|
|
@@ -84,7 +84,7 @@ namespace LLama.Web.Common |
|
|
/// <summary> |
|
|
/// <summary> |
|
|
/// how split tensors should be distributed across GPUs |
|
|
/// how split tensors should be distributed across GPUs |
|
|
/// </summary> |
|
|
/// </summary> |
|
|
public nint TensorSplits { get; set; } |
|
|
|
|
|
|
|
|
public float[] TensorSplits { get; set; } |
|
|
|
|
|
|
|
|
/// <summary> |
|
|
/// <summary> |
|
|
/// Grouped-Query Attention |
|
|
/// Grouped-Query Attention |
|
|
|