Top-K Token Decoding Without PyTorch | Dataford Interview Questions

Decode tokens from LLM using top K sampling without using PyTorch

Asked in the Technical interview (ML coding) stage. No documentation was allowed; candidates were advised to brush up on NumPy.

Implement decode_tokens(logits, k, random_values). logits is a list of token-logit rows, k is the number of candidates, and random_values contains one deterministic value in [0, 1) per row. For each row, return a token ID sampled from its top k logits using softmax probabilities. Break equal-logit ties by lower token ID.

Constraints

1 <= len(logits) <= 10^4
1 <= len(logits[i]) <= 10^5
1 <= k <= len(logits[i])
len(random_values) == len(logits)
0 <= random_values[i] < 1
All logits are finite numbers

Function Signature

def decode_tokens(logits, k, random_values):

Your question is Top-K Token Decoding Without PyTorch. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem

Decode tokens from LLM using top K sampling without using PyTorch

Asked in the Technical interview (ML coding) stage. No documentation was allowed; candidates were advised to brush up on NumPy.

Constraints

1 <= len(logits) <= 10^4
1 <= len(logits[i]) <= 10^5
1 <= k <= len(logits[i])
len(random_values) == len(logits)
0 <= random_values[i] < 1
All logits are finite numbers

Function Signature

def decode_tokens(logits, k, random_values):

Your question is Top-K Token Decoding Without PyTorch. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.