Block-wise attention forward pass without the full score matrixCodingImplement scaled dot-product attention that processes queries and keys in blocks with a running softmax normalization, so the quadratic score matrix is never materialized…
Hugging Face technical phone screen interview
Round 2 of 6 · 2 questions reported
- Length
- 60 min
- Format
- Live coding
- They grade
- Medium-level coding, plus ML or systems depth depending on the role
- Heads-up
- Only one of the three guides lists this round. ML roles reported on 1point3acres were asked to write blocked attention without the full score matrix.
Questions Hugging Face asks in this round 2
- Speculative decoding for faster generationCodingImplement speculative decoding: a small draft model proposes tokens and the large model verifies them, accepting matching runs to save forward passes. Reported on 1point3acres'…