File size: 7,500 Bytes
178abc4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
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
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
119
120
121
122
123
124
125
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
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
"""
Training configuration schemas — Pydantic v2.

All training jobs are validated against these models before execution.
No raw dicts escape into the pipeline; everything is typed and constrained.
"""

from __future__ import annotations

from enum import StrEnum
from typing import Annotated

from pydantic import BaseModel, Field, HttpUrl, model_validator
from pydantic import PositiveFloat, PositiveInt


# ---------------------------------------------------------------------------
# Enums
# ---------------------------------------------------------------------------
class EvalStrategy(StrEnum):
    NO = "no"
    STEPS = "steps"
    EPOCH = "epoch"


class Precision(StrEnum):
    FP32 = "fp32"
    FP16 = "fp16"
    BF16 = "bf16"
    INT8 = "int8"


class OptimizerType(StrEnum):
    ADAMW = "adamw_torch"
    ADAMW_8BIT = "adamw_8bit"
    PAGED_ADAMW_8BIT = "paged_adamw_8bit"
    SGD = "sgd"


class EvalMetric(StrEnum):
    PASS_AT_1 = "pass_at_1"
    PASS_AT_10 = "pass_at_10"
    BLEU = "bleu"
    EXECUTION_ACCURACY = "execution_accuracy"
    EXACT_MATCH = "exact_match"


# ---------------------------------------------------------------------------
# Sub-configs
# ---------------------------------------------------------------------------
class LoRAConfig(BaseModel):
    """LoRA adapter configuration. Omit to disable LoRA (full fine-tune)."""

    enabled: bool = True
    r: Annotated[int, Field(ge=1, le=256)] = 16
    alpha: Annotated[int, Field(ge=1)] = 32
    dropout: Annotated[float, Field(ge=0.0, lt=1.0)] = 0.05
    target_modules: list[str] = Field(
        default_factory=lambda: ["q_proj", "v_proj"],
        min_length=1,
    )
    bias: str = "none"

    @model_validator(mode="after")
    def alpha_geq_r(self) -> "LoRAConfig":
        if self.alpha < self.r:
            raise ValueError(f"lora.alpha ({self.alpha}) should be >= lora.r ({self.r})")
        return self


class TrainingHyperparams(BaseModel):
    num_epochs: Annotated[int, Field(ge=1, le=100)] = 3
    batch_size: Annotated[int, Field(ge=1, le=256)] = 8
    gradient_accumulation_steps: Annotated[int, Field(ge=1, le=128)] = 4
    learning_rate: Annotated[float, Field(gt=0.0, lt=1.0)] = 2e-5
    weight_decay: Annotated[float, Field(ge=0.0, lt=1.0)] = 0.01
    warmup_ratio: Annotated[float, Field(ge=0.0, lt=1.0)] = 0.1
    max_seq_length: Annotated[int, Field(ge=64, le=32768)] = 1024
    max_grad_norm: Annotated[float, Field(gt=0.0)] = 1.0
    optimizer: OptimizerType = OptimizerType.ADAMW
    precision: Precision = Precision.BF16
    lr_scheduler: str = "cosine"
    seed: int = 42
    dataloader_num_workers: Annotated[int, Field(ge=0, le=32)] = 4

    @property
    def effective_batch_size(self) -> int:
        return self.batch_size * self.gradient_accumulation_steps


class EvaluationConfig(BaseModel):
    enabled: bool = True
    strategy: EvalStrategy = EvalStrategy.EPOCH
    eval_steps: PositiveInt | None = None  # required when strategy=STEPS
    metrics: list[EvalMetric] = Field(
        default_factory=lambda: [EvalMetric.PASS_AT_1, EvalMetric.BLEU]
    )
    num_samples_per_problem: Annotated[int, Field(ge=1, le=200)] = 10
    timeout_seconds: Annotated[int, Field(ge=1, le=60)] = 10
    load_best_model_at_end: bool = True
    metric_for_best_model: EvalMetric = EvalMetric.PASS_AT_1
    greater_is_better: bool = True

    @model_validator(mode="after")
    def eval_steps_required_for_steps_strategy(self) -> "EvaluationConfig":
        if self.strategy == EvalStrategy.STEPS and self.eval_steps is None:
            raise ValueError("evaluation.eval_steps is required when strategy='steps'")
        return self


class CheckpointConfig(BaseModel):
    save_strategy: EvalStrategy = EvalStrategy.EPOCH
    save_steps: PositiveInt | None = None
    save_total_limit: Annotated[int, Field(ge=1, le=20)] = 3
    output_dir: str = "./checkpoints"
    resume_from_checkpoint: str | None = None

    @model_validator(mode="after")
    def save_steps_required_for_steps_strategy(self) -> "CheckpointConfig":
        if self.save_strategy == EvalStrategy.STEPS and self.save_steps is None:
            raise ValueError("checkpoint.save_steps required when save_strategy='steps'")
        return self


class HubConfig(BaseModel):
    push_to_hub: bool = False
    repo_id: str | None = None
    private: bool = True
    commit_message: str = "Training checkpoint"

    @model_validator(mode="after")
    def repo_id_required_if_pushing(self) -> "HubConfig":
        if self.push_to_hub and not self.repo_id:
            raise ValueError("hub.repo_id is required when hub.push_to_hub=true")
        return self


class DatasetConfig(BaseModel):
    dataset_id: str  # internal UUID or HF Hub dataset path
    split_ratio: Annotated[float, Field(gt=0.0, lt=1.0)] = 0.9  # train split
    max_samples: PositiveInt | None = None  # None = use all
    text_column: str = "content"
    shuffle: bool = True
    shuffle_seed: int = 42


# ---------------------------------------------------------------------------
# Root job config
# ---------------------------------------------------------------------------
class TrainingJobConfig(BaseModel):
    """
    Complete training job specification.

    Validated at job submission time. If validation passes, the job is
    guaranteed to reach the pipeline with a coherent configuration.
    """

    job_name: Annotated[str, Field(min_length=1, max_length=128, pattern=r"^[\w\-]+$")]
    base_model: str = Field(
        description="HuggingFace model ID or local path",
        examples=["Salesforce/codegen-350M-mono", "deepseek-ai/deepseek-coder-1.3b-base"],
    )
    dataset: DatasetConfig
    training: TrainingHyperparams = Field(default_factory=TrainingHyperparams)
    lora: LoRAConfig | None = Field(default_factory=LoRAConfig)
    evaluation: EvaluationConfig = Field(default_factory=EvaluationConfig)
    checkpoint: CheckpointConfig = Field(default_factory=CheckpointConfig)
    hub: HubConfig = Field(default_factory=HubConfig)
    tags: list[str] = Field(default_factory=list, max_length=20)
    notes: str | None = None

    model_config = {
        "json_schema_extra": {
            "examples": [
                {
                    "job_name": "codegen-finetune-v1",
                    "base_model": "Salesforce/codegen-350M-mono",
                    "dataset": {"dataset_id": "ds_abc123"},
                    "training": {
                        "num_epochs": 3,
                        "batch_size": 8,
                        "learning_rate": 2e-5,
                    },
                    "hub": {
                        "push_to_hub": True,
                        "repo_id": "your-org/codegen-finetune-v1",
                    },
                }
            ]
        }
    }


# ---------------------------------------------------------------------------
# Inference config (served separately but validated here for consistency)
# ---------------------------------------------------------------------------
class InferenceConfig(BaseModel):
    model_id: str
    max_new_tokens: Annotated[int, Field(ge=1, le=4096)] = 256
    temperature: Annotated[float, Field(ge=0.0, le=2.0)] = 0.2
    top_p: Annotated[float, Field(ge=0.0, le=1.0)] = 0.95
    top_k: Annotated[int, Field(ge=0, le=1000)] = 50
    do_sample: bool = True
    num_return_sequences: Annotated[int, Field(ge=1, le=200)] = 1
    stop_sequences: list[str] = Field(default_factory=list)
    precision: Precision = Precision.BF16