Softmax with Temperature

Every LLM turns raw logits into a probability distribution over the vocabulary with softmax. A temperature $T$ controls how sharp that distribution is:

Implement softmax(logits, temperature=1.0) returning a NumPy array of probabilities.

Requirements

• Must be numerically stable — logits like [1000, 1001] must not overflow.
• Lower temperature → sharper (more confident) distribution; higher → flatter.

Example

softmax(np.array([1.0, 2.0, 3.0]))
# array([0.0900, 0.2447, 0.6652])