arbitor/config.py

VOCAB=288
AUDIO_VOCAB=288
AUDIO_SR=16000
AUDIO_FRAME_RATE=50
THRESHOLD=0.05

# -- 3B Target Dimensions --
EMBEDDING_DIM=1536
CODEBOOK_DIM=1024
CODEBOOK_SIZE=524288            # Base unit
# Shared multimodal VQ (256K entries × 1024-dim)
SHARED_VQ_SIZE = 262144
HIDDEN_DIM=8192             # Main hidden dimension
FFN_HIDDEN=16384              # 2× HIDDEN_DIM
CTX=256

# MoEGraph (256 experts, centroid routing, unified ACT)
MG_N_EXPERTS = 256
MG_CORE_RANK = 384
MG_SHARED_INTER = 1536
MG_ACT_ITERS = 4
MG_WORKSPACE_DIM = 768
MG_TOP_K = 2

# VQ
# MemGram (32 heads × ~65K slots ≈ 2M total associative slots)
MEMGRAM_STRUCT_PRIMES = [64901, 64919, 64921, 64927, 64937, 64951, 64969, 64997,
                         65003, 65011, 65027, 65029, 65033, 65053, 65063, 65071,
                         65101, 65119, 65123, 65129, 65141, 65147, 65167, 65171,
                         65173, 65179, 65183, 65203, 65213, 65239, 65257, 65269]
MEMGRAM_CONV_PRIMES = [8009, 8011, 8017, 8039, 8081, 8087, 8089, 8093]
MEMGRAM_EMBED_DIM = 64
MEMGRAM_KEY_DIM = 32

# KV Ledger
KV_LEDGER_SIZE = 262144
SLIDING_WINDOW_SIZE = 32768
KQ_CACHE_SIZE = 8192

# MLA Attention dimensions
MLA_N_HEADS = 32
MLA_QK_NOPE_HEAD_DIM = 96
MLA_QK_ROPE_HEAD_DIM = 32
MLA_V_HEAD_DIM = 96
MLA_SLIDE_DIM = 64
MLA_FULL_DIM = 32
MLA_N_LAYERS = 24

# RoPE
MLA_ROPE_THETA = 10000.0

# Attention
ATTENTION_STRIDE = 8
KV_CONTEXT_LENGTH = 33554432

# CSA / HCA compression (DeepSeek V4 hybrid attention)
MLA_CSA_DIM = 16
MLA_HCA_DIM = 16
MLA_HCA_STRIDE = 32

# KG EMA — Phase 17
KG_EMA_ALPHA=0.99
KG_REQUANT_EVERY=50
KG_TERNARY_THRESHOLD=0.3

# Composite Motif VQ — Phase 17 (64K entries × 1024-dim)
KGVQ_CODEBOOK_SIZE=65536
KGVQ_CODEBOOK_DIM=1024
KGVQ_DECAY=0.99
KGVQ_COMMITMENT_WEIGHT=1.0
KGVQ_DEAD_CODE_THRESHOLD=2
K_MAX_COMPOSITES=20

# VideoHead (Open-Sora VAE: 4 latent channels, 8× spatial + 4× temporal compression)
VIDEO_LATENT_CHANNELS = 4
VIDEO_MAX_STEPS = 8
VIDEO_HEIGHT = 64
VIDEO_WIDTH = 64

# -- Open-Sora 3D VAE (Phase 19) --
OPEN_SORA_VAE_PATH = "arbitor/encoders/models/opensora-vae"
OPEN_SORA_VAE_REPO = "hpcai-tech/OpenSora-VAE-v1.2"
OPEN_SORA_LATENT_CHANNELS = 4
OPEN_SORA_SCALE_FACTOR_SPATIAL = 8
OPEN_SORA_SCALE_FACTOR_TEMPORAL = 4

# -- ACT Loop Parameters (Phase 19) --
BYTEHEAD_ACT_MAX_ITERS = 3
BYTEHEAD_ACT_HALT_CONSECUTIVE = 2
BYTEHEAD_ACT_PONDER_LAMBDA = 0.01

VIDEOHEAD_ACT_MIN_FPS = 1
VIDEOHEAD_ACT_MAX_FPS = 60
VIDEOHEAD_ACT_FRAME_CHUNK = 8

TALKERHEAD_ACT_CHUNK_FRAMES = 500

# -- Timestamp Encoding (Phase 19) --
TIMESTAMP_MAX_PERIOD = 10000.0

# -- Temporal Frame Buffer (Phase 19) --
FRAME_BUFFER_LOCAL_SIZE = 3
FRAME_BUFFER_CACHE_STRIDE = 4

SPECIAL_VOCAB = {
    # Control
    'PAD': 256, 'BOS': 257, 'EOS': 258, 'STOP': 259,
    # Roles
    'SYSTEM': 260, 'USER': 261, 'ASSISTANT': 262,
    # Reasoning
    'SCRATCHPAD': 263, 'PLAN': 264, 'REFLECTION': 265, 'SUMMARY': 266,
    # Tool use
    'ACTION': 267, 'TOOL': 268, 'TOOL_RESULT': 269,
    # Code
    'CODE': 270, 'CODE_BLOCK': 271, 'EXECUTION': 272,
    # RAG
    'SEARCH': 273, 'CONTEXT': 274, 'CITATION': 275,
    # Quality / format
    'ERROR': 276, 'FORMAT': 277,
    # Multimodal
    'IMAGE': 278, 'TEXT': 279, 'AUDIO': 280,
    'VIDEO': 281, 'SPEAK': 282, 'IMG_GEN': 283,
    # Future
    'RES1': 284, 'RES2': 285, 'RES3': 286, 'RESERVED': 287,
}