Serving a model with a very long context window, well into the hundreds of thousands of tokens, is expensive specifically because self-attention's compute cost grows quadratically with sequence length, and one common workaround is to shard that long context across multiple GPUs or hosts rather than trying to process it on a single device. The existing standard approach for this, called Star Attention, handles the sharding by prepending a static, content-blind copy of the first block of context to every host, which works but wastes compute on a fixed anchor block that doesn't actually adapt to what's in the rest of the context. This new paper proposes Pulsar Attention as a replacement: instead of a static anchor, it uses two lightweight, content-aware pieces, a small attention-sink prefix that keeps the underlying softmax calculation numerically stable, and compact cross-block summaries built using a Max-IDF heuristic that specifically picks out chunks containing globally rare, and therefore likely more informationally important, tokens. The measured result is a genuinely large efficiency win: up to a 3.3x reduction in per-GPU compute during the first processing phase compared to Star Attention, while keeping an identical memory footprint for the key-value cache, and on long-context benchmarks up to 128,000 tokens it actually outperforms both Star Attention and standard dense attention in accuracy, not just efficiency. For teams building or operating infrastructure that serves long-context LLM workloads at scale, whether that's document analysis, large codebase understanding, or long conversational memory, this is directly relevant because the per-GPU compute savings compound across every request, and the accuracy improvement means this isn't purely a cost-versus-quality tradeoff, it's a case where the more efficient approach also happens to perform better. It's early-stage academic work rather than a shipped production system, but the technique is a genuine candidate for teams running their own distributed long-context inference to evaluate against their current sharding approach.