Skip to content

unitorch.cli.models.llama¤

LlamaProcessor¤

Tip

core/process/llama is the section for configuration of LlamaProcessor.

Bases: LlamaProcessor

Processor for Llama models.

Initialize the LlamaProcessor.

Parameters:

Name Type Description Default
vocab_path str

The path to the vocabulary file.

 required
max_seq_length int

The maximum sequence length. Defaults to 128.

 128
max_gen_seq_length int

The maximum generated sequence length. Defaults to 128.

 128
Source code in src/unitorch/cli/models/llama/processing.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
def __init__(
    self,
    vocab_path,
    tokenizer_file,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 128,
):
    """
    Initialize the LlamaProcessor.

    Args:
        vocab_path (str): The path to the vocabulary file.
        max_seq_length (int, optional): The maximum sequence length. Defaults to 128.
        max_gen_seq_length (int, optional): The maximum generated sequence length. Defaults to 128.
    """
    super().__init__(
        vocab_path=vocab_path,
        tokenizer_file=tokenizer_file,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

from_core_configure classmethod ¤

from_core_configure(config, **kwargs)

Create an instance of LlamaProcessor from a core configuration.

Parameters:

Name Type Description Default
config

The core configuration.

 required
**kwargs

Additional keyword arguments.

 {}

Returns:

Name Type Description
LlamaProcessor

An instance of LlamaProcessor.
Source code in src/unitorch/cli/models/llama/processing.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
@classmethod
@add_default_section_for_init("core/process/llama")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of LlamaProcessor from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        LlamaProcessor: An instance of LlamaProcessor.
    """
    config.set_default_section("core/process/llama")
    pretrained_name = config.getoption("pretrained_name", "llama-7b")
    vocab_path = config.getoption("vocab_path", None)
    vocab_path = pop_value(
        vocab_path,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "vocab"),
        check_none=False,
    )
    vocab_path = cached_path(vocab_path) if vocab_path is not None else None

    tokenizer_file = config.getoption("tokenizer_file", None)
    tokenizer_file = pop_value(
        tokenizer_file,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "tokenizer"),
        check_none=False,
    )
    tokenizer_file = (
        cached_path(tokenizer_file) if tokenizer_file is not None else None
    )

    return {
        "vocab_path": vocab_path,
        "tokenizer_file": tokenizer_file,
    }

LlamaForClassification¤

Tip

core/model/classification/llama is the section for configuration of LlamaForClassification.

Bases: LlamaForClassification

Llama model for classification tasks.

Initialize the LlamaForClassification model.

Parameters:

Name Type Description Default
config_path str

The path to the model configuration file.

 required
num_classes int

The number of classes for classification. Defaults to 1.

 1
gradient_checkpointing bool

Whether to use gradient checkpointing during training. Defaults to False.

 False
Source code in src/unitorch/cli/models/llama/modeling.py 
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    num_classes: Optional[int] = 1,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initialize the LlamaForClassification model.

    Args:
        config_path (str): The path to the model configuration file.
        num_classes (int, optional): The number of classes for classification. Defaults to 1.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing during training. Defaults to False.
    """
    super().__init__(
        config_path=config_path,
        quant_config_path=quant_config_path,
        num_classes=num_classes,
        gradient_checkpointing=gradient_checkpointing,
    )

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Perform a forward pass on the LlamaForClassification model.

Parameters:

Name Type Description Default
input_ids Tensor

The input tensor containing the input IDs.

 required
attention_mask Tensor

The attention mask tensor. Defaults to None.

 None
position_ids Tensor

The position IDs tensor. Defaults to None.

 None

Returns:

Name Type Description
ClassificationOutputs

The output of the classification model.
Source code in src/unitorch/cli/models/llama/modeling.py 
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Perform a forward pass on the LlamaForClassification model.

    Args:
        input_ids (torch.Tensor): The input tensor containing the input IDs.
        attention_mask (torch.Tensor, optional): The attention mask tensor. Defaults to None.
        position_ids (torch.Tensor, optional): The position IDs tensor. Defaults to None.

    Returns:
        ClassificationOutputs: The output of the classification model.
    """
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )
    return ClassificationOutputs(outputs=outputs)

from_core_configure classmethod ¤

from_core_configure(config, **kwargs)

Create an instance of LlamaForClassification from a core configuration.

Parameters:

Name Type Description Default
config

The core configuration.

 required
**kwargs

Additional keyword arguments.

 {}

Returns:

Name Type Description
LlamaForClassification

An instance of LlamaForClassification.
Source code in src/unitorch/cli/models/llama/modeling.py 
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
@classmethod
@add_default_section_for_init("core/model/classification/llama")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of LlamaForClassification from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        LlamaForClassification: An instance of LlamaForClassification.
    """
    config.set_default_section("core/model/classification/llama")
    pretrained_name = config.getoption("pretrained_name", "llama-7b")
    pretrained_lora_name = config.getoption("pretrained_lora_name", "llama-7b-lora")
    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    quant_config_path = config.getoption("quant_config_path", None)
    if quant_config_path is not None:
        quant_config_path = cached_path(quant_config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)
    num_classes = config.getoption("num_classes", 1)

    inst = cls(
        config_path,
        quant_config_path=quant_config_path,
        num_classes=num_classes,
        gradient_checkpointing=gradient_checkpointing,
    )

    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "weight"),
        check_none=False,
    )

    if weight_path is not None:
        inst.from_pretrained(
            weight_path=weight_path,
        )

    pretrained_lora_weight_path = config.getoption(
        "pretrained_lora_weight_path", None
    )
    lora_weight_path = pop_value(
        pretrained_lora_weight_path,
        nested_dict_value(pretrained_llama_extensions_infos, pretrained_lora_name),
        check_none=False,
    )
    pretrained_lora_weight = config.getoption("pretrained_lora_weight", 1.0)
    pretrained_lora_alpha = config.getoption("pretrained_lora_alpha", 32.0)
    if lora_weight_path is not None:
        inst.load_lora_weights(
            lora_weight_path,
            lora_weights=pretrained_lora_weight,
            lora_alphas=pretrained_lora_alpha,
        )

    return inst

LlamaForGeneration¤

Tip

core/model/generation/llama is the section for configuration of LlamaForGeneration.

Bases: LlamaForGeneration

Llama model for generation tasks.

Initialize the LlamaForGeneration model.

Parameters:

Name Type Description Default
config_path str

The path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing during training. Defaults to False.

 False
Source code in src/unitorch/cli/models/llama/modeling.py 
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initialize the LlamaForGeneration model.

    Args:
        config_path (str): The path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing during training. Defaults to False.
    """
    super().__init__(
        config_path=config_path,
        quant_config_path=quant_config_path,
        gradient_checkpointing=gradient_checkpointing,
    )

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Perform a forward pass on the LlamaForGeneration model.

Parameters:

Name Type Description Default
input_ids Tensor

The input tensor containing the input IDs. Defaults to None.

 required
attention_mask Tensor

The attention mask tensor. Defaults to None.

 None
position_ids Tensor

The position IDs tensor. Defaults to None.

 None

Returns:

Name Type Description
GenerationOutputs

The output of the generation model.
Source code in src/unitorch/cli/models/llama/modeling.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Perform a forward pass on the LlamaForGeneration model.

    Args:
        input_ids (torch.Tensor, optional): The input tensor containing the input IDs. Defaults to None.
        attention_mask (torch.Tensor, optional): The attention mask tensor. Defaults to None.
        position_ids (torch.Tensor, optional): The position IDs tensor. Defaults to None.

    Returns:
        GenerationOutputs: The output of the generation model.
    """
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )
    return GenerationOutputs(sequences=outputs)

from_core_configure classmethod ¤

from_core_configure(config, **kwargs)

Create an instance of LlamaForGeneration from a core configuration.

Parameters:

Name Type Description Default
config

The core configuration.

 required
**kwargs

Additional keyword arguments.

 {}

Returns:

Name Type Description
LlamaForGeneration

An instance of LlamaForGeneration.
Source code in src/unitorch/cli/models/llama/modeling.py 
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
@classmethod
@add_default_section_for_init("core/model/generation/llama")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of LlamaForGeneration from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        LlamaForGeneration: An instance of LlamaForGeneration.
    """
    config.set_default_section("core/model/generation/llama")
    pretrained_name = config.getoption("pretrained_name", "llama-7b")
    pretrained_lora_name = config.getoption("pretrained_lora_name", "llama-7b-lora")
    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    quant_config_path = config.getoption("quant_config_path", None)
    if quant_config_path is not None:
        quant_config_path = cached_path(quant_config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)

    inst = cls(
        config_path,
        quant_config_path=quant_config_path,
        gradient_checkpointing=gradient_checkpointing,
    )
    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "weight"),
        check_none=False,
    )

    if weight_path is not None:
        inst.from_pretrained(
            weight_path=weight_path,
        )

    pretrained_lora_weight_path = config.getoption(
        "pretrained_lora_weight_path", None
    )
    lora_weight_path = pop_value(
        pretrained_lora_weight_path,
        nested_dict_value(pretrained_llama_extensions_infos, pretrained_lora_name),
        check_none=False,
    )
    pretrained_lora_weight = config.getoption("pretrained_lora_weight", 1.0)
    pretrained_lora_alpha = config.getoption("pretrained_lora_alpha", 32.0)
    if lora_weight_path is not None:
        inst.load_lora_weights(
            lora_weight_path,
            lora_weights=pretrained_lora_weight,
            lora_alphas=pretrained_lora_alpha,
        )

    return inst

generate ¤

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
)

Generate sequences using the Llama model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

Decoder start token ID. Defaults to 1.

 1
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 2
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum generation sequence length. Defaults to 0.

 0
max_gen_seq_length int

Maximum generation sequence length. Defaults to 48.

 48
repetition_penalty float

Repetition penalty. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to prevent repetition. Defaults to 0.

 0
early_stopping bool

Whether to perform early stopping. Defaults to True.

 True
length_penalty float

Length penalty. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Diversity penalty for diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k sampling parameter. Defaults to 50.

 50
top_p float

Top-p sampling parameter. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenerationOutputs

The generation outputs.
Source code in src/unitorch/cli/models/llama/modeling.py 
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
@add_default_section_for_function("core/model/generation/llama")
@torch.no_grad()
@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
):
    """
    Generate sequences using the Llama model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): Decoder start token ID. Defaults to 1.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum generation sequence length. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum generation sequence length. Defaults to 48.
        repetition_penalty (float, optional): Repetition penalty. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to prevent repetition. Defaults to 0.
        early_stopping (bool, optional): Whether to perform early stopping. Defaults to True.
        length_penalty (float, optional): Length penalty. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Diversity penalty for diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k sampling parameter. Defaults to 50.
        top_p (float, optional): Top-p sampling parameter. Defaults to 1.0.

    Returns:
        GenerationOutputs: The generation outputs.
    """
    outputs = super().generate(
        input_ids,
        num_beams=num_beams,
        decoder_start_token_id=decoder_start_token_id,
        decoder_end_token_id=decoder_end_token_id,
        num_return_sequences=num_return_sequences,
        min_gen_seq_length=min_gen_seq_length,
        max_gen_seq_length=max_gen_seq_length,
        repetition_penalty=repetition_penalty,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        do_sample=do_sample,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
    )

    return GenerationOutputs(
        sequences=outputs.sequences,
        sequences_scores=outputs.sequences_scores,
    )