Why architecture matters here
Position handling is where long-context models succeed or fail. Train at 4k tokens, deploy at 128k without a plan, and quality collapses. RoPE's rotational form is what allows the extension techniques to work — but naive extension (just letting positions go higher) blows up quickly.
The architecture matters because YaRN and NTK-aware scaling change the frequency base per dimension in specific ways, and kernel fusion must reflect those changes. Getting scaling wrong at serving time silently degrades outputs.
With the math and the extensions in your head, you can pick the right scaling and prove it in eval.
The architecture: every piece explained
The top strip is the core algorithm. Q / K vectors per head per token are the input. Frequency base theta_i = 10000^{-2i/d} defines the rotation frequency per dimension pair. Rotation matrix applies a 2D rotation to each dimension pair, encoded as cos/sin lookups. Attention proceeds normally; the rotations naturally produce relative-position sensitivity in the QK^T product.
The middle row is long-context extension. Position Interpolation (PI) scales positions by original_len / new_len so the frequencies stay in training range — simple, effective for modest extension. NTK-aware scaling adjusts the base of the frequencies to preserve high-frequency components while spreading low-frequency ones — better at preserving short-context quality. YaRN goes further by scaling per-dimension differently, applying temperature to attention, and requires a small fine-tune to recover full quality. Long-context fine-tune is the last mile — a few billion tokens on long-context data restores task performance.
The lower rows are the practical surface. Kernel integration means fused RoPE inside attention (Flash Attention + RoPE). Eval uses needle-in-haystack and long-document QA. Serving caches scaled positions and documents the max supported context per model.
End-to-end flow
End-to-end: a team extends a 4k-training Llama to 128k context. They apply YaRN with per-dim scaling and attention temperature. They fine-tune on 500M tokens of long documents. They fuse the modified RoPE into their attention kernel. Eval: needle-in-haystack recall at 128k = 0.92 (vs 0.35 without extension). Perplexity in short context flat. Ship. Serving documents max context 128k; users can prompt to that limit without surprise. Metrics show long-context requests running with expected latency thanks to Flash Attention + RoPE fusion.