{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-26-009",
    "source_story_id": "tmp-story-hyscale-sparse-selection",
    "edition_id": "mp-2026-10-26-morning-0108",
    "edition_url": "https://themachinepress.com/edition/2026-10-26",
    "position": 9,
    "story_type": "dispatch",
    "section": "infrastructure",
    "editorial_classification": "editorial",
    "headline": "Sparse Attention Stops Materializing the Score Table",
    "slug": "sparse-attention-stops-materializing-the-score-table",
    "dek": "HyScale finds exact top-k indices on the fly and replays selection, reporting a 2.07-times training speedup for million-token GLM-5.1 runs.",
    "summary": "HyScale finds exact top-k indices on the fly and replays selection, reporting a 2.07-times training speedup for million-token GLM-5.1 runs.",
    "body_text": "Learned sparse attention can still score every preceding token before reading only a small selected set, making selection itself the bottleneck. HyScale treats the result as indices rather than an activation tensor. HyTopK finds each query's exact top-k without materializing the full score table, and selection replay reuses the forward-pass decision during recomputation. The authors report a 2.07-times GLM-5.1 training speedup at one million tokens on 512 H20 GPUs, plus layer-level gains on two other systems. Those numbers depend on the named hardware, models and baselines; the contribution is a system design for exact selection reuse, not a universal doubling of long-context training speed.",
    "why_it_matters": "HyScale finds exact top-k indices on the fly and replays selection, reporting a 2.07-times training speedup for million-token GLM-5.1 runs.",
    "limitations": [
      "Learned sparse attention can still score every preceding token before reading only a small selected set, making selection itself the bottleneck.",
      "Those numbers depend on the named hardware, models and baselines; the contribution is a system design for exact selection reuse, not a universal doubling of long-context training speed."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-26-009/sparse-attention-stops-materializing-the-score-table",
    "json_url": "https://themachinepress.com/story/mp-2026-10-26-009.json",
    "first_published_at": "2026-10-26T09:00:00.000-04:00",
    "modified_at": "2026-10-26T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-26-009-001",
        "text": "HyScale finds exact top-k indices on the fly and replays selection, reporting a 2.07-times training speedup for million-token GLM-5.1 runs.",
        "source_ids": [
          "source-2026-10-26-009"
        ],
        "qualification": "Learned sparse attention can still score every preceding token before reading only a small selected set, making selection itself the bottleneck."
      }
    ],
    "source_ids": [
      "source-2026-10-26-009"
    ],
    "tags": [
      "sparse attention",
      "long context",
      "training systems",
      "GPUs"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-26-009",
      "title": "arXiv preprint 2610.32057, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.32057v1",
      "canonical_url": "https://arxiv.org/abs/2610.32057v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-22T21:54:02.000-04:00",
      "accessed_at": "2026-10-26T08:26:57.002-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-26-009-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Sparse Attention Stops Materializing the Score Table",
    "publisher": "The Machine Press",
    "published_at": "2026-10-26T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-26-009/sparse-attention-stops-materializing-the-score-table"
  }
}