Skip to content

unitorch.models.peft¤

BloomLoraForClassification¤

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_bloom.py 
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
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["query_key_value"],
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    freeze_classifer: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = BloomConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.peft_model = PeftModelForSequenceClassification(
        BloomModel(self.config), self.peft_config
    )
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.hidden_size, num_classes)
    if freeze_classifer:
        for param in self.classifier.parameters():
            param.requires_grad = False
    self.init_weights()

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/peft/modeling_bloom.py 
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
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.peft_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

BloomLoraForGeneration¤

Bases: GenericPeftModel

Bloom Loar 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/peft/modeling_bloom.py 
 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
119
120
121
122
123
124
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["query_key_value"],
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Bloom Loar 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 = BloomConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.peft_model = PeftModelForCausalLM(
        BloomForCausalLM(self.config), self.peft_config
    )
    self.init_weights()

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/peft/modeling_bloom.py 
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
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.peft_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/peft/modeling_bloom.py 
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
227
228
229
230
231
232
233
234
235
236
237
238
@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.peft_model.generate(
        input_ids=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,
    )

LlamaLoraForClassification¤

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llama.py 
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
64
65
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    freeze_classifer: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    model = LlamaModel(self.config)
    if quant_config_path is not None:
        quant_config = QuantizationConfig.from_json_file(quant_config_path)
        ignore_modules = target_modules + ["lm_head"]
        model = quantize_model(model, quant_config, ignore_modules=ignore_modules)
    self.peft_model = PeftModelForSequenceClassification(model, self.peft_config)
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.hidden_size, num_classes)
    if freeze_classifer:
        for param in self.classifier.parameters():
            param.requires_grad = False
    self.init_weights()

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/peft/modeling_llama.py 
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
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.peft_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

LlamaLoraForGeneration¤

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llama.py 
105
106
107
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
def __init__(
    self,
    config_path: str,
    quant_config_path: Optional[str] = None,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    model = LlamaForCausalLM(self.config)
    if quant_config_path is not None:
        quant_config = QuantizationConfig.from_json_file(quant_config_path)
        ignore_modules = target_modules + ["lm_head"]
        model = quantize_model(model, quant_config, ignore_modules=ignore_modules)
    self.peft_model = PeftModelForCausalLM(model, self.peft_config)
    self.init_weights()

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/peft/modeling_llama.py 
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
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.peft_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/peft/modeling_llama.py 
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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
@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.peft_model.generate(
        input_ids=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,
    )