kv_cache_scales.json 3.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990
  1. {
  2. "model_type": "llama",
  3. "kv_cache": {
  4. "dtype": "float8_e4m3fn",
  5. "scaling_factor": {
  6. "0": {
  7. "0": 0.0230364128947258,
  8. "1": 0.01979283057153225,
  9. "2": 0.0241350457072258,
  10. "3": 0.0308314748108387,
  11. "4": 0.0430733822286129,
  12. "5": 0.0370396226644516,
  13. "6": 0.0306222103536129,
  14. "7": 0.0357491634786129,
  15. "8": 0.0358189195394516,
  16. "9": 0.0443289652466774,
  17. "10": 0.0433175228536129,
  18. "11": 0.0416782945394516,
  19. "12": 0.0366908498108387,
  20. "13": 0.0432477705180645,
  21. "14": 0.0410505048930645,
  22. "15": 0.0457589291036129,
  23. "16": 0.0418526791036129,
  24. "17": 0.0432477705180645,
  25. "18": 0.0469447560608387,
  26. "19": 0.0514787957072258,
  27. "20": 0.0541294664144516,
  28. "21": 0.0587681382894516,
  29. "22": 0.0625,
  30. "23": 0.0585588738322258,
  31. "24": 0.0600237175822258,
  32. "25": 0.0588030144572258,
  33. "26": 0.0531180277466774,
  34. "27": 0.06396484375,
  35. "28": 0.0603027381002903,
  36. "29": 0.0582101047039032,
  37. "30": 0.0625348836183548,
  38. "31": 0.0585588738322258,
  39. "32": 0.0582798570394516,
  40. "33": 0.0575125589966774,
  41. "34": 0.0590820349752903,
  42. "35": 0.0614188089966774,
  43. "36": 0.0631975457072258,
  44. "37": 0.0615931935608387,
  45. "38": 0.0601283498108387,
  46. "39": 0.0571986623108387,
  47. "40": 0.0670340433716774,
  48. "41": 0.0523507259786129,
  49. "42": 0.0547223798930645,
  50. "43": 0.0631975457072258,
  51. "44": 0.0663713738322258,
  52. "45": 0.0603376142680645,
  53. "46": 0.0652204304933548,
  54. "47": 0.0734514519572258,
  55. "48": 0.0693708211183548,
  56. "49": 0.0725446492433548,
  57. "50": 0.0627790242433548,
  58. "51": 0.0691266804933548,
  59. "52": 0.0688825398683548,
  60. "53": 0.068429134786129,
  61. "54": 0.0605119988322258,
  62. "55": 0.0799386203289032,
  63. "56": 0.0853097140789032,
  64. "57": 0.0661969929933548,
  65. "58": 0.0689871683716774,
  66. "59": 0.0724051371216774,
  67. "60": 0.0541643425822258,
  68. "61": 0.0626743882894516,
  69. "62": 0.0628487765789032,
  70. "63": 0.0607212632894516,
  71. "64": 0.0589076466858387,
  72. "65": 0.0451660193502903,
  73. "66": 0.0453055277466774,
  74. "67": 0.0414341539144516,
  75. "68": 0.0385044664144516,
  76. "69": 0.0414341539144516,
  77. "70": 0.0466308631002903,
  78. "71": 0.0399693101644516,
  79. "72": 0.0437011756002903,
  80. "73": 0.0434221550822258,
  81. "74": 0.0428989976644516,
  82. "75": 0.0401785746216774,
  83. "76": 0.0431082621216774,
  84. "77": 0.0484444759786129,
  85. "78": 0.0417829267680645,
  86. "79": 0.0418178029358387
  87. }
  88. }
  89. }
  90. }