qnn/config.json

{
  "model_parameters": {
    "n_layers": 16,
    "vocab_size": 128256,
    "embed_dim": 2048,
    "ffn_hidden_dim": 8192,
    "head_dim": 64,
    "n_kv_heads": 8,
    "rope_theta": 500000.0,
    "rms_norm_eps": 1e-05,
    "attention_mask_value": -100000.0,
    "tie_embedding": true
  },
  "qnn_parameters": {
    "n_hvx_threads": 4
  },
  "graphs": [
    {
      "type": "transformers",
      "start_layer_id": 0,
      "end_layer_id": 16,
      "batch_size": 1,
      "cache_size": 1920,
      "context_size": 2048,
      "graph_name": "batch_1",
      "model_path": "llama3_2_1b_0.bin",
      "kv_path_format": "kv/layer_{layer_id}_{kv_type}_{head_id}.raw",
      "kv_size": 13,
      "x_name": "x",
      "out_name": "out"
    },
    {
      "type": "transformers",
      "start_layer_id": 0,
      "end_layer_id": 16,
      "batch_size": 128,
      "cache_size": 1920,
      "context_size": 2048,
      "graph_name": "batch_128",
      "model_path": "llama3_2_1b_0.bin",
      "kv_path_format": "kv/layer_{layer_id}_{kv_type}_{head_id}.raw",
      "kv_size": 13,
      "x_name": "x",
      "out_name": "out"
    }
  ],
  "embeddings": [
    {
      "graph_name": "batch_1",
      "model_path": "lm_head.bin",
      "batch_size": 1,
      "x_name": "x",
      "out_name": "logits"
    },
    {
      "graph_name": "batch_128",
      "model_path": "lm_head.bin",
      "batch_size": 128,
      "x_name": "x",
      "out_name": "logits"
    }
  ]
}