File size: 25,250 Bytes
2d7e335 9bdac3f 2d7e335 9bdac3f 2d7e335 9bdac3f 2d7e335 9bdac3f 2d7e335 9bdac3f 2d7e335 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 | """
AAM Diffusion LLM — Model Configuration
Defines all hyperparameters for the diffusion model architecture,
training process, and inference pipeline.
Design Philosophy:
- Small model (100M-500M params) — specialized, not general
- Sentence-level tokenization — not subword, because AAM arranges
sentences, not individual tokens
- Graph-conditioned — the model MUST receive graph structure as input
- Non-sequential generation — diffusion, not autoregressive
Analogi: Seperti tubuh Jin Soun, model ini kecil tapi KKHUSUS
dilatih untuk satu tugas: menarasikan dari graph. Tidak perlu
7B params kalau tugasku hanya menyusun kalimat dari data yang
sudah terstruktur.
"""
from __future__ import annotations
import json
from dataclasses import dataclass, field, asdict
from pathlib import Path
from typing import Optional
@dataclass
class ModelConfig:
"""Architecture hyperparameters for the Diffusion Transformer.
Target: 100M-500M parameters total.
Calculation:
params ≈ d_model^2 * (12 * n_layers) for transformer
d_model=512, n_layers=8 → ~50M core params
d_model=768, n_layers=12 → ~170M core params
d_model=1024, n_layers=12 → ~300M core params
"""
# --- Core Transformer ---
d_model: int = 768
"""Hidden dimension of the transformer."""
n_layers: int = 12
"""Number of transformer blocks."""
n_heads: int = 12
"""Number of attention heads (d_model must be divisible by n_heads)."""
d_ff: int = 3072
"""Feed-forward hidden dimension (typically 4x d_model)."""
dropout: float = 0.1
"""Dropout rate for attention and feed-forward layers."""
activation: str = "gelu"
"""Activation function: 'gelu' or 'relu'."""
# --- Sequence ---
max_seq_len: int = 512
"""Maximum sequence length (in sentence-level tokens)."""
# --- Vocabulary ---
vocab_size: int = 32000
"""Vocabulary size for the tokenizer.
Since we use sentence-level tokens + subword BPE hybrid,
this includes special tokens + subword units.
"""
# --- Positional Encoding ---
pos_encoding_type: str = "rotary"
"""Positional encoding type: 'rotary' (RoPE) or 'learned'."""
# --- Attention ---
use_flash_attention: bool = True
"""Whether to use Flash Attention 2 if available."""
# --- Normalization ---
norm_type: str = "rmsnorm"
"""Normalization type: 'rmsnorm' or 'layernorm'."""
norm_eps: float = 1e-6
"""Epsilon for normalization layers."""
# --- Initialization ---
init_std: float = 0.02
"""Standard deviation for weight initialization."""
def estimate_params(self) -> str:
"""Estimate total parameter count."""
# Embedding: vocab_size * d_model
embed_params = self.vocab_size * self.d_model
# Per layer: 4 * d_model^2 (QKV + O) + 2 * d_model * d_ff (FF)
layer_params = 4 * self.d_model ** 2 + 2 * self.d_model * self.d_ff
total = embed_params + self.n_layers * layer_params
if total >= 1e9:
return f"{total / 1e9:.1f}B"
elif total >= 1e6:
return f"{total / 1e6:.1f}M"
else:
return f"{total / 1e3:.1f}K"
@dataclass
class DiffusionConfig:
"""Hyperparameters for the diffusion process.
The diffusion process works on the latent representation of text:
1. Forward: Add Gaussian noise to text embeddings over T timesteps
2. Reverse: Learn to denoise step by step
3. At inference: Start from pure noise, denoise to coherent text
This is DIFFERENT from image diffusion because:
- We operate in a learned latent space (not pixel space)
- Text has discrete structure (sentences, not pixels)
- We use a text-specific noise schedule
"""
# --- Noise Schedule ---
n_timesteps: int = 1000
"""Total number of diffusion timesteps for training."""
n_inference_steps: int = 50
"""Number of denoising steps at inference (fewer = faster, less quality)."""
schedule_type: str = "cosine"
"""Noise schedule type: 'linear', 'cosine', or 'sigmoid'."""
beta_start: float = 1e-4
"""Starting beta for linear schedule."""
beta_end: float = 0.02
"""Ending beta for linear schedule."""
# --- Noise Prediction ---
prediction_type: str = "epsilon"
"""What the model predicts: 'epsilon' (noise), 'x0' (clean data),
or 'v' (velocity). Epsilon prediction is most stable for text."""
# --- Sampling ---
sampling_method: str = "ddim"
"""Sampling method: 'ddpm' (slow, stochastic) or 'ddim' (fast, deterministic)."""
eta_ddim: float = 0.0
"""DDIM stochasticity parameter (0 = deterministic, 1 = full stochastic)."""
# --- Clipping ---
clip_sample_max: float = 5.0
"""Maximum value for clipped samples during inference."""
clip_sample_min: float = -5.0
"""Minimum value for clipped samples during inference."""
# --- Loss ---
loss_type: str = "mse"
"""Loss function: 'mse' (L2) or 'mae' (L1) or 'huber'."""
loss_weighting: str = "min_snr"
"""Loss weighting strategy: 'none', 'min_snr', or 'p2'."""
p2_gamma: float = 1.0
"""P2 weighting gamma (only used if loss_weighting='p2')."""
p2_k: float = 1.0
"""P2 weighting k (only used if loss_weighting='p2')."""
@dataclass
class GraphEncoderConfig:
"""Configuration for the Graph Conditioning Encoder.
The graph encoder takes structured graph data (evidence nodes,
compositions, confidence scores, anomalies, reasoning chains)
and produces a conditioning vector that guides the diffusion process.
This is the KEY differentiator from general LLMs:
the model is conditioned on GRAPH STRUCTURE, not just text prompts.
"""
# --- Graph Encoder Architecture ---
d_graph: int = 512
"""Hidden dimension for graph encoding."""
n_graph_layers: int = 4
"""Number of graph attention layers."""
n_graph_heads: int = 8
"""Number of attention heads for graph encoding."""
# --- Input Dimensions ---
max_evidence_nodes: int = 50
"""Maximum number of evidence nodes to encode."""
max_compositions: int = 20
"""Maximum number of compositions to encode."""
max_anomalies: int = 10
"""Maximum number of anomalies to encode."""
max_reasoning_steps: int = 15
"""Maximum number of reasoning steps to encode."""
# --- Conditioning Injection ---
conditioning_method: str = "cross_attention"
"""How to inject graph conditioning into the diffusion model:
'cross_attention' (separate encoder, cross-attn in transformer)
'ada_ln' (adaptive layer norm, conditioning modulates scale/shift)
'concat' (concatenate conditioning to input sequence)
"""
# --- Confidence Embedding ---
embed_confidence: bool = True
"""Whether to embed confidence scores as part of the conditioning."""
# --- Temporal Embedding ---
embed_temporal: bool = True
"""Whether to embed temporal context (time-based relationships)."""
@dataclass
class TokenizerConfig:
"""Configuration for the AAM Sentence-Level Tokenizer.
Unlike standard BPE tokenizers that operate at subword level,
AAM's tokenizer is designed for SENTENCE ARRANGEMENT:
- Sentences are the primary unit of generation
- Within sentences, subword BPE handles individual words
- Special tokens for graph structure (evidence, anomaly, etc.)
"""
# --- BPE ---
bpe_vocab_size: int = 28000
"""Subword BPE vocabulary size (within the total vocab_size)."""
# --- Sentence-Level ---
max_sentences: int = 32
"""Maximum number of sentences in one generation."""
sentence_boundary_token: str = "<sent>"
"""Token marking sentence boundaries."""
# --- Special Tokens ---
pad_token: str = "<pad>"
bos_token: str = "<bos>"
eos_token: str = "<eos>"
mask_token: str = "<mask>"
noise_token: str = "<noise>"
# --- Graph-Structure Tokens ---
evidence_token: str = "<evidence>"
anomaly_token: str = "<anomaly>"
confidence_token: str = "<confidence>"
reasoning_token: str = "<reasoning>"
composition_token: str = "<composition>"
temporal_token: str = "<temporal>"
# --- Training ---
min_frequency: int = 2
"""Minimum frequency for BPE merge operations."""
dropout_rate: float = 0.0
"""BPE dropout rate (0 = no dropout, regularization during training)."""
@dataclass
class TrainingConfig:
"""Training hyperparameters and settings."""
# --- Optimizer ---
learning_rate: float = 1e-4
"""Peak learning rate."""
weight_decay: float = 0.01
"""Weight decay for AdamW."""
adam_beta1: float = 0.9
"""Adam beta1."""
adam_beta2: float = 0.999
"""Adam beta2."""
adam_eps: float = 1e-8
"""Adam epsilon."""
# --- Learning Rate Schedule ---
lr_schedule: str = "cosine"
"""LR schedule: 'cosine', 'linear', or 'constant'."""
warmup_steps: int = 2000
"""Number of warmup steps."""
# --- Training ---
batch_size: int = 32
"""Training batch size (per GPU)."""
gradient_accumulation_steps: int = 4
"""Gradient accumulation steps (effective batch = batch_size * this)."""
max_steps: int = 500000
"""Maximum training steps."""
max_epochs: int = 100
"""Maximum training epochs."""
# --- Regularization ---
dropout: float = 0.1
"""Training dropout rate."""
grad_clip_norm: float = 1.0
"""Gradient clipping max norm."""
# --- Mixed Precision ---
use_amp: bool = True
"""Whether to use Automatic Mixed Precision (fp16/bf16)."""
amp_dtype: str = "bf16"
"""AMP data type: 'fp16' or 'bf16'."""
# --- Checkpointing ---
save_every_steps: int = 5000
"""Save checkpoint every N steps."""
eval_every_steps: int = 1000
"""Evaluate every N steps."""
keep_last_n_checkpoints: int = 3
"""Keep only the last N checkpoints."""
# --- EMA ---
use_ema: bool = True
"""Whether to use Exponential Moving Average for inference weights."""
ema_decay: float = 0.9999
"""EMA decay rate."""
# --- Data ---
train_data_path: str = ""
"""Path to training data (JSONL format)."""
val_data_path: str = ""
"""Path to validation data (JSONL format)."""
num_workers: int = 4
"""Number of data loading workers."""
# --- Logging ---
log_every_steps: int = 100
"""Log training metrics every N steps."""
wandb_project: str = "aam-diffusion-llm"
"""Weights & Biases project name."""
wandb_run_name: str = ""
"""Weights & Biases run name (auto-generated if empty)."""
@dataclass
class InferenceConfig:
"""Inference-time configuration."""
n_steps: int = 50
"""Number of denoising steps (more = better quality, slower)."""
temperature: float = 1.0
"""Sampling temperature (1.0 = standard, <1 = more deterministic)."""
top_k: int = 50
"""Top-k sampling for token decoding."""
top_p: float = 0.95
"""Nucleus sampling threshold."""
repetition_penalty: float = 1.2
"""Penalty for repeating tokens."""
max_output_sentences: int = 16
"""Maximum number of sentences in output."""
language: str = "id"
"""Output language: 'id' (Indonesian) or 'en' (English)."""
# ---------------------------------------------------------------------------
# v2.0 Upgrade — New Module Configurations (from Losion)
# ---------------------------------------------------------------------------
@dataclass
class AnchoredDecoderConfig:
"""Configuration for Anchored Diffusion Decoder."""
d_model: int = 768
d_vocab: int = 32000
n_refine_steps: int = 3
d_refine: int = 512
use_evoformer_feedback: bool = True
n_feedback_iterations: int = 2
disambiguation_heads: int = 8
@dataclass
class FlowMatchingConfig:
"""Configuration for Flow Matching Decoder."""
d_model: int = 768
d_vocab: int = 32000
num_steps: int = 3
@dataclass
class EvoformerConfig:
"""Configuration for Evoformer Feedback System."""
d_model: int = 768
n_recycling_steps: int = 3
dropout: float = 0.0
use_layer_recycling: bool = True
use_token_recycling: bool = True
use_decoder_feedback: bool = True
use_prediction_recycling: bool = True
use_router_coevolve: bool = True
d_pair: int = 0 # 0 = use d_model
min_recycling_improvement: float = 1e-4
@dataclass
class DualMemoryConfig:
"""Configuration for Dual Memory System."""
d_model: int = 768
working_memory_size: int = 512
long_term_memory_dim: int = 256
consolidation_method: str = "attention"
retrieval_method: str = "attention"
n_retrieval_heads: int = 4
dropout: float = 0.0
@dataclass
class MCTSConfig:
"""Configuration for MCTS Reasoning Engine."""
num_simulations: int = 64
c_puct: float = 1.5
temperature: float = 1.0
max_depth: int = 10
use_value_network: bool = True
max_children: int = 8
@dataclass
class ThinkingToggleConfig:
"""Configuration for Thinking Toggle."""
d_model: int = 768
threshold: float = 0.5
@dataclass
class MatryoshkaConfig:
"""Configuration for Matryoshka Elastic Inference."""
d_model: int = 768
d_ff: int = 3072
granularity_factors: list = None # will use default_factory in __post_init__
matryoshka_loss_weight: float = 0.1
use_adaptive: bool = True
def __post_init__(self):
if self.granularity_factors is None:
self.granularity_factors = [0.25, 0.5, 0.75, 1.0]
@dataclass
class AamDiffusionConfig:
"""Master configuration for the AAM Diffusion LLM.
Combines all sub-configurations into a single object.
This is the entry point for configuring the entire framework.
"""
model: ModelConfig = field(default_factory=ModelConfig)
diffusion: DiffusionConfig = field(default_factory=DiffusionConfig)
graph_encoder: GraphEncoderConfig = field(default_factory=GraphEncoderConfig)
tokenizer: TokenizerConfig = field(default_factory=TokenizerConfig)
training: TrainingConfig = field(default_factory=TrainingConfig)
inference: InferenceConfig = field(default_factory=InferenceConfig)
# --- v2.0 Upgrades from Losion ---
anchored_decoder: AnchoredDecoderConfig = field(default_factory=AnchoredDecoderConfig)
flow_matching: FlowMatchingConfig = field(default_factory=FlowMatchingConfig)
evoformer: EvoformerConfig = field(default_factory=EvoformerConfig)
dual_memory: DualMemoryConfig = field(default_factory=DualMemoryConfig)
mcts: MCTSConfig = field(default_factory=MCTSConfig)
thinking_toggle: ThinkingToggleConfig = field(default_factory=ThinkingToggleConfig)
matryoshka: MatryoshkaConfig = field(default_factory=MatryoshkaConfig)
# --- v2.0 Feature Flags ---
use_anchored_decoder: bool = True
use_flow_matching: bool = True
use_evoformer: bool = True
use_dual_memory: bool = True
use_mcts: bool = False # Future — needs custom state representation
use_thinking_toggle: bool = True
use_matryoshka: bool = True
use_swiglu_ffn: bool = True # Replace GELU with SwiGLU
# --- Meta ---
model_name: str = "aam-diffusion-v2.0"
"""Model name for saving/loading."""
output_dir: str = "./output"
"""Base output directory."""
seed: int = 42
"""Random seed for reproducibility."""
# --- AAM Philosophy ---
aam_mind_source: str = "rsvs_graph"
"""Source of the 'mind' that conditions this 'body'.
Always 'rsvs_graph' for AAM — the model CANNOT generate
information not present in the graph conditioning."""
aam_body_type: str = "specialized_diffusion"
"""Type of the 'body'. Always 'specialized_diffusion' for AAM.
This is NOT a general LLM — it only arranges sentences
based on graph-structured evidence."""
def to_dict(self) -> dict:
"""Serialize config to dictionary."""
return asdict(self)
def to_json(self, path: str | Path) -> None:
"""Save config to JSON file."""
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
with open(path, "w", encoding="utf-8") as f:
json.dump(self.to_dict(), f, indent=2, ensure_ascii=False)
@classmethod
def from_json(cls, path: str | Path) -> AamDiffusionConfig:
"""Load config from JSON file."""
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
return cls(
model=ModelConfig(**data.get("model", {})),
diffusion=DiffusionConfig(**data.get("diffusion", {})),
graph_encoder=GraphEncoderConfig(**data.get("graph_encoder", {})),
tokenizer=TokenizerConfig(**data.get("tokenizer", {})),
training=TrainingConfig(**data.get("training", {})),
inference=InferenceConfig(**data.get("inference", {})),
# v2.0 sub-configs
anchored_decoder=AnchoredDecoderConfig(**data.get("anchored_decoder", {})),
flow_matching=FlowMatchingConfig(**data.get("flow_matching", {})),
evoformer=EvoformerConfig(**data.get("evoformer", {})),
dual_memory=DualMemoryConfig(**data.get("dual_memory", {})),
mcts=MCTSConfig(**data.get("mcts", {})),
thinking_toggle=ThinkingToggleConfig(**data.get("thinking_toggle", {})),
matryoshka=MatryoshkaConfig(**data.get("matryoshka", {})),
# v2.0 feature flags
use_anchored_decoder=data.get("use_anchored_decoder", True),
use_flow_matching=data.get("use_flow_matching", True),
use_evoformer=data.get("use_evoformer", True),
use_dual_memory=data.get("use_dual_memory", True),
use_mcts=data.get("use_mcts", False),
use_thinking_toggle=data.get("use_thinking_toggle", True),
use_matryoshka=data.get("use_matryoshka", True),
use_swiglu_ffn=data.get("use_swiglu_ffn", True),
# Meta
model_name=data.get("model_name", "aam-diffusion-v2.0"),
output_dir=data.get("output_dir", "./output"),
seed=data.get("seed", 42),
aam_mind_source=data.get("aam_mind_source", "rsvs_graph"),
aam_body_type=data.get("aam_body_type", "specialized_diffusion"),
)
def summary(self) -> str:
"""Print a summary of the configuration."""
lines = [
"=" * 60,
f" AAM Diffusion LLM Configuration: {self.model_name}",
"=" * 60,
"",
f" Model Architecture:",
f" d_model={self.model.d_model}, n_layers={self.model.n_layers}, "
f"n_heads={self.model.n_heads}",
f" d_ff={self.model.d_ff}, vocab_size={self.model.vocab_size}",
f" max_seq_len={self.model.max_seq_len}",
f" Estimated params: {self.model.estimate_params()}",
"",
f" Diffusion Process:",
f" Timesteps (train)={self.diffusion.n_timesteps}",
f" Timesteps (inference)={self.diffusion.n_inference_steps}",
f" Schedule={self.diffusion.schedule_type}",
f" Prediction={self.diffusion.prediction_type}",
f" Sampling={self.diffusion.sampling_method}",
"",
f" Graph Encoder:",
f" d_graph={self.graph_encoder.d_graph}",
f" n_layers={self.graph_encoder.n_graph_layers}",
f" Conditioning={self.graph_encoder.conditioning_method}",
f" Max evidence nodes={self.graph_encoder.max_evidence_nodes}",
"",
f" Training:",
f" LR={self.training.learning_rate}",
f" Batch={self.training.batch_size} x {self.training.gradient_accumulation_steps} accum",
f" Max steps={self.training.max_steps}",
f" AMP={self.training.use_amp} ({self.training.amp_dtype})",
"",
f" v2.0 Modules (Losion Upgrade):",
f" Anchored Decoder: {self.use_anchored_decoder} "
f"(n_refine={self.anchored_decoder.n_refine_steps})",
f" Flow Matching: {self.use_flow_matching} "
f"(num_steps={self.flow_matching.num_steps})",
f" Evoformer: {self.use_evoformer} "
f"(n_recycle={self.evoformer.n_recycling_steps})",
f" Dual Memory: {self.use_dual_memory} "
f"(working={self.dual_memory.working_memory_size})",
f" MCTS: {self.use_mcts} "
f"(simulations={self.mcts.num_simulations})",
f" Thinking Toggle: {self.use_thinking_toggle} "
f"(threshold={self.thinking_toggle.threshold})",
f" Matryoshka: {self.use_matryoshka} "
f"(factors={self.matryoshka.granularity_factors})",
f" SwiGLU FFN: {self.use_swiglu_ffn}",
"",
f" AAM Philosophy:",
f" Mind = {self.aam_mind_source} (RSVS Knowledge Graph)",
f" Body = {self.aam_body_type} (This Model)",
f" Identity = 1 Mind + 1 Body (NOT rented LLM)",
"",
"=" * 60,
]
return "\n".join(lines)
def get_default_config(
model_size: str = "base",
) -> AamDiffusionConfig:
"""Get a default configuration for different model sizes.
Args:
model_size: One of 'tiny', 'small', 'base', 'medium'.
- tiny: ~25M params (for quick testing)
- small: ~70M params (for development)
- base: ~170M params (recommended for training)
- medium: ~300M params (for final training)
Returns:
AamDiffusionConfig with appropriate settings.
"""
configs = {
"tiny": AamDiffusionConfig(
model=ModelConfig(
d_model=256,
n_layers=4,
n_heads=4,
d_ff=1024,
vocab_size=16000,
max_seq_len=256,
),
graph_encoder=GraphEncoderConfig(
d_graph=256,
n_graph_layers=2,
n_graph_heads=4,
),
diffusion=DiffusionConfig(
n_timesteps=500,
n_inference_steps=20,
),
training=TrainingConfig(
batch_size=16,
learning_rate=3e-4,
warmup_steps=500,
max_steps=100000,
),
model_name="aam-diffusion-tiny",
),
"small": AamDiffusionConfig(
model=ModelConfig(
d_model=512,
n_layers=8,
n_heads=8,
d_ff=2048,
vocab_size=24000,
max_seq_len=384,
),
graph_encoder=GraphEncoderConfig(
d_graph=384,
n_graph_layers=4,
n_graph_heads=8,
),
diffusion=DiffusionConfig(
n_timesteps=1000,
n_inference_steps=30,
),
training=TrainingConfig(
batch_size=24,
learning_rate=2e-4,
warmup_steps=1000,
max_steps=200000,
),
model_name="aam-diffusion-small",
),
"base": AamDiffusionConfig(
model=ModelConfig(
d_model=768,
n_layers=12,
n_heads=12,
d_ff=3072,
vocab_size=32000,
max_seq_len=512,
),
graph_encoder=GraphEncoderConfig(
d_graph=512,
n_graph_layers=4,
n_graph_heads=8,
),
diffusion=DiffusionConfig(
n_timesteps=1000,
n_inference_steps=50,
),
training=TrainingConfig(
batch_size=32,
learning_rate=1e-4,
warmup_steps=2000,
max_steps=500000,
),
model_name="aam-diffusion-base",
),
"medium": AamDiffusionConfig(
model=ModelConfig(
d_model=1024,
n_layers=12,
n_heads=16,
d_ff=4096,
vocab_size=32000,
max_seq_len=768,
),
graph_encoder=GraphEncoderConfig(
d_graph=768,
n_graph_layers=6,
n_graph_heads=12,
),
diffusion=DiffusionConfig(
n_timesteps=1000,
n_inference_steps=50,
),
training=TrainingConfig(
batch_size=16,
learning_rate=5e-5,
warmup_steps=5000,
max_steps=1000000,
),
model_name="aam-diffusion-medium",
),
}
if model_size not in configs:
raise ValueError(
f"Unknown model_size '{model_size}'. "
f"Choose from: {list(configs.keys())}"
)
return configs[model_size]
|