Add challenge 88: Prefix-Cached Attention (Medium)

[claude]

Summary

• Adds challenge 88: Prefix-Cached Attention (Medium difficulty)
• Models the chunked-prefill attention pattern used in LLM inference systems (vLLM, TensorRT-LLM): new query tokens attend to a full KV cache prefix (bidirectional) plus causally to each other (lower-triangular)
• Teaches solvers about block-causal attention masks, online softmax for mixed causal/non-causal patterns, and memory access trade-offs when tiling over a long KV cache

Challenge description

Given:

• Q shape (num_heads, new_len, head_dim) — queries for a chunk of new tokens
• K, V shape (num_heads, cache_len + new_len, head_dim) — packed key/value buffer (cache prefix followed by new tokens)

Compute scaled dot-product attention where query token i (absolute position cache_len + i) attends to key j iff j ≤ cache_len + i. This gives full access to the cached prefix and causal access within the new chunk.

Validation

• Validated with run_challenge.py --action run against the live T4 GPU platform — all tests pass
• Pre-commit lint passes (black, isort, flake8, clang-format)

Test plan

• All functional tests pass (10 cases: edge cases, zeros, cache_len=0, decode step new_len=1, power-of-2 sizes, non-power-of-2 sizes, realistic dimensions)
• Performance test: num_heads=32, cache_len=1024, new_len=512, head_dim=128 — fits well within 16 GB VRAM (≈64 MB per test)
• All 6 starter files present and lint-clean

🤖 Generated with Claude Code