Skip to content

unitorch.models.llama¤

LlamaProcessor¤

Bases: HfTextClassificationProcessor, HfTextGenerationProcessor

Initialize the LlamaProcessor.

Parameters:

Name Type Description Default
vocab_file str

Path to the vocabulary file.

 required
max_seq_length int

Maximum sequence length for text classification. Defaults to 128.

 128
max_gen_seq_length int

Maximum sequence length for text generation. Defaults to 48.

 48
Source code in src/unitorch/models/llama/processing.py 
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
def __init__(
    self,
    vocab_path: Optional[str] = None,
    tokenizer_file: Optional[str] = None,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 48,
):
    """
    Initialize the LlamaProcessor.

    Args:
        vocab_file (str): Path to the vocabulary file.
        max_seq_length (int, optional): Maximum sequence length for text classification. Defaults to 128.
        max_gen_seq_length (int, optional): Maximum sequence length for text generation. Defaults to 48.
    """
    if vocab_path is not None:
        tokenizer = LlamaTokenizer(vocab_file=vocab_path)
    elif tokenizer_file is not None:
        tokenizer = LlamaTokenizerFast(
            tokenizer_file=tokenizer_file,
            bos_token="<|begin_of_text|>",
            eos_token="<|end_of_text|>",
        )
    else:
        raise ValueError("Either vocab_path or tokenizer_file must be provided")
    tokenizer.cls_token = tokenizer.bos_token
    tokenizer.sep_token = tokenizer.eos_token
    tokenizer.pad_token = tokenizer.unk_token
    tokenizer.cls_token_id = tokenizer.bos_token_id
    tokenizer.sep_token_id = tokenizer.eos_token_id
    tokenizer.pad_token_id = tokenizer.unk_token_id
    HfTextClassificationProcessor.__init__(
        self,
        tokenizer=tokenizer,
        max_seq_length=max_seq_length,
    )
    HfTextGenerationProcessor.__init__(
        self,
        tokenizer=tokenizer,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

classification ¤

classification(
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
)

Process text for classification.

Parameters:

Name Type Description Default
text str

Input text.

 required
text_pair str

Input text pair. Defaults to None.

 None
max_seq_length int

Maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed input_ids and attention_mask tensors.
Source code in src/unitorch/models/llama/processing.py 
 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
def classification(
    self,
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
):
    """
    Process text for classification.

    Args:
        text (str): Input text.
        text_pair (str, optional): Input text pair. Defaults to None.
        max_seq_length (int, optional): Maximum sequence length. Defaults to None.

    Returns:
        GenericOutputs: Processed input_ids and attention_mask tensors.
    """
    max_seq_length = pop_value(
        max_seq_length,
        self.max_seq_length,
    )

    tokens = self.tokenizer.tokenize(str(text))
    if text_pair is None:
        tokens = tokens[:max_seq_length]
        input_ids = self.tokenizer.convert_tokens_to_ids(tokens)
    else:
        tokens_pair = self.tokenizer.tokenize(str(text_pair))
        truncate_sequence_pair(tokens, tokens_pair, max_seq_length)
        tokens = tokens + tokens_pair
        input_ids = self.tokenizer.convert_tokens_to_ids(tokens)

    padding = [0] * (max_seq_length - len(input_ids))
    attention_mask = [0] * len(padding) + [1] * len(input_ids)
    input_ids = len(padding) * [self.pad_token_id] + input_ids

    assert len(input_ids) == max_seq_length
    assert len(attention_mask) == max_seq_length
    return GenericOutputs(
        input_ids=torch.tensor(input_ids, dtype=torch.long),
        attention_mask=torch.tensor(attention_mask, dtype=torch.long),
    )

generation ¤

generation(
    text: str,
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)

Process text for generation.

Parameters:

Name Type Description Default
text str

Input text.

 required
text_pair str

Input text pair.

 required
max_seq_length int

Maximum sequence length. Defaults to None.

 None
max_gen_seq_length int

Maximum generation sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed input_ids, attention_mask, input_ids_label, and attention_mask_label tensors.
Source code in src/unitorch/models/llama/processing.py 
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
def generation(
    self,
    text: str,
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    """
    Process text for generation.

    Args:
        text (str): Input text.
        text_pair (str): Input text pair.
        max_seq_length (int, optional): Maximum sequence length. Defaults to None.
        max_gen_seq_length (int, optional): Maximum generation sequence length. Defaults to None.

    Returns:
        GenericOutputs: Processed input_ids, attention_mask, input_ids_label, and attention_mask_label tensors.
    """
    max_seq_length = pop_value(
        max_seq_length,
        self.max_seq_length,
    )
    max_gen_seq_length = pop_value(
        max_gen_seq_length,
        self.max_gen_seq_length,
    )

    tokens = [self.bos_token] + self.tokenizer.tokenize(str(text))[
        1 - max_seq_length :
    ]
    tokens_pair = self.tokenizer.tokenize(str(text_pair))[
        : max_gen_seq_length - 1
    ] + [self.eos_token]
    padding_a = [self.pad_token] * (max_seq_length - len(tokens))
    padding_b = [self.pad_token] * (max_gen_seq_length - len(tokens_pair))
    attention_mask = (
        [0] * len(padding_a)
        + [1] * (len(tokens) + len(tokens_pair))
        + [0] * len(padding_b)
    )
    _tokens = padding_a + tokens + tokens_pair + padding_b
    input_ids = self.tokenizer.convert_tokens_to_ids(_tokens)

    tokens_label = tokens_pair + [self.pad_token] * (
        max_gen_seq_length - len(tokens_pair) + 1
    )
    input_ids_label = self.tokenizer.convert_tokens_to_ids(tokens_label)
    input_ids_label = [0] * (max_seq_length - 1) + input_ids_label
    attention_mask_label = [1] * len(tokens_pair) + [0] * (
        max_gen_seq_length - len(tokens_pair) + 1
    )
    attention_mask_label = [0] * (max_seq_length - 1) + attention_mask_label

    return GenericOutputs(
        input_ids=torch.tensor(input_ids, dtype=torch.long),
        attention_mask=torch.tensor(attention_mask, dtype=torch.long),
        input_ids_label=torch.tensor(input_ids_label, dtype=torch.long),
        attention_mask_label=torch.tensor(attention_mask_label, dtype=torch.long),
    )

generation_inputs ¤

generation_inputs(
    text: str, max_seq_length: Optional[int] = None
)

Process text for generation inputs.

Parameters:

Name Type Description Default
text str

Input text.

 required
max_seq_length int

Maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed input_ids tensor.
Source code in src/unitorch/models/llama/processing.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
def generation_inputs(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
):
    """
    Process text for generation inputs.

    Args:
        text (str): Input text.
        max_seq_length (int, optional): Maximum sequence length. Defaults to None.

    Returns:
        GenericOutputs: Processed input_ids tensor.
    """
    max_seq_length = pop_value(
        max_seq_length,
        self.max_seq_length,
    )
    tokens = [self.bos_token] + self.tokenizer.tokenize(str(text))[
        1 - max_seq_length :
    ]
    padding = [self.pad_token] * (max_seq_length - len(tokens))
    tokens = padding + tokens
    input_ids = self.tokenizer.convert_tokens_to_ids(tokens)

    assert len(input_ids) == max_seq_length
    return GenericOutputs(
        input_ids=torch.tensor(input_ids, dtype=torch.long),
    )

generation_labels ¤

generation_labels(
    text: str, max_gen_seq_length: Optional[int] = None
)

Process text for generation labels.

Parameters:

Name Type Description Default
text str

Input text.

 required
max_gen_seq_length int

Maximum generation sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs

Processed input_ids and attention_mask tensors.
Source code in src/unitorch/models/llama/processing.py 
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
def generation_labels(
    self,
    text: str,
    max_gen_seq_length: Optional[int] = None,
):
    """
    Process text for generation labels.

    Args:
        text (str): Input text.
        max_gen_seq_length (int, optional): Maximum generation sequence length. Defaults to None.

    Returns:
        GenericOutputs: Processed input_ids and attention_mask tensors.
    """
    max_gen_seq_length = pop_value(
        max_gen_seq_length,
        self.max_gen_seq_length,
    )
    tokens = self.tokenizer.tokenize(str(text))[: max_gen_seq_length - 1] + [
        self.eos_token
    ]
    padding = [self.pad_token] * (max_gen_seq_length - len(tokens))
    input_ids = self.tokenizer.convert_tokens_to_ids(tokens)
    attention_mask = [1] * len(input_ids)

    padding = [0] * (max_gen_seq_length - len(input_ids))
    input_ids += [self.pad_token_id] * len(padding)
    attention_mask += padding

    assert len(input_ids) == max_gen_seq_length
    assert len(attention_mask) == max_gen_seq_length
    return GenericOutputs(
        input_ids=torch.tensor(input_ids, dtype=torch.long),
        attention_mask=torch.tensor(attention_mask, dtype=torch.long),
    )

instruction_generation_inputs ¤

instruction_generation_inputs(
    instruction: str,
    input: str,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs

Preprocesses text as generation inputs.

Parameters:

Name Type Description Default
text str

The input text for generation.

 required
max_seq_length Optional[int]

The maximum sequence length. Defaults to None.

 None

Returns:

Name Type Description
GenericOutputs GenericOutputs

The processed input IDs tensor.
Source code in src/unitorch/models/llama/processing.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
def instruction_generation_inputs(
    self,
    instruction: str,
    input: str,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    """
    Preprocesses text as generation inputs.

    Args:
        text (str): The input text for generation.
        max_seq_length (Optional[int]): The maximum sequence length. Defaults to None.

    Returns:
        GenericOutputs: The processed input IDs tensor.
    """
    max_seq_length = pop_value(
        max_seq_length,
        self.max_seq_length,
    )
    tokens = self._instrution_tokenize(instruction, input, max_seq_length)
    padding = [self.pad_token] * (max_seq_length - len(tokens))
    tokens = padding + tokens
    input_ids = self.tokenizer.convert_tokens_to_ids(tokens)

    assert len(input_ids) == max_seq_length
    assert len(input_ids) == max_seq_length
    return GenericOutputs(
        input_ids=torch.tensor(input_ids, dtype=torch.long),
    )

LlamaForClassification¤

Bases: GenericModel, QuantizationMixin, PeftWeightLoaderMixin

Llama model for classification tasks.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
num_classes int

Number of classes for classification. Defaults to 1.

 1
hidden_dropout_prob float

Dropout probability for hidden layers. Defaults to 0.1.

 0.1
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/llama/modeling.py 
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Llama model for classification tasks.

    Args:
        config_path (str): Path to the model configuration file.
        num_classes (int, optional): Number of classes for classification. Defaults to 1.
        hidden_dropout_prob (float, optional): Dropout probability for hidden layers. Defaults to 0.1.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.model = LlamaModel(self.config)
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.hidden_size, num_classes)
    self.init_weights()

    if quant_config_path is not None:
        self.quant_config = QuantizationConfig.from_json_file(quant_config_path)
        self.quantize(self.quant_config, ignore_modules=["lm_head"])

forward ¤

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

Forward pass of the classification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch Output logits.Tensor: tensor of shape (batch_size, num_classes).
Source code in src/unitorch/models/llama/modeling.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the classification model.

    Args:
        input_ids (torch.Tensor): Input tensor of shape (batch_size, sequence_length).
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch Output logits.Tensor: tensor of shape (batch_size, num_classes).
    """
    outputs = self.model(
        input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )[0]
    pooled_output = outputs[:, -1]
    pooled_output = self.dropout(pooled_output)
    logits = self.classifier(pooled_output)
    return logits

LlamaForGeneration¤

Bases: GenericModel, QuantizationMixin, PeftWeightLoaderMixin

Llama model for text generation tasks.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/llama/modeling.py 
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Llama model for text generation tasks.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.base_model = LlamaForCausalLM(self.config)
    self.init_weights()

    if quant_config_path is not None:
        self.quant_config = QuantizationConfig.from_json_file(quant_config_path)
        self.quantize(self.quant_config, ignore_modules=["lm_head"])

forward ¤

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

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch Output logits.Tensor: tensor of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/llama/modeling.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch Output logits.Tensor: tensor of shape (batch_size, sequence_length, vocab_size).
    """
    outputs = self.base_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
        return_dict=True,
    )
    logits = outputs.logits
    return logits

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 text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 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 length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. 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

Penalty for diverse sequences in 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 value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/llama/modeling.py 
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
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
@torch.no_grad()
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 text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        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 length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in 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 value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.base_model.generate(
        input_ids,
        max_length=max_gen_seq_length + input_seq_length,
        min_length=min_gen_seq_length + input_seq_length,
        num_beams=num_beams,
        do_sample=do_sample,
        decoder_start_token_id=decoder_start_token_id,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        num_return_sequences=num_return_sequences,
        bos_token_id=decoder_start_token_id,
        eos_token_id=decoder_end_token_id,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        return_dict_in_generate=True,
        output_scores=True,
    )

    sequences = outputs.sequences.reshape(
        -1, num_return_sequences, outputs.sequences.size(-1)
    )
    outputs.sequences = torch.zeros(
        sequences.size(0), num_return_sequences, max_gen_seq_length
    ).to(device=sequences.device)
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

    if num_return_sequences == 1:
        outputs.sequences = outputs.sequences.reshape(-1, max_gen_seq_length)

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )