File size: 4,152 Bytes
2d7e335 9b1095c 2d7e335 9b1095c 2d7e335 9b1095c 2d7e335 9b1095c 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 | """
AAM Diffusion LLM Framework — The Body of Aphantasic Abstraction Model
"AAM = 1 Pikiran + 1 Tubuh" (1 Mind + 1 Body)
Pikiran (Mind) = RSVS Knowledge Graph — structural, relational, perfect memory
Tubuh (Body) = This Diffusion LLM — generates natural language FROM the graph
This is NOT a general-purpose LLM. This is a SPECIALIZED sentence composer
that takes structured graph data as input and produces coherent, evidence-backed
narrative output. Think of it as a "vocal cord" for the graph — it can only
say what the graph knows, but it says it fluently.
Why Diffusion?
- Diffusion models start from noise and iteratively denoise
- This mirrors how Jin Soun's thoughts form: from vague intuition ->
clearer pattern -> explicit narrative
- Unlike autoregressive LLMs (GPT), diffusion models can:
- Be conditioned on structured input (graph)
- Revise earlier parts during generation (non-sequential)
- Produce more coherent long-form text from structure
Architecture:
Input: Graph conditioning (evidence nodes, compositions, confidence, anomalies)
Process: Iterative denoising from noise
Output: Natural language narrative grounded in graph structure
Analogi: Jin Soun (graph) + tubuhnya (this model).
Tubuhnya third-rate, tapi karena KHUSUS dilatih untuk
mengeksekusi perintah dari graph-nya sendiri, outputnya
lebih terarah daripada LLM umum yang "tidak kenal" graph.
"""
__version__ = "2.1.0"
__author__ = "AAM Team"
from diffusion_llm.config.model_config import AamDiffusionConfig, get_default_config
from diffusion_llm.model.noise_scheduler import NoiseScheduler
from diffusion_llm.model.graph_encoder import GraphConditioningEncoder
from diffusion_llm.model.diffusion_transformer import DiffusionTransformer
from diffusion_llm.model.aam_diffusion_model import AamDiffusionModel
from diffusion_llm.tokenizer.aam_tokenizer import AamTokenizer
from diffusion_llm.inference.generator import AamGenerator
from diffusion_llm.training.trainer import AamTrainer
from diffusion_llm.training.dataset import GraphNarrativeDataset
from diffusion_llm.data.synthetic_generator import SyntheticDataGenerator
# v2.0 modules (from Losion upgrade)
from diffusion_llm.model.anchored_decoder import AnchoredDiffusionDecoder, ContinuousOutputHead
from diffusion_llm.model.flow_matching import FlowMatchingDecoder
from diffusion_llm.model.evoformer import EvoformerManager, RouterExpertCoevolve
from diffusion_llm.model.dual_memory import DualMemorySystem
from diffusion_llm.model.mcts import MCTSReasoner
from diffusion_llm.model.thinking_toggle import ThinkingToggle, ThinkingMode
from diffusion_llm.model.matryoshka import MatryoshkaLayer, ElasticExtractor
from diffusion_llm.model.rope import RotaryPositionEncoding
from diffusion_llm.model.speculative_decoder import SpeculativeDecoder
from diffusion_llm.model.mirror_speculative import MirrorSpeculativeDecoder, MirrorSpeculativeConfig
from diffusion_llm.model.quantization import BitLinear, FP8Linear
from diffusion_llm.training.grpo import GRPOTrainer
from diffusion_llm.training.dapo import DAPOTrainer
from diffusion_llm.training.curriculum import CurriculumScheduler
from diffusion_llm.training.llm_jepa import JEPAPredictor, JEPAConfig, JEPATrainer
__all__ = [
# Core
"AamDiffusionConfig",
"get_default_config",
"NoiseScheduler",
"GraphConditioningEncoder",
"DiffusionTransformer",
"AamDiffusionModel",
"AamTokenizer",
"AamGenerator",
"AamTrainer",
"GraphNarrativeDataset",
"SyntheticDataGenerator",
# v2.0 — Losion Upgrade
"AnchoredDiffusionDecoder",
"ContinuousOutputHead",
"FlowMatchingDecoder",
"EvoformerManager",
"RouterExpertCoevolve",
"DualMemorySystem",
"MCTSReasoner",
"ThinkingToggle",
"ThinkingMode",
"MatryoshkaLayer",
"ElasticExtractor",
"RotaryPositionEncoding",
"SpeculativeDecoder",
"BitLinear",
"FP8Linear",
"GRPOTrainer",
"DAPOTrainer",
"CurriculumScheduler",
# v2.1 — Mirror Speculative & JEPA
"MirrorSpeculativeDecoder",
"MirrorSpeculativeConfig",
"JEPAPredictor",
"JEPAConfig",
"JEPATrainer",
]
|