models:
  "meta-llama/Meta-Llama-3-70B-Instruct":
    hidden_size: 8192
    num_hidden_layers: 80
    num_attention_heads: 64
    num_key_value_heads: 8
    vocab_size: 128256
    max_position_embeddings: 8192
    intermediate_size: 28672

  "meta-llama/Meta-Llama-3-8B-Instruct":
    hidden_size: 4096
    num_hidden_layers: 32
    num_attention_heads: 32
    num_key_value_heads: 8
    vocab_size: 128256
    max_position_embeddings: 8192
    intermediate_size: 14336

  "mistralai/Mixtral-8x7B-Instruct-v0.1":
    hidden_size: 4096
    num_hidden_layers: 32
    num_attention_heads: 32
    num_key_value_heads: 8
    vocab_size: 32000
    max_position_embeddings: 32768
    intermediate_size: 14336
    notes:
      moe:
        num_local_experts: 8
        num_experts_per_tok: 2

  "mistralai/Mistral-7B-Instruct-v0.3":
    hidden_size: 4096
    num_hidden_layers: 32
    num_attention_heads: 32
    num_key_value_heads: 8
    vocab_size: 32768
    max_position_embeddings: 32768
    intermediate_size: 14336

  "google/gemma-7b":
    hidden_size: 3072
    num_hidden_layers: 28
    num_attention_heads: 16
    num_key_value_heads: 16
    vocab_size: 256000
    max_position_embeddings: 8192
    intermediate_size: 24576

  "openai/gpt-oss-20b":
    hidden_size: 2880
    num_hidden_layers: 24
    num_attention_heads: 64
    num_key_value_heads: 8
    vocab_size: 201088
    max_position_embeddings: 131072
    intermediate_size: 2880
    num_local_experts: 32
    num_experts_per_tok: 4

  "openai/gpt-oss-120b":
    hidden_size: 2880
    num_hidden_layers: 36
    num_attention_heads: 64
    num_key_value_heads: 8
    vocab_size: 201088
    max_position_embeddings: 131072
    intermediate_size: 2880
    num_local_experts: 128
    num_experts_per_tok: 4

  "Qwen/Qwen3-VL-235B-A22B-Thinking":
    text_config:
      hidden_size: 8192
      num_hidden_layers: 96
      num_attention_heads: 64
      num_key_value_heads: 8
      vocab_size: 151936
      max_position_embeddings: 262144
      intermediate_size: 24576
    torch_dtype: bfloat16
    notes:
      moe:
        num_experts: 512
        num_experts_per_tok: 10

  "Qwen/Qwen3-VL-235B-A22B-Instruct":
    text_config:
      hidden_size: 8192
      num_hidden_layers: 96
      num_attention_heads: 64
      num_key_value_heads: 8
      vocab_size: 151936
      max_position_embeddings: 262144
      intermediate_size: 24576
    torch_dtype: bfloat16
    notes:
      moe:
        num_experts: 512
        num_experts_per_tok: 10

  "Qwen/Qwen3-VL-30B-A3B-Thinking":
    text_config:
      hidden_size: 6144
      num_hidden_layers: 80
      num_attention_heads: 48
      num_key_value_heads: 8
      vocab_size: 151936
      max_position_embeddings: 262144
      intermediate_size: 16384
    torch_dtype: bfloat16

  "Qwen/Qwen3-VL-30B-A3B-Instruct":
    text_config:
      hidden_size: 6144
      num_hidden_layers: 80
      num_attention_heads: 48
      num_key_value_heads: 8
      vocab_size: 151936
      max_position_embeddings: 262144
      intermediate_size: 16384
    torch_dtype: bfloat16