"""Validate codec speech preparation and adaptation settings.""" from __future__ import annotations import hashlib import re from pathlib import Path from typing import Literal, Self from pydantic import BaseModel, ConfigDict, Field, model_validator DEFAULT_RECIPE = Path('resources/codec_tts/recipe.json') REVISION_LENGTH = 40 SMOKE_PARTS = ('parity', 'throughput', 'resume', 'evaluate', 'failure') def is_run_name(value: str) -> bool: return re.fullmatch('[A-Za-z0-9][A-Za-z0-9_.-]{0,127}', value) is not None class Geometry(BaseModel): model_config = ConfigDict(extra='forbid', frozen=True, allow_inf_nan=False) hidden_size: int = Field(default=2560, ge=1) intermediate_size: int = Field(default=9728, ge=1) layers: int = Field(default=36, ge=1) heads: int = Field(default=32, ge=1) kv_heads: int = Field(default=8, ge=1) head_dim: int = Field(default=128, ge=2, multiple_of=2) text_vocabulary: int = Field(default=151936, ge=1) rms_epsilon: float = Field(default=1e-06, gt=0) rope_theta: float = Field(default=1000000, gt=0) @model_validator(mode='after') def validate_heads(self) -> Self: if self.heads % self.kv_heads: message = f'heads={self.heads} must be divisible by kv_heads={self.kv_heads}' raise ValueError(message) return self @property def adapter_parameters(self) -> int: width, query, key = (self.hidden_size, self.heads * self.head_dim, self.kv_heads * self.head_dim) return self.layers * 32 * (2 * (width + query) + 2 * (width + key) + 3 * (width + self.intermediate_size)) class Specification(BaseModel): model_config = ConfigDict(extra='forbid', frozen=True, allow_inf_nan=False) class Recipe(Specification): version: Literal[1] = 1 source: Path = Path('data/raw_v2/taqbaylit.common-voice-scripted-speech-kab-26-huge') source_revision: str = 'dc3f6c81050794a2653c76c6d2ef6518a7cd87e2' source_id: str = 'hf.boffire.common-voice-scripted-speech-kab-26-huge' license: str = 'cc0-1.0' output: Path = Path('artifacts/runs/codec-tts-data/prepared') base_id: str = 'bosonai/higgs-tts-3-4b' base_revision: str = '239f63fb7b02b1aa085f98d9efae5e35cc5523e8' codec_id: str = 'bosonai/higgs-audio-v2-tokenizer' codec_revision: str = '815a03f1b57dcb5ae918db415f22cd0c8a6f0f54' base_directory: Path = Path('artifacts/runs/codec-tts-base/native') codec_configuration: Path = Path('resources/codec_tts/codec.json') evaluation_examples: int = Field(default=16, ge=2) seed: int = Field(default=20261102, ge=0) rate_hz: Literal[24000] = 24000 min_duration_s: float = Field(default=0.25, gt=0) max_duration_s: float = Field(default=120, gt=0) max_embedded_bytes: int = Field(default=32000000, ge=1) validation_fraction: float = Field(default=0.05, gt=0, lt=1) test_fraction: float = Field(default=0.05, gt=0, lt=1) rank: Literal[32] = 32 alpha: Literal[64] = 64 dropout: float = Field(default=0, ge=0, le=0) peak_lr: float = Field(default=5e-05, gt=0) minimum_lr: float = Field(default=5e-06, gt=0) warmup_fraction: float = Field(default=0.02, gt=0, lt=1) gradient_clip: float = Field(default=1, gt=0) update_examples: int = Field(default=32, ge=1) micro_examples: int = Field(default=1, ge=1) max_mixed_positions: int = Field(default=8192, ge=1, le=8192) workers: int = Field(default=2, ge=0, le=12) checkpoint_seconds: float = Field(default=300, gt=0) memory_reserve_fraction: float = Field(default=0.1, gt=0, lt=1) activation_checkpointing: bool = True @model_validator(mode='after') def validate_relationships(self) -> Self: if self.validation_fraction + self.test_fraction >= 1: message = 'validation_fraction + test_fraction must be below 1' raise ValueError(message) if self.min_duration_s >= self.max_duration_s: message = 'min_duration_s must be below max_duration_s' raise ValueError(message) if self.minimum_lr > self.peak_lr or self.micro_examples > self.update_examples: message = 'minimum_lr/peak_lr or micro_examples/update_examples is inconsistent' raise ValueError(message) source, output = (self.source.resolve(), self.output.resolve()) if output == source or source in output.parents or output in source.parents: message = f'source and output overlap: {source}, {output}' raise ValueError(message) for name, revision in (('source_revision', self.source_revision), ('base_revision', self.base_revision), ('codec_revision', self.codec_revision)): if len(revision) != REVISION_LENGTH or any((char not in '0123456789abcdef' for char in revision)): message = f'{name} must be a full lowercase git SHA, got {revision!r}' raise ValueError(message) return self @property def digest(self) -> str: return hashlib.sha256(self.model_dump_json().encode()).hexdigest() def read_recipe(path: Path=DEFAULT_RECIPE) -> Recipe: return Recipe.model_validate_json(path.read_bytes())