ attention_mask = attention_mask[:, -max_cache_length:]
1242
+
1243
+ position_ids = kwargs.get("position_ids", None)
1244
+ if attention_mask is not None and position_ids is None:
1245
+ # create position_ids on the fly for batch generation
1246
+ position_ids = attention_mask.long().cumsum(-1) - 1
1247
+ position_ids.masked_fill_(attention_mask == 0, 1)
1248
+ if past_key_values:
1249
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1250
+
1251
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1252
+ if inputs_embeds is not None and past_key_values is None:
1253
+ model_inputs = {"inputs_embeds": inputs_embeds}
1254
+ else:
1255
+ model_inputs = {"input_ids": input_ids}
1256
+
1257
+ model_inputs.update(
1258
+ {
1259
+ "position_ids": position_ids,
1260
+ "past_key_values": past_key_values,
1261
+ "use_cache": kwargs.get("use_cache"),
1262
+ "attention_mask": attention_mask,
1263
+ }
1264
+ )
1265
+ return model_inputs
1266
+
1267
+ @staticmethod
1268
+ def _reorder_cache(past_key_values, beam_idx):
1269
+ reordered_past = ()
1270
+ for layer_past in past_key_values:
1271
+ reordered_past += (
1272
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1273
+ )
1274
+ return reordered_past
1275
+
1276
+
1277
+ @add_start_docstrings(
1278
+ """
1279
+ The JetMoE Model transformer with a sequence classification head on top (linear layer).
1280
+
1281
+ [`JetMoEForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1282
+ (e.g. GPT-2) do.
1283
+
1284
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1285
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1286
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1287
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1288
+ each row of the batch).
1289
+ """,
1290
+ MODULEFORMER_START_DOCSTRING,
1291
+ )
1292
+ # Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->JetMoE, LLAMA->MODULEFORMER
1293
+ class JetMoEForSequenceClassification(JetMoEPreTrainedModel):
1294
+ def __init__(self, config):
1295
+ super().__init__(config)
1296
+ self.num_labels = config.num_labels
1297
+ self.model = JetMoEModel(config)
1298
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1299
+
1300
+ # Initialize weights and apply final processing
1301
+ self.post_init()
1302
+
1303
+ def get_input_embeddings(self):
1304
+ return self.model.embed_tokens
1305
+
1306
+ def set_input_embeddings(self, value):
1307
+ self.model.embed_tokens = value
1308
+
1309
+ @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
1310
+ def forward(
1311
+ self,
1312
+ input_ids: torch.LongTensor = None,
1313
+ attention_mask: Optional[torch.Tensor] = None,
1314
+ position_ids: Optional[torch.LongTensor] = None,
1315
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1316
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1317
+ labels: Optional[torch.LongTensor] = None,
1318
+ use_cache: Optional[bool] = None,
1319
+ output_attentions: Optional[bool] = None,
1320
+ output_hidden_states: Optional[bool] = None,
1321
+ return_dict: Optional[bool] = None,
1322
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1323
+ r"""
1324
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1325
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1326
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1327
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1328
+ """
1329
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1330
+
1331
+ transformer_outputs = self.model(
1332
+ input_ids,
1333
+ attention_mask=attention_mask,
1334
+ position_ids=position_ids,
1335
+ past_key_values=past_key_values,
1336
+ inputs_embeds=inputs_embeds,
1337
+ use_cache=use_cache,
1338
+ output_attentions=output_attentions,
1339
+ output_hidden_states=output_hidden_states,
1340
+ return_dict=return_dict,
1341
+ )
1342
+ hidden_states = transformer_outputs[0]
1343
+ logits = self.score(hidden_states)
1344
+
1345
+ if input_ids is not None:
1346
+ batch_size = input_ids.shape[0]
1347
+ else:
1348
+ batch_size = inputs_embeds.shape[0]
1349
+
1350
+ if self.config.pad_token_id is None and batch_size != 1:
1351
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1352
+ if self.config.pad_token_id is None:
1353
+ sequence_lengths = -1
1354
+ else:
1355
+ if input_ids is not None:
1356
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1357
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1358
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1359
+ sequence_lengths = sequence_lengths.to(logits.device)
1360
+ else:
1361
+ sequence_lengths = -1
1362
+
1363
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1364
+
1365
+ loss = None
1366
+ if labels is not None:
1367
+ labels = labels.to(logits.device)
1368
+ if self.config.problem_type is None:
1369
+ if self.num_labels == 1:
1370
+ self.config.problem_type = "regression"
1371
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1372
+ self.config.problem_type = "single_label_classification"
1373
+ else:
1374
+ self.config.problem_type = "multi_label_classification"
1375
+
1376
+ if self.config.problem_type == "regression":
1377
+ loss_fct = MSELoss()
1378
+ if self.num_labels == 1:
1379
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1380
+ else:
1381
+ loss = loss_fct(pooled_logits, labels)
1382
+ elif self.config.problem_type == "single_label_classification":
1383
+ loss_fct = CrossEntropyLoss()
1384
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1385
+ elif self.config.problem_type == "multi_label_classification":
1386
+ loss_fct = BCEWithLogitsLoss()
1387
+ loss = loss_fct(pooled_logits, labels)
1388
+ if not return_dict:
1389
+ output = (pooled_logits,) + transformer_outputs[1:]
1390
+ return ((loss,) + output) if loss is not None else output
1391
+
1392
+ return JetMoESequenceClassifierOutputWithPast(
1393
+ loss=loss,
1394
+ logits=pooled_logits,
1395
+ past_key_values=transformer_outputs.past_key_values,
1396
+ hidden_states=transformer_outputs.hidden_states,
1397
+ attentions=transformer_outputs.attentions,
1398
+ aux_loss=transformer_outputs.aux_loss,
1399
+ )