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

min_p_sampling

def min_p_sampling[dtype: DType, out_idx_type: DType, //, _test_sort: Bool = False](min_ps: TileTensor[dtype, Storage=min_ps.Storage, address_space=min_ps.address_space, linear_idx_type=min_ps.linear_idx_type], input_logits: TileTensor[dtype, Storage=input_logits.Storage, address_space=input_logits.address_space, linear_idx_type=input_logits.linear_idx_type], out_token_ids: TileTensor[out_idx_type, Storage=out_token_ids.Storage, address_space=out_token_ids.address_space, linear_idx_type=out_token_ids.linear_idx_type], temperature: Scalar[dtype] = 1)

Naive CPU implementation of Min-P sampling for token selection. This function applies temperature scaling, softmax, a merge sort, and then samples tokens based on the calculated probability threshold (Min-P).

Parameters:

dtype (DType): Element type of input_logits, min_ps, and temperature (inferred).
out_idx_type (DType): Element type of the out_token_ids tensor (inferred).
_test_sort (Bool): When true, copies the sorted probabilities back into input_logits to verify descending order (defaults to false).