inference-cache-config/llama-variants.json

{
   "defog/sqlcoder-7b-2": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
  "m-a-p/OpenCodeInterpreter-DS-6.7B": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
  "ibm/labradorite-13b": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 8,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "fp16"
    }
  ],
  "abacusai/Smaug-72B-v0.1": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 24,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 24,
      "auto_cast_type": "fp16"
    }
  ],
  "gorilla-llm/gorilla-openfunctions-v2": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
   "m-a-p/ChatMusician": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
   "LargeWorldModel/LWM-Text-Chat-1M": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
   "princeton-nlp/Sheared-LLaMA-1.3B": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "fp16"
    }
  ],
  "01-ai/Yi-34B-200K": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 24,
      "auto_cast_type": "fp16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 24,
      "auto_cast_type": "fp16"
    }
  ]
}