"""Built-in model architecture presets for offline quick-start."""

from __future__ import annotations

from .report import ModelArch


PRESETS: dict[str, ModelArch] = {
    "Llama-3 8B": ModelArch(
        name="Llama-3 8B", architecture="llama",
        n_layer=32, n_embd=4096, n_head=32, n_head_kv=8,
        training_ctx=8192, params=8_030_000_000, rope_freq_base=500000.0,
    ),
    "Llama-3 70B": ModelArch(
        name="Llama-3 70B", architecture="llama",
        n_layer=80, n_embd=8192, n_head=64, n_head_kv=8,
        training_ctx=8192, params=70_000_000_000, rope_freq_base=500000.0,
    ),
    "Llama-3.1 8B": ModelArch(
        name="Llama-3.1 8B", architecture="llama",
        n_layer=32, n_embd=4096, n_head=32, n_head_kv=8,
        training_ctx=131072, params=8_030_000_000, rope_freq_base=500000.0,
    ),
    "Llama-3.1 70B": ModelArch(
        name="Llama-3.1 70B", architecture="llama",
        n_layer=80, n_embd=8192, n_head=64, n_head_kv=8,
        training_ctx=131072, params=70_000_000_000, rope_freq_base=500000.0,
    ),
    "Mistral 7B": ModelArch(
        name="Mistral 7B", architecture="llama",
        n_layer=32, n_embd=4096, n_head=32, n_head_kv=8,
        training_ctx=32768, params=7_240_000_000, rope_freq_base=10000.0,
    ),
    "Qwen2 7B": ModelArch(
        name="Qwen2 7B", architecture="qwen2",
        n_layer=28, n_embd=3584, n_head=28, n_head_kv=4,
        training_ctx=32768, params=7_620_000_000, rope_freq_base=1000000.0,
    ),
    "Qwen2 72B": ModelArch(
        name="Qwen2 72B", architecture="qwen2",
        n_layer=80, n_embd=8192, n_head=64, n_head_kv=8,
        training_ctx=131072, params=72_700_000_000, rope_freq_base=1000000.0,
    ),
    "DeepSeek-V3 (MoE)": ModelArch(
        name="DeepSeek-V3", architecture="deepseek2",
        n_layer=61, n_embd=7168, n_head=128, n_head_kv=128,
        training_ctx=131072, params=671_000_000_000, rope_freq_base=10000.0,
        n_expert=256, n_expert_used=8, n_mtp=1,
    ),
    "Phi-3 mini 3.8B": ModelArch(
        name="Phi-3 mini", architecture="phi3",
        n_layer=32, n_embd=3072, n_head=32, n_head_kv=32,
        training_ctx=131072, params=3_800_000_000, rope_freq_base=10000.0,
    ),
    "Gemma-2 9B": ModelArch(
        name="Gemma-2 9B", architecture="gemma2",
        n_layer=42, n_embd=3584, n_head=16, n_head_kv=8,
        training_ctx=8192, params=9_240_000_000, rope_freq_base=10000.0,
    ),
    # --- Current-gen hybrid-attention models (qwen3_5 / qwen3_5_moe). ---
    # These interleave linear/recurrent (Gated DeltaNet) layers with full-
    # attention layers every 4th position (layer_types = [L,L,L,F]×N). Only the
    # full-attention layers + MTP heads carry O(n_ctx) KV; n_full_attn_layers
    # encodes that count. head_dim is explicit (256) and ≠ n_embd/n_head.
    "Qwen3.6 27B": ModelArch(
        name="Qwen3.6 27B", architecture="qwen3_5",
        n_layer=64, n_embd=5120, n_head=24, n_head_kv=4,
        head_dim=256, n_full_attn_layers=16, full_attention_interval=4,
        training_ctx=262144, params=27_000_000_000, rope_freq_base=10000000.0,
        n_mtp=1,
    ),
    "Qwen3.6 35B-A3B (MoE)": ModelArch(
        name="Qwen3.6 35B-A3B", architecture="qwen3_5_moe",
        n_layer=40, n_embd=2048, n_head=16, n_head_kv=2,
        head_dim=256, n_full_attn_layers=10, full_attention_interval=4,
        training_ctx=262144, params=35_000_000_000, rope_freq_base=10000000.0,
        n_expert=256, n_expert_used=8, n_mtp=1,
    ),
    "Ornith 1.0 35B": ModelArch(
        name="Ornith 1.0 35B", architecture="qwen3_5_moe",
        n_layer=40, n_embd=2048, n_head=16, n_head_kv=2,
        head_dim=256, n_full_attn_layers=10, full_attention_interval=4,
        training_ctx=262144, params=35_000_000_000, rope_freq_base=10000000.0,
        n_expert=256, n_expert_used=8,
        # Ornith ships without MTP by default; leave n_mtp=0 unless the GGUF
        # sets it.
    ),
}

PRESET_NAMES = list(PRESETS.keys())