inference-cache-config/mistral.json

{
  "mistralai/Mistral-7B-Instruct-v0.3": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 4,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 8,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 8,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 16,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "bf16"
    },
    {
      "batch_size": 32,
      "sequence_length": 4096,
      "num_cores": 8,
      "auto_cast_type": "bf16"
    }
  ],
  "mistralai/Mistral-7B-Instruct-v0.1": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "bf16"
    }
  ],
  "mistralai/Mistral-7B-Instruct-v0.2": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 2,
      "auto_cast_type": "bf16"
    }
  ],
  "mistralai/Mistral-Small-Instruct-2409": [
    {
      "batch_size": 1,
      "sequence_length": 4096,
      "num_cores": 12,
      "auto_cast_type": "bf16",
      "task": "text-generation"
    }
  ],
  "mistralai/Mistral-Small-Instruct-2409": [
    {
      "batch_size": 4,
      "sequence_length": 4096,
      "num_cores": 12,
      "auto_cast_type": "bf16",
      "task": "text-generation"
    }
  ]
}