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.

Source code in src/unitorch/cli/models/llama/processing.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def __init__(
    self,
    tokenizer_file: str,
    tokenizer_config: Optional[str] = None,
    special_tokens_map: Optional[str] = None,
    chat_template: Optional[str] = None,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 128,
):
    super().__init__(
        tokenizer_file=tokenizer_file,
        tokenizer_config=tokenizer_config,
        special_tokens_map=special_tokens_map,
        chat_template=chat_template,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/llama/processing.py 
43
44
45
46
47
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
85
86
87
88
89
90
91
92
93
94
95
@classmethod
@config_defaults_init("core/process/llama")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/process/llama")
    pretrained_name = config.getoption("pretrained_name", "llama-7b")

    tokenizer_file = config.getoption("tokenizer_file", None)
    tokenizer_file = pop_value(
        tokenizer_file,
        nested_dict_value(pretrained_llama_infos, pretrained_name, "tokenizer"),
    )
    tokenizer_file = cached_path(tokenizer_file)

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

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

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

    return {
        "tokenizer_file": tokenizer_file,
        "tokenizer_config": tokenizer_config,
        "special_tokens_map": special_tokens_map,
        "chat_template": chat_template,
    }

_chat_template ¤

_chat_template(messages: List[Dict[str, Any]])
Source code in src/unitorch/cli/models/llama/processing.py 
 97
 98
 99
100
101
102
@register_process("core/process/llama/chat_template")
def _chat_template(
    self,
    messages: List[Dict[str, Any]],
):
    return super().chat_template(messages=messages)

_classification ¤

_classification(
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/llama/processing.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
@register_process("core/process/llama/classification")
def _classification(
    self,
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
):
    outputs = super().classification(
        text=text,
        text_pair=text_pair,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    )

_generation_inputs ¤

_generation_inputs(
    text: str, max_seq_length: Optional[int] = None
)
Source code in src/unitorch/cli/models/llama/processing.py 
121
122
123
124
125
126
127
128
129
130
131
@register_process("core/process/llama/generation/inputs")
def _generation_inputs(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
):
    outputs = super().generation_inputs(
        text=text,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(input_ids=outputs.input_ids)

_generation_labels ¤

_generation_labels(
    text: str, max_gen_seq_length: Optional[int] = None
)
Source code in src/unitorch/cli/models/llama/processing.py 
133
134
135
136
137
138
139
140
141
142
143
144
145
146
@register_process("core/process/llama/generation/labels")
def _generation_labels(
    self,
    text: str,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation_labels(
        text=text,
        max_gen_seq_length=max_gen_seq_length,
    )
    return GenerationTargets(
        refs=outputs.input_ids,
        masks=outputs.attention_mask,
    )

_generation ¤

_generation(
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/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
@register_process("core/process/llama/generation")
def _generation(
    self,
    text: str,
    text_pair: Optional[str] = None,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation(
        text=text,
        text_pair=text_pair,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_messages_generation ¤

_messages_generation(
    messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/llama/processing.py 
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
@register_process("core/process/llama/messages/generation")
def _messages_generation(
    self,
    messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
):
    outputs = super().messages_generation(
        messages=messages,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_detokenize ¤

_detokenize(outputs: GenerationOutputs)
Source code in src/unitorch/cli/models/llama/processing.py 
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
@register_process("core/postprocess/llama/detokenize")
def _detokenize(
    self,
    outputs: GenerationOutputs,
):
    results = outputs.to_pandas()
    assert results.shape[0] == 0 or results.shape[0] == outputs.sequences.shape[0]

    decoded = super().detokenize(sequences=outputs.sequences)
    cleanup_string = lambda text: re.sub(r"\n", " ", text)
    if isinstance(decoded[0], list):
        decoded = [list(map(cleanup_string, sequence)) for sequence in decoded]
    elif isinstance(decoded[0], str):
        decoded = list(map(cleanup_string, decoded))
    else:
        raise ValueError(
            f"Unsupported type for llama detokenize: {type(decoded[0])}"
        )
    results["decoded"] = decoded
    return WriterOutputs(results)

LlamaForClassification¤

Tip

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

Bases: LlamaForClassification

Llama model for classification tasks.

Source code in src/unitorch/cli/models/llama/modeling.py 
30
31
32
33
34
35
36
37
38
39
40
def __init__(
    self,
    config_path: str,
    num_classes: Optional[int] = 1,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__(
        config_path=config_path,
        num_classes=num_classes,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/llama/modeling.py 
42
43
44
45
46
47
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
85
86
87
88
89
90
91
92
93
94
95
@classmethod
@config_defaults_init("core/model/classification/llama")
def from_config(cls, config, **kwargs):
    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)

    gradient_checkpointing = config.getoption("gradient_checkpointing", False)
    num_classes = config.getoption("num_classes", 1)

    inst = cls(
        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,
            save_base_state=False,
        )

    return inst

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/llama/modeling.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
@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,
):
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )
    return ClassificationOutputs(outputs=outputs)

LlamaForGeneration¤

Tip

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

Bases: LlamaForGeneration

Llama model for generation tasks.

Source code in src/unitorch/cli/models/llama/modeling.py 
116
117
118
119
120
121
122
123
124
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__(
        config_path=config_path,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/llama/modeling.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
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
@classmethod
@config_defaults_init("core/model/generation/llama")
def from_config(cls, config, **kwargs):
    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)

    gradient_checkpointing = config.getoption("gradient_checkpointing", False)

    inst = cls(
        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,
            save_base_state=False,
        )

    return inst

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/llama/modeling.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
@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,
):
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )
    return GenerationOutputs(sequences=outputs)

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,
)
Source code in src/unitorch/cli/models/llama/modeling.py 
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
@config_defaults_method("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,
):
    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,
    )