Spaces:
Runtime error
Runtime error
Update Feather H200 training runtime image
Browse files
overlay/scripts/__pycache__/benchmark_hyena_stack.cpython-312.pyc
CHANGED
|
Binary files a/overlay/scripts/__pycache__/benchmark_hyena_stack.cpython-312.pyc and b/overlay/scripts/__pycache__/benchmark_hyena_stack.cpython-312.pyc differ
|
|
|
overlay/scripts/benchmark_hyena_stack.py
CHANGED
|
@@ -78,6 +78,69 @@ CONFIGS = {
|
|
| 78 |
# best config to get more aggressive reclamation.
|
| 79 |
"HYDRA_HTM_SUBSAMPLE": "128",
|
| 80 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
}
|
| 82 |
|
| 83 |
|
|
|
|
| 78 |
# best config to get more aggressive reclamation.
|
| 79 |
"HYDRA_HTM_SUBSAMPLE": "128",
|
| 80 |
},
|
| 81 |
+
# A10 diagnostic profiles. These are intentionally explicit so that a
|
| 82 |
+
# benchmark result says exactly which expensive subsystem was active.
|
| 83 |
+
"a10_full": {
|
| 84 |
+
"HYDRA_BATCH_SIZE": "8",
|
| 85 |
+
"HYDRA_HYENA_LAYERS": "3,7",
|
| 86 |
+
"HYDRA_HYENA_FLASH_FFT": "0",
|
| 87 |
+
"HYDRA_HYENA_TRAIN_CACHE": "0",
|
| 88 |
+
"HYDRA_HYENA_FILTER_CACHE": "0",
|
| 89 |
+
"HYDRA_ABLATE_HTM": "0",
|
| 90 |
+
"HYDRA_ABLATE_MIDSTACK": "0",
|
| 91 |
+
"HYDRA_ABLATE_SDR_PROJECT": "0",
|
| 92 |
+
"HYDRA_DISABLE_TRITON_SDR_PROJECT": "1",
|
| 93 |
+
"HYDRA_FORCE_HTM_CPU": "0",
|
| 94 |
+
},
|
| 95 |
+
"a10_no_htm": {
|
| 96 |
+
"HYDRA_BATCH_SIZE": "8",
|
| 97 |
+
"HYDRA_HYENA_LAYERS": "3,7",
|
| 98 |
+
"HYDRA_HYENA_FLASH_FFT": "0",
|
| 99 |
+
"HYDRA_HYENA_TRAIN_CACHE": "0",
|
| 100 |
+
"HYDRA_HYENA_FILTER_CACHE": "0",
|
| 101 |
+
"HYDRA_ABLATE_HTM": "1",
|
| 102 |
+
"HYDRA_ABLATE_MIDSTACK": "0",
|
| 103 |
+
"HYDRA_ABLATE_SDR_PROJECT": "0",
|
| 104 |
+
"HYDRA_DISABLE_TRITON_SDR_PROJECT": "1",
|
| 105 |
+
"HYDRA_FORCE_HTM_CPU": "0",
|
| 106 |
+
},
|
| 107 |
+
"a10_no_htm_no_sdr": {
|
| 108 |
+
"HYDRA_BATCH_SIZE": "8",
|
| 109 |
+
"HYDRA_HYENA_LAYERS": "3,7",
|
| 110 |
+
"HYDRA_HYENA_FLASH_FFT": "0",
|
| 111 |
+
"HYDRA_HYENA_TRAIN_CACHE": "0",
|
| 112 |
+
"HYDRA_HYENA_FILTER_CACHE": "0",
|
| 113 |
+
"HYDRA_ABLATE_HTM": "1",
|
| 114 |
+
"HYDRA_ABLATE_MIDSTACK": "0",
|
| 115 |
+
"HYDRA_ABLATE_SDR_PROJECT": "1",
|
| 116 |
+
"HYDRA_DISABLE_TRITON_SDR_PROJECT": "1",
|
| 117 |
+
"HYDRA_FORCE_HTM_CPU": "0",
|
| 118 |
+
},
|
| 119 |
+
"a10_no_htm_no_midstack": {
|
| 120 |
+
"HYDRA_BATCH_SIZE": "8",
|
| 121 |
+
"HYDRA_HYENA_LAYERS": "3,7",
|
| 122 |
+
"HYDRA_HYENA_FLASH_FFT": "0",
|
| 123 |
+
"HYDRA_HYENA_TRAIN_CACHE": "0",
|
| 124 |
+
"HYDRA_HYENA_FILTER_CACHE": "0",
|
| 125 |
+
"HYDRA_ABLATE_HTM": "1",
|
| 126 |
+
"HYDRA_ABLATE_MIDSTACK": "1",
|
| 127 |
+
"HYDRA_ABLATE_SDR_PROJECT": "1",
|
| 128 |
+
"HYDRA_DISABLE_TRITON_SDR_PROJECT": "1",
|
| 129 |
+
"HYDRA_FORCE_HTM_CPU": "0",
|
| 130 |
+
},
|
| 131 |
+
"a10_all_hyena_fast": {
|
| 132 |
+
"HYDRA_BATCH_SIZE": "4",
|
| 133 |
+
"HYDRA_TOTAL_BATCH": "4096",
|
| 134 |
+
"HYDRA_HYENA_LAYERS": "0,1,2,3",
|
| 135 |
+
"HYDRA_HYENA_FLASH_FFT": "0",
|
| 136 |
+
"HYDRA_HYENA_TRAIN_CACHE": "0",
|
| 137 |
+
"HYDRA_HYENA_FILTER_CACHE": "0",
|
| 138 |
+
"HYDRA_ABLATE_HTM": "1",
|
| 139 |
+
"HYDRA_ABLATE_MIDSTACK": "1",
|
| 140 |
+
"HYDRA_ABLATE_SDR_PROJECT": "1",
|
| 141 |
+
"HYDRA_DISABLE_TRITON_SDR_PROJECT": "1",
|
| 142 |
+
"HYDRA_FORCE_HTM_CPU": "0",
|
| 143 |
+
},
|
| 144 |
}
|
| 145 |
|
| 146 |
|