{
 "device": {
  "jax": "0.11.2",
  "backend": "cpu",
  "devices": [
   "cpu:0"
  ],
  "python": "3.14.8",
  "device_kind": "cpu",
  "gpu": "NVIDIA GeForce RTX 3090",
  "driver": "615.78.08",
  "gpu_memory_gb": 24.0,
  "torch": "2.14.1+cu130",
  "torch_cuda": "13.0"
 },
 "config": {
  "sources": [
   {
    "file": "cpu_A.json",
    "code": "final",
    "note": "all CPU paths, 1k/4k tokens, E8 top2, numerics vs fp32 dense",
    "config": {
     "device": "cpu",
     "suites": [
      "layer"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      1024,
      4096
     ],
     "experts": [
      8
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": null,
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "slab_elems": null,
     "check": true
    }
   },
   {
    "file": "cpu_B.json",
    "code": "final",
    "note": "16k tokens E8 top2",
    "config": {
     "device": "cpu",
     "suites": [
      "layer"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      16384
     ],
     "experts": [
      8
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": null,
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "slab_elems": null,
     "check": true
    }
   },
   {
    "file": "cpu_C.json",
    "code": "final",
    "note": "4k tokens E64 top2 (ragged_dot skipped: too slow)",
    "config": {
     "device": "cpu",
     "suites": [
      "layer"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      4096
     ],
     "experts": [
      64
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": null,
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 131072,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "slab_elems": null,
     "check": true
    }
   },
   {
    "file": "cpu_D.json",
    "code": "final",
    "note": "top_k 1/4, skew 0/0.9",
    "config": {
     "device": "cpu",
     "suites": [
      "layer"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      4096
     ],
     "experts": [
      8
     ],
     "top_k": [
      1,
      4
     ],
     "skew": [
      0.0,
      0.9
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": [
      "dropless:chunked:sorted",
      "dropless:blocked:sorted",
      "capacity:auto:sorted"
     ],
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "slab_elems": null,
     "check": true
    }
   },
   {
    "file": "cpu_E.json",
    "code": "final",
    "note": "bfloat16",
    "config": {
     "device": "cpu",
     "suites": [
      "layer"
     ],
     "dtypes": [
      "bfloat16"
     ],
     "tokens": [
      4096
     ],
     "experts": [
      8
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": [
      "dropless:chunked:sorted",
      "dropless:blocked:sorted",
      "capacity:auto:sorted",
      "dense:auto:sorted"
     ],
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "slab_elems": null,
     "check": true
    }
   },
   {
    "file": "cpu_chunk_v1.json",
    "code": "final-kernel",
    "note": "chunk sweep of the landed chunked kernel (fixed chunk)",
    "config": {
     "device": "cpu",
     "suites": [
      "chunk"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      4096,
      16384
     ],
     "experts": [
      8,
      64
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": null,
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      256,
      512
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "check": true
    }
   },
   {
    "file": "cpu_chunk_v2.json",
    "code": "rejected",
    "note": "chunk sweep of a variant with concatenated [Wg|Wu] (slower, more memory; not landed)",
    "config": {
     "device": "cpu",
     "suites": [
      "chunk"
     ],
     "dtypes": [
      "float32"
     ],
     "tokens": [
      1024,
      4096,
      16384
     ],
     "experts": [
      8,
      64
     ],
     "top_k": [
      2
     ],
     "skew": [
      0.0
     ],
     "d_model": 512,
     "d_hidden": 1024,
     "impls": null,
     "block_m": 64,
     "block_ms": [
      32,
      64,
      128
     ],
     "chunks": [
      128,
      256,
      512,
      1024
     ],
     "dense_limit": 524288,
     "ragged_limit": 262144,
     "maxtok_start": 32768,
     "maxtok_stop": 2097152,
     "check": true
    }
   }
  ]
 },
 "results": [
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 7.518226623535156,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.034751927996694576,
   "inner": 1,
   "median_ms": 29.592196995508857,
   "q1_ms": 28.95863800222287,
   "q3_ms": 29.80539499912993,
   "rel_err_out": 3.622434892596369e-08,
   "rel_err_grad_params": 8.803933099013328e-07,
   "rel_err_grad_x": 7.436895473179166e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 3.0723953247070312,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.13100317100179382,
   "inner": 1,
   "median_ms": 124.50293300207704,
   "q1_ms": 122.47849200139171,
   "q3_ms": 125.439308998466,
   "rel_err_out": 3.622434892596369e-08,
   "rel_err_grad_params": 8.803933099013328e-07,
   "rel_err_grad_x": 7.436895473179166e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 162.02371978759766,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.039450849006243516,
   "inner": 1,
   "median_ms": 80.40560300287325,
   "q1_ms": 79.34636900245096,
   "q3_ms": 81.84200549658271,
   "rel_err_out": 2.9113461235546828e-08,
   "rel_err_grad_params": 8.643502811389138e-07,
   "rel_err_grad_x": 6.642576977418919e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 154.07083892822266,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.12190630900295218,
   "inner": 1,
   "median_ms": 125.2352180017624,
   "q1_ms": 124.33601649900083,
   "q3_ms": 128.62963000225136,
   "rel_err_out": 2.9113461235546828e-08,
   "rel_err_grad_params": 8.643502811389138e-07,
   "rel_err_grad_x": 6.642576977418919e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 104.01582336425781,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.08920898799988208,
   "inner": 1,
   "median_ms": 107.92212199885398,
   "q1_ms": 106.34662150187069,
   "q3_ms": 109.56663850083714,
   "rel_err_out": 8.885152879540563e-07,
   "rel_err_grad_params": 1.0679636687015801e-06,
   "rel_err_grad_x": 9.970541072635397e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 176.05494689941406,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.222221048999927,
   "inner": 1,
   "median_ms": 245.48067500290927,
   "q1_ms": 242.72967550132307,
   "q3_ms": 247.4165579988039,
   "rel_err_out": 8.885152879540563e-07,
   "rel_err_grad_params": 1.0679636687015801e-06,
   "rel_err_grad_x": 9.970541072635397e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 25.033279418945312,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.016544090998650063,
   "inner": 4,
   "median_ms": 11.233725999773014,
   "q1_ms": 11.075157250161283,
   "q3_ms": 11.395415750484972
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 9.078323364257812,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.07476567300182069,
   "inner": 1,
   "median_ms": 77.22497500071768,
   "q1_ms": 75.4667535002227,
   "q3_ms": 79.57911400080775
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 80.04689025878906,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.048503525998967234,
   "inner": 1,
   "median_ms": 50.55977599840844,
   "q1_ms": 47.78074299974833,
   "q3_ms": 54.83068600005936,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 160.04701232910156,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.1848508370021591,
   "inner": 1,
   "median_ms": 168.43855600018287,
   "q1_ms": 167.17835700546857,
   "q3_ms": 174.51539300236618,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.067298889160156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.08509956299531041,
   "inner": 1,
   "median_ms": 78.180312004406,
   "q1_ms": 74.8595490003936,
   "q3_ms": 81.15370300220093,
   "rel_err_out": 3.643038377928492e-08,
   "rel_err_grad_params": 8.326905203527818e-07,
   "rel_err_grad_x": 7.494935373176036e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.409553527832031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.2804833179980051,
   "inner": 1,
   "median_ms": 277.6387470003101,
   "q1_ms": 275.14475799762295,
   "q3_ms": 282.67194949876284,
   "rel_err_out": 3.643038377928492e-08,
   "rel_err_grad_params": 8.326905203527818e-07,
   "rel_err_grad_x": 7.494935373176036e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 486.09403228759766,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.11516291099542286,
   "inner": 1,
   "median_ms": 151.2849520004238,
   "q1_ms": 150.32041749873315,
   "q3_ms": 153.34926750074374,
   "rel_err_out": 3.010823583447513e-08,
   "rel_err_grad_params": 8.138714448906908e-07,
   "rel_err_grad_x": 6.678746844065834e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 520.1257095336914,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.30851139599690214,
   "inner": 1,
   "median_ms": 339.37753299687756,
   "q1_ms": 335.48474800409167,
   "q3_ms": 346.6281159999198,
   "rel_err_out": 3.010823583447513e-08,
   "rel_err_grad_params": 8.138714448906908e-07,
   "rel_err_grad_x": 6.678746844065834e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 368.0626983642578,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.3955393680007546,
   "inner": 1,
   "median_ms": 434.12148299830733,
   "q1_ms": 432.8891110017139,
   "q3_ms": 444.87211749947164,
   "rel_err_out": 8.888168490111711e-07,
   "rel_err_grad_params": 1.0245490236982073e-06,
   "rel_err_grad_x": 9.971220245044604e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 864.0940093994141,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 1.08176582899614,
   "inner": 1,
   "median_ms": 1129.3402669980424,
   "q1_ms": 1119.824763998622,
   "q3_ms": 1165.5410314997425,
   "rel_err_out": 8.888168490111711e-07,
   "rel_err_grad_params": 1.0245490236982073e-06,
   "rel_err_grad_x": 9.971220245044604e-07
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 100.13288879394531,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.06093009599862853,
   "inner": 1,
   "median_ms": 63.51177099713823,
   "q1_ms": 62.3777600012545,
   "q3_ms": 68.42512149887625
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 260.2501983642578,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.18487081600324018,
   "inner": 1,
   "median_ms": 197.4341640016064,
   "q1_ms": 195.87169350052136,
   "q3_ms": 200.60530400223797
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 320.18751525878906,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.17521748199942522,
   "inner": 1,
   "median_ms": 186.50014699960593,
   "q1_ms": 182.98099449748406,
   "q3_ms": 187.35459900199203,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_A",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 896.1876373291016,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.4745331800004351,
   "inner": 1,
   "median_ms": 530.6009859996266,
   "q1_ms": 528.3921724949323,
   "q3_ms": 532.0897429992328,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.255531311035156,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.20783904200652614,
   "inner": 1,
   "median_ms": 192.1932819968788,
   "q1_ms": 191.8869729997823,
   "q3_ms": 194.0037829990615,
   "rel_err_out": 3.6346865519365234e-08,
   "rel_err_grad_params": 6.690686349193516e-07,
   "rel_err_grad_x": 7.493093606613383e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 65.75183868408203,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.6265772550032125,
   "inner": 1,
   "median_ms": 619.0481899975566,
   "q1_ms": 611.2032544988324,
   "q3_ms": 620.4193999983545,
   "rel_err_out": 3.6346865519365234e-08,
   "rel_err_grad_params": 6.690686349193516e-07,
   "rel_err_grad_x": 7.493093606613383e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 1782.3752822875977,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.34851564199925633,
   "inner": 1,
   "median_ms": 505.5638450066908,
   "q1_ms": 503.72283199976664,
   "q3_ms": 519.8205315027735,
   "rel_err_out": 2.987022463273257e-08,
   "rel_err_grad_params": 6.511336607787061e-07,
   "rel_err_grad_x": 6.699250489344965e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 1912.6889419555664,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.8863690340003814,
   "inner": 1,
   "median_ms": 1014.6151469962206,
   "q1_ms": 1009.9403630010784,
   "q3_ms": 1015.5006229979335,
   "rel_err_out": 2.987022463273257e-08,
   "rel_err_grad_params": 6.511336607787061e-07,
   "rel_err_grad_x": 6.699250489344965e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 1424.250259399414,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 1.672552320997056,
   "inner": 1,
   "median_ms": 1733.083336999698,
   "q1_ms": 1689.661824500945,
   "q3_ms": 1773.6377800029004,
   "rel_err_out": 8.895264845321679e-07,
   "rel_err_grad_params": 8.844662090549437e-07,
   "rel_err_grad_x": 9.97753076888933e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dropless",
   "impl": "ragged_dot",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 3584.437759399414,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 3.809876705003262,
   "inner": 1,
   "median_ms": 4206.578657998762,
   "q1_ms": 4143.519679004385,
   "q3_ms": 4239.152378002473,
   "rel_err_out": 8.895264845321679e-07,
   "rel_err_grad_params": 8.844662090549437e-07,
   "rel_err_grad_x": 9.97753076888933e-07
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 400.5313262939453,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.22794039099971997,
   "inner": 1,
   "median_ms": 261.9926729967119,
   "q1_ms": 257.96743250248255,
   "q3_ms": 265.6318530025601
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 1216.5001983642578,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.685470031996374,
   "inner": 1,
   "median_ms": 767.4295679971692,
   "q1_ms": 759.9184624959889,
   "q3_ms": 770.6717410037527
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 1280.750015258789,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.7288266960022156,
   "inner": 1,
   "median_ms": 842.0050639979308,
   "q1_ms": 834.7959849961626,
   "q3_ms": 849.8313600030087,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_B",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 3584.8751373291016,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 1.8891490009991685,
   "inner": 1,
   "median_ms": 2142.9531419998966,
   "q1_ms": 2122.8516530027264,
   "q3_ms": 2195.7585680029297,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 6.815834045410156,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.18509890399582218,
   "inner": 1,
   "median_ms": 173.1781790003879,
   "q1_ms": 172.8025844968215,
   "q3_ms": 174.918330500077,
   "rel_err_out": 3.696287004493774e-08,
   "rel_err_grad_params": 1.091342274739996e-06,
   "rel_err_grad_x": 8.502482506676492e-07
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 2.6908035278320312,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.8173136199984583,
   "inner": 1,
   "median_ms": 796.6608120041201,
   "q1_ms": 791.4946224991581,
   "q3_ms": 802.4821859980875,
   "rel_err_out": 3.696287004493774e-08,
   "rel_err_grad_params": 1.091342274739996e-06,
   "rel_err_grad_x": 8.502482506676492e-07
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 864.0940322875977,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.18145616500260076,
   "inner": 1,
   "median_ms": 280.256632999226,
   "q1_ms": 277.9315959996893,
   "q3_ms": 284.7494995003217,
   "rel_err_out": 2.9776251168037518e-08,
   "rel_err_grad_params": 1.0709739302952364e-06,
   "rel_err_grad_x": 7.731838856008419e-07
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 817.2040176391602,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.7138261270010844,
   "inner": 1,
   "median_ms": 831.6442919967812,
   "q1_ms": 826.5148660029809,
   "q3_ms": 838.3468635001918,
   "rel_err_out": 2.9776251168037518e-08,
   "rel_err_grad_params": 1.0709739302952364e-06,
   "rel_err_grad_x": 7.731838856008419e-07
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 100.13288879394531,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.059918606995779555,
   "inner": 1,
   "median_ms": 87.28792699548649,
   "q1_ms": 84.38269200269133,
   "q3_ms": 88.8655699964147
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 37.21894836425781,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.40747126800124533,
   "inner": 1,
   "median_ms": 491.91463500028476,
   "q1_ms": 489.05251199539634,
   "q3_ms": 497.59587349399226
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 2561.062515258789,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 1.840855663002003,
   "inner": 1,
   "median_ms": 2033.9283970024553,
   "q1_ms": 2032.5426510025864,
   "q3_ms": 2036.9231555014267,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_C",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 7169.015762329102,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 4.383147002001351,
   "inner": 1,
   "median_ms": 4962.793783000961,
   "q1_ms": 4945.769717498479,
   "q3_ms": 4974.681444495218,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 0.0,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.035804748535156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.05082278900226811,
   "inner": 1,
   "median_ms": 43.094542998005636,
   "q1_ms": 42.7133369994408,
   "q3_ms": 43.801003997941734,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.9319049105728285e-07,
   "rel_err_grad_x": 9.104503311908768e-08
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.206184387207031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.1909933870047098,
   "inner": 1,
   "median_ms": 175.99471100402297,
   "q1_ms": 174.88427149874042,
   "q3_ms": 176.5027019973786,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.9319049105728285e-07,
   "rel_err_grad_x": 9.104503311908768e-08
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 270.06278228759766,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.06007374500040896,
   "inner": 1,
   "median_ms": 115.48591300379485,
   "q1_ms": 113.81842950140708,
   "q3_ms": 117.622649500845,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.2033422707411313e-07,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 260.0943374633789,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.16472381499625044,
   "inner": 1,
   "median_ms": 190.07265500113135,
   "q1_ms": 186.4535415006685,
   "q3_ms": 192.39360550272977,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.2033422707411313e-07,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 50.06648254394531,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.030143347998091485,
   "inner": 1,
   "median_ms": 53.25621100200806,
   "q1_ms": 49.006379002094036,
   "q3_ms": 54.91263900330523
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.0,
   "max_expert_frac": 0.146240234375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 100.20722961425781,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.11188037499960046,
   "inner": 1,
   "median_ms": 114.1642239963403,
   "q1_ms": 111.45227449742379,
   "q3_ms": 116.14475950045744
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.035804748535156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.04842023600212997,
   "inner": 1,
   "median_ms": 41.67179199430393,
   "q1_ms": 41.60183499698178,
   "q3_ms": 41.80602149790502,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.347319809108445e-07,
   "rel_err_grad_x": 9.101745225544391e-08
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.206184387207031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.17983058500249172,
   "inner": 1,
   "median_ms": 170.41618999792263,
   "q1_ms": 169.80217449963675,
   "q3_ms": 172.17642949981382,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.347319809108445e-07,
   "rel_err_grad_x": 9.101745225544391e-08
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 270.06278228759766,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.056227510001917835,
   "inner": 1,
   "median_ms": 117.42531100026099,
   "q1_ms": 114.16007600200828,
   "q3_ms": 118.26769899926148,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.0929493151736697e-07,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 260.0943374633789,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.1641831520028063,
   "inner": 1,
   "median_ms": 183.29796299804002,
   "q1_ms": 181.05425050089252,
   "q3_ms": 185.1726935019542,
   "rel_err_out": 0.0,
   "rel_err_grad_params": 2.0929493151736697e-07,
   "rel_err_grad_x": 0.0
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 50.06648254394531,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.028633312998863403,
   "inner": 1,
   "median_ms": 34.87290300108725,
   "q1_ms": 30.934915994293988,
   "q3_ms": 50.0807305033959
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 1,
   "skew": 0.9,
   "max_expert_frac": 0.939208984375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 100.20722961425781,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.10386902300524525,
   "inner": 1,
   "median_ms": 109.59828799968818,
   "q1_ms": 107.67798099550419,
   "q3_ms": 110.12508950079791
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.130043029785156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.11419548199773999,
   "inner": 1,
   "median_ms": 106.35676799574867,
   "q1_ms": 105.18197100100224,
   "q3_ms": 106.63356149962055,
   "rel_err_out": 4.910403036421828e-08,
   "rel_err_grad_params": 5.556341312281158e-07,
   "rel_err_grad_x": 5.983950580064187e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.691047668457031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.35792605999449734,
   "inner": 1,
   "median_ms": 350.9891299981973,
   "q1_ms": 349.72827050296473,
   "q3_ms": 353.185363997909,
   "rel_err_out": 4.910403036421828e-08,
   "rel_err_grad_params": 5.556341312281158e-07,
   "rel_err_grad_x": 5.983950580064187e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 918.1877822875977,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.19140548699942883,
   "inner": 1,
   "median_ms": 276.3590159956948,
   "q1_ms": 272.50105399798485,
   "q3_ms": 280.6844805018045,
   "rel_err_out": 5.0458248764433676e-08,
   "rel_err_grad_params": 4.769874283499208e-07,
   "rel_err_grad_x": 5.434796952955984e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 984.2509536743164,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.4903914859969518,
   "inner": 1,
   "median_ms": 538.7859920010669,
   "q1_ms": 537.5439680028649,
   "q3_ms": 544.0149714995641,
   "rel_err_out": 5.0458248764433676e-08,
   "rel_err_grad_params": 4.769874283499208e-07,
   "rel_err_grad_x": 5.434796952955984e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 200.2657012939453,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.10285999500047183,
   "inner": 1,
   "median_ms": 128.55084300099406,
   "q1_ms": 121.94624149924493,
   "q3_ms": 145.44218400260434
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.0,
   "max_expert_frac": 0.12750244140625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 608.2501983642578,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.32557485600409564,
   "inner": 1,
   "median_ms": 369.7201149989269,
   "q1_ms": 368.1224080028187,
   "q3_ms": 372.2371155017754
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.130043029785156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.11543101000279421,
   "inner": 1,
   "median_ms": 104.50309199950425,
   "q1_ms": 104.44885900142253,
   "q3_ms": 104.69830099827959,
   "rel_err_out": 7.819331721318397e-09,
   "rel_err_grad_params": 6.457078373355385e-07,
   "rel_err_grad_x": 4.1879971707650814e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.691047668457031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.3651731360005215,
   "inner": 1,
   "median_ms": 347.86637499928474,
   "q1_ms": 346.60781450293143,
   "q3_ms": 349.97498750090017,
   "rel_err_out": 7.819331721318397e-09,
   "rel_err_grad_params": 6.457078373355385e-07,
   "rel_err_grad_x": 4.1879971707650814e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 918.1877822875977,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.19090229800349334,
   "inner": 1,
   "median_ms": 269.7674330047448,
   "q1_ms": 268.0828324992035,
   "q3_ms": 271.78008000191767,
   "rel_err_out": 7.996113723388197e-09,
   "rel_err_grad_params": 6.352205148415315e-07,
   "rel_err_grad_x": 3.7670320881274073e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 984.2509536743164,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.4875185400014743,
   "inner": 1,
   "median_ms": 537.8023099983693,
   "q1_ms": 532.5787059991853,
   "q3_ms": 544.4485245025135,
   "rel_err_out": 7.996113723388197e-09,
   "rel_err_grad_params": 6.352205148415315e-07,
   "rel_err_grad_x": 3.7670320881274073e-07
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 200.2657012939453,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.10450596900045639,
   "inner": 1,
   "median_ms": 120.3567779957666,
   "q1_ms": 120.12849549864768,
   "q3_ms": 131.8776465013798
  },
  {
   "code": "final",
   "run": "cpu_D",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 4,
   "skew": 0.9,
   "max_expert_frac": 0.2366943359375,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 608.2501983642578,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.318213457998354,
   "inner": 1,
   "median_ms": 370.8682430005865,
   "q1_ms": 365.23426699932315,
   "q3_ms": 372.33641449711286
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 40.065162658691406,
   "xla_arg_mb": 52.015625,
   "xla_out_mb": 4.0,
   "compile_s": 0.07624156399833737,
   "inner": 1,
   "median_ms": 78.20294299745001,
   "q1_ms": 77.3424704966601,
   "q3_ms": 78.47377500365837,
   "rel_err_out": 0.019680057815701008,
   "rel_err_grad_params": 0.03446605749880788,
   "rel_err_grad_x": 0.028784485418086102
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 18.40955352783203,
   "xla_arg_mb": 148.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.2855342330003623,
   "inner": 1,
   "median_ms": 264.26139999966836,
   "q1_ms": 261.6576754953712,
   "q3_ms": 266.540676002478,
   "rel_err_out": 0.019680057815701008,
   "rel_err_grad_params": 0.03446605749880788,
   "rel_err_grad_x": 0.028784485418086102
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 486.09403228759766,
   "xla_arg_mb": 52.015625,
   "xla_out_mb": 4.0,
   "compile_s": 0.11650494200148387,
   "inner": 1,
   "median_ms": 182.65382800018415,
   "q1_ms": 181.55242050488596,
   "q3_ms": 183.9273264995427,
   "rel_err_out": 0.019462141239644078,
   "rel_err_grad_params": 0.03613020020045399,
   "rel_err_grad_x": 0.03002466229453848
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dropless",
   "impl": "blocked",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 520.2972793579102,
   "xla_arg_mb": 148.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.2780543960034265,
   "inner": 1,
   "median_ms": 314.0101790049812,
   "q1_ms": 313.86219849446206,
   "q3_ms": 320.8559404993139,
   "rel_err_out": 0.019462141239644078,
   "rel_err_grad_params": 0.03613020020045399,
   "rel_err_grad_x": 0.03002466229453848
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 66.11726379394531,
   "xla_arg_mb": 52.015625,
   "xla_out_mb": 4.0,
   "compile_s": 0.058156530001724605,
   "inner": 1,
   "median_ms": 60.64490599965211,
   "q1_ms": 60.601108001719695,
   "q3_ms": 61.26609600323718
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "capacity",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 188.3283233642578,
   "xla_arg_mb": 148.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.17622143299377058,
   "inner": 1,
   "median_ms": 184.6230529990862,
   "q1_ms": 181.421136498102,
   "q3_ms": 186.2596275022952
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 176.25001525878906,
   "xla_arg_mb": 52.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.1588482950028265,
   "inner": 1,
   "median_ms": 175.06805100128986,
   "q1_ms": 174.6410529995046,
   "q3_ms": 177.55269500412396,
   "rel_err_out": 0.02007995096813788,
   "rel_err_grad_params": 0.034566805832186985,
   "rel_err_grad_x": 0.02912048349213149
  },
  {
   "code": "final",
   "run": "cpu_E",
   "suite": "layer",
   "routing": "dense",
   "impl": "auto",
   "dispatch": "sorted",
   "dtype": "bfloat16",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "skew": 0.0,
   "max_expert_frac": 0.1356201171875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 608.1016998291016,
   "xla_arg_mb": 148.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.47491592499864055,
   "inner": 1,
   "median_ms": 507.31106299645035,
   "q1_ms": 504.56552050309256,
   "q3_ms": 511.20578900008695,
   "rel_err_out": 0.02007995096813788,
   "rel_err_grad_params": 0.034566805832186985,
   "rel_err_grad_x": 0.02912048349213149
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 7.565589904785156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.08256722999794874,
   "inner": 1,
   "median_ms": 72.31805499759503,
   "q1_ms": 71.78520550223766,
   "q3_ms": 72.73568299933686,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 3.4088211059570312,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.25239926100039156,
   "inner": 1,
   "median_ms": 244.1966800033697,
   "q1_ms": 238.3479534983053,
   "q3_ms": 248.3584850015177,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.067298889160156,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.07315425199340098,
   "inner": 1,
   "median_ms": 63.426116001210175,
   "q1_ms": 63.06453599972883,
   "q3_ms": 63.793539000471355,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.409553527832031,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.24891398000181653,
   "inner": 1,
   "median_ms": 242.49441100255353,
   "q1_ms": 240.33303000032902,
   "q3_ms": 243.5905050006113,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 7.566322326660156,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.19026808100170456,
   "inner": 1,
   "median_ms": 173.89001000265125,
   "q1_ms": 173.56613049923908,
   "q3_ms": 174.1505439968023,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 4.190803527832031,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.8158665539958747,
   "inner": 1,
   "median_ms": 815.6506860032096,
   "q1_ms": 811.7816755002423,
   "q3_ms": 817.5440220002201,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.068031311035156,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.22306175400444772,
   "inner": 1,
   "median_ms": 210.53273500001524,
   "q1_ms": 210.19215900014387,
   "q3_ms": 211.80148800340248,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 5.191535949707031,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.9733634350050124,
   "inner": 1,
   "median_ms": 971.614968002541,
   "q1_ms": 968.3541614977003,
   "q3_ms": 977.1380415004387,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 7.754554748535156,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.25790061799489195,
   "inner": 1,
   "median_ms": 242.58556299901102,
   "q1_ms": 241.7051084994455,
   "q3_ms": 243.33514550016844,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 65.75281524658203,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.7159169079968706,
   "inner": 1,
   "median_ms": 713.3391090028454,
   "q1_ms": 710.8025659981649,
   "q3_ms": 714.25619799993,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.255531311035156,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.20822272300574696,
   "inner": 1,
   "median_ms": 195.0889469953836,
   "q1_ms": 194.02403649655753,
   "q3_ms": 195.62042500183452,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 65.75183868408203,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.6166178370040143,
   "inner": 1,
   "median_ms": 622.5881530044717,
   "q1_ms": 620.1841120018798,
   "q3_ms": 625.833309499285,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 7.755287170410156,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.3613681179995183,
   "inner": 1,
   "median_ms": 348.71176700107753,
   "q1_ms": 347.74598500007414,
   "q3_ms": 349.70621400134405,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 7.879768371582031,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.2952177109982586,
   "inner": 1,
   "median_ms": 1305.583775996638,
   "q1_ms": 1297.4154295006883,
   "q3_ms": 1309.237101500912,
   "chunk": 256
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 9.256263732910156,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.35635775200353237,
   "inner": 1,
   "median_ms": 342.04202399996575,
   "q1_ms": 341.02299100050004,
   "q3_ms": 343.2001509972906,
   "chunk": 512
  },
  {
   "code": "final-kernel",
   "run": "cpu_chunk_v1",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 8.879768371582031,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.3603021730014007,
   "inner": 1,
   "median_ms": 1356.7087759947754,
   "q1_ms": 1348.9534280015505,
   "q3_ms": 1359.212565501366,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 39.267486572265625,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.042025137998280115,
   "inner": 1,
   "median_ms": 46.70879799959948,
   "q1_ms": 44.363386998156784,
   "q3_ms": 47.366623002744745,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 64.32213592529297,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.2562185920032789,
   "inner": 1,
   "median_ms": 170.13734400097746,
   "q1_ms": 166.21820950240362,
   "q3_ms": 180.86971299999277,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 40.518218994140625,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.04116075199999614,
   "inner": 1,
   "median_ms": 34.183021998615004,
   "q1_ms": 33.68053999656695,
   "q3_ms": 34.85781749986927,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 64.57238006591797,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.14528596800664673,
   "inner": 1,
   "median_ms": 163.13701999752084,
   "q1_ms": 161.59567900103866,
   "q3_ms": 164.5534650015179,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 43.020172119140625,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.04838820599979954,
   "inner": 1,
   "median_ms": 37.35087400127668,
   "q1_ms": 36.711739499878604,
   "q3_ms": 38.97282300022198,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 65.07335662841797,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.16473429100005887,
   "inner": 1,
   "median_ms": 176.52194300171686,
   "q1_ms": 175.5062660013209,
   "q3_ms": 181.9057385000633,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 48.024078369140625,
   "xla_arg_mb": 50.015625,
   "xla_out_mb": 2.0,
   "compile_s": 0.06772828599787317,
   "inner": 1,
   "median_ms": 68.86914400092792,
   "q1_ms": 68.59020649790182,
   "q3_ms": 69.78073800200946,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.14013671875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 66.07530975341797,
   "xla_arg_mb": 146.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.21780660900549265,
   "inner": 1,
   "median_ms": 233.4318620050908,
   "q1_ms": 227.53166850088746,
   "q3_ms": 297.7554755016172,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 263.2682189941406,
   "xla_arg_mb": 386.125,
   "xla_out_mb": 2.0,
   "compile_s": 0.15708473999984562,
   "inner": 1,
   "median_ms": 175.40315599762835,
   "q1_ms": 173.22005100140814,
   "q3_ms": 176.9396885028982,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 512.541618347168,
   "xla_arg_mb": 1154.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.7349766869956511,
   "inner": 1,
   "median_ms": 829.5133409992559,
   "q1_ms": 809.3238399997063,
   "q3_ms": 937.6304334982706,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 264.5191955566406,
   "xla_arg_mb": 386.125,
   "xla_out_mb": 2.0,
   "compile_s": 0.19094630199833773,
   "inner": 1,
   "median_ms": 210.96381400275277,
   "q1_ms": 204.34846349962754,
   "q3_ms": 213.2167310010118,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 512.792106628418,
   "xla_arg_mb": 1154.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.9365816040008212,
   "inner": 1,
   "median_ms": 1039.6210760009126,
   "q1_ms": 948.5641445025976,
   "q3_ms": 1156.9303994983784,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 267.0211486816406,
   "xla_arg_mb": 386.125,
   "xla_out_mb": 2.0,
   "compile_s": 0.24550705600267975,
   "inner": 1,
   "median_ms": 270.38285899470793,
   "q1_ms": 262.1337524979026,
   "q3_ms": 370.536850499775,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 513.293083190918,
   "xla_arg_mb": 1154.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.9761786279996159,
   "inner": 1,
   "median_ms": 1118.09177500254,
   "q1_ms": 1108.6604570009513,
   "q3_ms": 1129.740693999338,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 272.0250549316406,
   "xla_arg_mb": 386.125,
   "xla_out_mb": 2.0,
   "compile_s": 0.5520571509987349,
   "inner": 1,
   "median_ms": 392.14981099939905,
   "q1_ms": 381.3289040008385,
   "q3_ms": 396.72711599996546,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 1024,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.02197265625,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 514.295036315918,
   "xla_arg_mb": 1154.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.5398904489993583,
   "inner": 1,
   "median_ms": 1524.689899997611,
   "q1_ms": 1519.989753000118,
   "q3_ms": 1551.8868744984502,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 39.315093994140625,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.10324880700500216,
   "inner": 1,
   "median_ms": 98.63283400045475,
   "q1_ms": 98.23972549929749,
   "q3_ms": 99.17416849930305,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 64.65880584716797,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.34074294400488725,
   "inner": 1,
   "median_ms": 358.94887799804565,
   "q1_ms": 347.2316644983948,
   "q3_ms": 361.9549395007198,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 40.565582275390625,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.08462487199722091,
   "inner": 1,
   "median_ms": 79.20127199759008,
   "q1_ms": 78.35833150238614,
   "q3_ms": 81.85148499978823,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 64.90880584716797,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.28813060600077733,
   "inner": 1,
   "median_ms": 292.93206999864196,
   "q1_ms": 289.84682300142595,
   "q3_ms": 299.6133505002945,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 43.067291259765625,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.08512971800519153,
   "inner": 1,
   "median_ms": 83.67923400510335,
   "q1_ms": 82.48299499973655,
   "q3_ms": 84.43150649691233,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 65.40953826904297,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.29323638600180857,
   "inner": 1,
   "median_ms": 294.0386429982027,
   "q1_ms": 292.40344950085273,
   "q3_ms": 297.11071550264023,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 48.070953369140625,
   "xla_arg_mb": 56.015625,
   "xla_out_mb": 8.0,
   "compile_s": 0.10426509900571546,
   "inner": 1,
   "median_ms": 99.39048800151795,
   "q1_ms": 98.53476399803185,
   "q3_ms": 101.18118050013436,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.135498046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 66.41124725341797,
   "xla_arg_mb": 152.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.3175877160028904,
   "inner": 1,
   "median_ms": 323.5857840045355,
   "q1_ms": 321.68434650156996,
   "q3_ms": 328.8211385006434,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 263.3158264160156,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.21839420199830784,
   "inner": 1,
   "median_ms": 239.59243599529145,
   "q1_ms": 237.69018250095542,
   "q3_ms": 241.16272750325152,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 513.440788269043,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.8605699780018767,
   "inner": 1,
   "median_ms": 976.0376319973147,
   "q1_ms": 969.7804089955753,
   "q3_ms": 976.9710875007149,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 264.5663146972656,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.23442862700176192,
   "inner": 1,
   "median_ms": 251.6735319950385,
   "q1_ms": 251.23244800124667,
   "q3_ms": 254.4154655006423,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 513.690788269043,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 0.8853308830002788,
   "inner": 1,
   "median_ms": 1043.548686997383,
   "q1_ms": 1022.7077624986123,
   "q3_ms": 1048.3271515004162,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 267.0680236816406,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.27438775599875953,
   "inner": 1,
   "median_ms": 329.5704569973168,
   "q1_ms": 300.6735410017427,
   "q3_ms": 413.08508350266493,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 514.191520690918,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.2378869160020258,
   "inner": 1,
   "median_ms": 1385.6522140049492,
   "q1_ms": 1337.612106999586,
   "q3_ms": 1422.0526259996404,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 272.0719299316406,
   "xla_arg_mb": 392.125,
   "xla_out_mb": 8.0,
   "compile_s": 0.3675665909977397,
   "inner": 1,
   "median_ms": 452.6463949950994,
   "q1_ms": 416.9979064972722,
   "q3_ms": 483.47012999875005,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 4096,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.0203857421875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 515.193473815918,
   "xla_arg_mb": 1160.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.472061863001727,
   "inner": 1,
   "median_ms": 1514.738613004738,
   "q1_ms": 1493.532898999547,
   "q3_ms": 1528.5661635025463,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 39.505523681640625,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.3150339950007037,
   "inner": 1,
   "median_ms": 291.8708229990443,
   "q1_ms": 290.9704789999523,
   "q3_ms": 293.585195500782,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 129.75316619873047,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.9644806869982858,
   "inner": 1,
   "median_ms": 944.4657909989473,
   "q1_ms": 943.5919675015612,
   "q3_ms": 948.1250190037827,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 40.754547119140625,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.24293056199530838,
   "inner": 1,
   "median_ms": 235.3532159977476,
   "q1_ms": 234.23315200125217,
   "q3_ms": 239.21584850540967,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 129.75279998779297,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.7135463039958267,
   "inner": 1,
   "median_ms": 759.4997120031621,
   "q1_ms": 758.7429855011578,
   "q3_ms": 769.7836084989831,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 43.255523681640625,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.20677501899626805,
   "inner": 1,
   "median_ms": 205.5939140045666,
   "q1_ms": 203.05599250059458,
   "q3_ms": 207.15401049892535,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 129.75182342529297,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.6664081010021619,
   "inner": 1,
   "median_ms": 695.8184229952167,
   "q1_ms": 684.2816669995955,
   "q3_ms": 709.2219119986112,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 48.258941650390625,
   "xla_arg_mb": 80.015625,
   "xla_out_mb": 32.0,
   "compile_s": 0.20409721300529782,
   "inner": 1,
   "median_ms": 202.55173899931833,
   "q1_ms": 201.28669549740152,
   "q3_ms": 204.17315650047385,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 8,
   "top_k": 2,
   "max_expert_frac": 0.129852294921875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 129.75133514404297,
   "xla_arg_mb": 176.04687881469727,
   "xla_out_mb": 144.04698944091797,
   "compile_s": 0.6277639600011753,
   "inner": 1,
   "median_ms": 659.9013009981718,
   "q1_ms": 650.9711865037389,
   "q3_ms": 662.8996495055617,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 263.5062561035156,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.49234649400023045,
   "inner": 1,
   "median_ms": 455.42641400243156,
   "q1_ms": 450.5328710001777,
   "q3_ms": 460.70072750080726,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 517.131217956543,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.6750608689981163,
   "inner": 1,
   "median_ms": 1613.8634590024594,
   "q1_ms": 1573.0325205040572,
   "q3_ms": 1623.0725510031334,
   "chunk": 128
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 264.7552795410156,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.3865265120039112,
   "inner": 1,
   "median_ms": 407.0726330028265,
   "q1_ms": 405.8855069997662,
   "q3_ms": 410.1975354969909,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 517.379753112793,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.3341965559957316,
   "inner": 1,
   "median_ms": 1466.4195000004838,
   "q1_ms": 1462.5535139966814,
   "q3_ms": 1519.1822855013015,
   "chunk": 256
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 267.2562561035156,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.3752814670006046,
   "inner": 1,
   "median_ms": 390.7224210051936,
   "q1_ms": 389.1647480013489,
   "q3_ms": 392.8782190014317,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 517.879753112793,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.3272362560019246,
   "inner": 1,
   "median_ms": 1539.6128830034286,
   "q1_ms": 1495.412350501283,
   "q3_ms": 1540.3840575017966,
   "chunk": 512
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "forward",
   "xla_temp_mb": 272.2596740722656,
   "xla_arg_mb": 416.125,
   "xla_out_mb": 32.0,
   "compile_s": 0.48095125499821734,
   "inner": 1,
   "median_ms": 528.2179480054765,
   "q1_ms": 513.3879919994797,
   "q3_ms": 532.6763585035224,
   "chunk": 1024
  },
  {
   "code": "rejected",
   "run": "cpu_chunk_v2",
   "suite": "chunk",
   "routing": "dropless",
   "impl": "chunked",
   "dispatch": "sorted",
   "dtype": "float32",
   "tokens": 16384,
   "d_model": 512,
   "d_hidden": 1024,
   "experts": 64,
   "top_k": 2,
   "max_expert_frac": 0.019744873046875,
   "empty_experts": 0,
   "op": "train_step",
   "xla_temp_mb": 518.881217956543,
   "xla_arg_mb": 1184.3750038146973,
   "xla_out_mb": 1152.375114440918,
   "compile_s": 1.7567978669976583,
   "inner": 1,
   "median_ms": 1880.3966240011505,
   "q1_ms": 1861.583335998148,
   "q3_ms": 1881.8787884993071,
   "chunk": 1024
  }
 ]
}