For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

sliced_add

def sliced_add[dtype: DType, //, target: StringSlice[StaticConstantOrigin]](c: TileTensor[dtype, Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type, element_size=c.element_size], a: TileTensor[dtype, Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type, element_size=a.element_size], b: TileTensor[dtype, Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type, element_size=b.element_size], lora_end_idx: TileTensor[DType.int64, Storage=lora_end_idx.Storage, address_space=lora_end_idx.address_space, linear_idx_type=lora_end_idx.linear_idx_type, element_size=lora_end_idx.element_size], ctx: DeviceContext)

Adds tensors a and b element-wise for rows < lora_end_idx, otherwise copies a.

This is used for LoRA where only some sequences have LoRA applied. For rows in [0, lora_end_idx): c = a + b For rows in [lora_end_idx, batch_seq_len): c = a

Args:

c (TileTensor[dtype, Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type, element_size=c.element_size]): Output tensor.
a (TileTensor[dtype, Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type, element_size=a.element_size]): First input tensor.
b (TileTensor[dtype, Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type, element_size=b.element_size]): Second input tensor.
lora_end_idx (TileTensor[DType.int64, Storage=lora_end_idx.Storage, address_space=lora_end_idx.address_space, linear_idx_type=lora_end_idx.linear_idx_type, element_size=lora_end_idx.element_size]): Scalar tensor with end index of LoRA token portion (rows to apply add).
ctx (DeviceContext): Device context for GPU operations.