sparse_repro_full_gpu.json

{
  "dense": {
    "loss": 3.8660676462056927,
    "ppl": 47.7542298479057,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": null,
    "config": {
      "name": "dense",
      "mode": "dense",
      "keep_frac": 1.0,
      "block_q": 8,
      "block_k": 64
    }
  },
  "k_only_tight": {
    "loss": 3.8801209951294755,
    "ppl": 48.43007451999271,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.7943548387096775,
    "config": {
      "name": "k_only_tight",
      "mode": "k_only",
      "keep_frac": 0.15,
      "block_q": 8,
      "block_k": 64
    }
  },
  "tile_sparse_tight": {
    "loss": 4.448680234276347,
    "ppl": 85.5140110354067,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.0,
    "config": {
      "name": "tile_sparse_tight",
      "mode": "tile_sparse",
      "keep_frac": 0.15,
      "block_q": 8,
      "block_k": 64
    }
  },
  "k_only_40pct_cut": {
    "loss": 3.8632751820537607,
    "ppl": 47.621063890068136,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.84651284348865,
    "config": {
      "name": "k_only_40pct_cut",
      "mode": "k_only",
      "keep_frac": 0.592,
      "block_q": 8,
      "block_k": 64
    }
  },
  "tile_sparse_40pct_cut": {
    "loss": 3.939278332578918,
    "ppl": 51.38150755741357,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.4973118279569892,
    "config": {
      "name": "tile_sparse_40pct_cut",
      "mode": "tile_sparse",
      "keep_frac": 0.592,
      "block_q": 8,
      "block_k": 64
    }
  },
  "tile_sparse_tight_matched_compute": {
    "loss": 3.9356542561932355,
    "ppl": 51.19563406252316,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.6209677419354838,
    "config": {
      "name": "tile_sparse_tight_matched_compute",
      "mode": "tile_sparse",
      "keep_frac": 0.65,
      "block_q": 8,
      "block_k": 64
    }
  },
  "tile_sparse_40pct_cut_matched_compute": {
    "loss": 3.9356542561932355,
    "ppl": 51.19563406252316,
    "n_tokens": 2897,
    "avg_tiles_touched_per_qblock": 1.6209677419354838,
    "config": {
      "name": "tile_sparse_40pct_cut_matched_compute",
      "mode": "tile_sparse",
      "keep_frac": 0.73,
      "block_q": 8,
      "block_k": 64
    }
  },
  "claim1_tight_budget_comparison": {
    "k_only_excess_loss": 0.014053348923782849,
    "tile_sparse_excess_loss": 0.5826125880706545,
    "tile_sparse_relative_improvement_over_k_only_AT_MATCHED_KEEP_FRAC": -40.457206480135426,
    "k_only_avg_tiles_touched_per_qblock": 1.7943548387096775,
    "tile_sparse_avg_tiles_touched_per_qblock": 1.0,
    "k_only_real_compute_vs_tile_sparse_x": 1.7943548387096775,
    "note": "tile_sparse's keep_frac is a real tile-compute budget (avg_tiles_touched matches keep_frac*n_kb by construction). k_only's keep_frac is a per-query memory budget only -- k_only_avg_tiles_touched_per_qblock shows the REAL tile count a tiled kernel must compute once independent per-query top-K choices are unioned across an MTP query block; if that is larger than tile_sparse's, the two configs are not at matched real compute despite equal keep_frac, and the accuracy comparison above understates tile_sparse's advantage per unit of actual FLOPs."
  },
  "claim1_matched_compute_comparison": {
    "k_only_avg_tiles_touched_per_qblock": 1.7943548387096775,
    "tile_sparse_avg_tiles_touched_per_qblock": 1.6209677419354838,
    "k_only_excess_loss": 0.014053348923782849,
    "tile_sparse_excess_loss": 0.06958660998754285,
    "tile_sparse_relative_improvement_over_k_only_AT_MATCHED_REAL_COMPUTE": -3.951603376884752,
    "tile_sparse_accuracy_retention_vs_dense": 0.9327793418787504
  },
  "claim2_40pct_cut_comparison": {
    "k_only_excess_loss": -0.002792464151931906,
    "tile_sparse_excess_loss": 0.07321068637322536,
    "tile_sparse_relative_improvement_over_k_only_AT_MATCHED_KEEP_FRAC": NaN,
    "k_only_avg_tiles_touched_per_qblock": 1.84651284348865,
    "tile_sparse_avg_tiles_touched_per_qblock": 1.4973118279569892,
    "k_only_real_compute_vs_tile_sparse_x": 1.2332186315579494,
    "note": "tile_sparse's keep_frac is a real tile-compute budget (avg_tiles_touched matches keep_frac*n_kb by construction). k_only's keep_frac is a per-query memory budget only -- k_only_avg_tiles_touched_per_qblock shows the REAL tile count a tiled kernel must compute once independent per-query top-K choices are unioned across an MTP query block; if that is larger than tile_sparse's, the two configs are not at matched real compute despite equal keep_frac, and the accuracy comparison above understates tile_sparse's advantage per unit of actual FLOPs.",
    "tile_sparse_accuracy_retention_vs_dense": 0.9294049964288268
  },
  "claim2_matched_compute_comparison": {
    "k_only_avg_tiles_touched_per_qblock": 1.84651284348865,
    "tile_sparse_avg_tiles_touched_per_qblock": 1.6209677419354838,
    "k_only_excess_loss": -0.002792464151931906,
    "tile_sparse_excess_loss": 0.06958660998754285,
    "tile_sparse_relative_improvement_over_k_only_AT_MATCHED_REAL_COMPUTE": 25.919428218765443,
    "tile_sparse_accuracy_retention_vs_dense": 0.9327793418787504
  }
}