Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C
2.7 KB
Raw Blame
{
  "architectures": [
    "MiMoV2ForCausalLM"
  ],
  "model_type": "mimo_v2",
  "vocab_size": 320,
  "hidden_size": 256,
  "intermediate_size": 256,
  "num_hidden_layers": 6,
  "hybrid_layer_pattern": [
    0,
    1,
    1,
    0,
    1,
    1
  ],
  "moe_layer_freq": [
    0,
    1,
    1,
    1,
    1,
    1
  ],
  "num_attention_heads": 4,
  "num_key_value_heads": 2,
  "head_dim": 48,
  "v_head_dim": 32,
  "swa_num_attention_heads": 4,
  "swa_num_key_value_heads": 4,
  "swa_head_dim": 48,
  "swa_v_head_dim": 32,
  "sliding_window": 8,
  "sliding_window_size": 8,
  "attention_chunk_size": 8,
  "add_full_attention_sink_bias": false,
  "add_swa_attention_sink_bias": true,
  "attention_value_scale": 0.707,
  "attention_projection_layout": "fused_qkv",
  "attention_bias": false,
  "partial_rotary_factor": 0.334,
  "rope_theta": 10000000.0,
  "swa_rope_theta": 10000.0,
  "rope_parameters": {
    "partial_rotary_factor": 0.334,
    "rope_theta": 10000000.0,
    "rope_type": "default",
    "type": "default"
  },
  "n_routed_experts": 16,
  "num_experts_per_tok": 4,
  "moe_intermediate_size": 64,
  "n_shared_experts": null,
  "n_group": 1,
  "topk_group": 1,
  "topk_method": "noaux_tc",
  "scoring_func": "sigmoid",
  "norm_topk_prob": true,
  "routed_scaling_factor": null,
  "moe_router_dtype": "bfloat16",
  "hidden_act": "silu",
  "layernorm_epsilon": 1e-06,
  "max_position_embeddings": 256,
  "tie_word_embeddings": false,
  "bos_token_id": null,
  "eos_token_id": 1,
  "pad_token_id": 0,
  "dtype": "bfloat16",
  "quantization_config": {
    "activation_scheme": "dynamic",
    "fmt": "e4m3",
    "quant_method": "fp8",
    "store_dtype": "mxfp4",
    "mxfp4_block_size": 32,
    "weight_block_size": [
      128,
      128
    ],
    "ignored_layers": [
      "model.layers.0.self_attn.o_proj",
      "model.layers.1.self_attn.o_proj",
      "model.layers.2.self_attn.o_proj",
      "model.layers.3.self_attn.o_proj",
      "model.layers.4.self_attn.o_proj",
      "model.layers.5.self_attn.o_proj"
    ]
  },
  "vision_config": {
    "depth": 6,
    "hidden_size": 64,
    "intermediate_size": 96,
    "num_heads": 4,
    "num_key_value_heads": 2,
    "qk_channels": 16,
    "out_hidden_size": 256,
    "patch_size": 16,
    "spatial_patch_size": 16,
    "temporal_patch_size": 2,
    "spatial_merge_size": 2,
    "in_chans": 3,
    "hidden_act": "silu",
    "fullatt_block_indexes": [
      0,
      5
    ],
    "vit_window_attn_types": [
      -1,
      0,
      1,
      1,
      0,
      -1
    ],
    "visual_token_window_size": 5,
    "use_sink": true,
    "window_size": 128,
    "tokens_per_second": 2
  },
  "image_token_id": 300,
  "vision_start_token_id": 301,
  "vision_end_token_id": 302,
  "vision_model_type": "mimovl"
}