|
{ |
|
"batch_size": 2, |
|
"sequence_length": 16, |
|
"num_new_tokens": 100, |
|
"num_return_sequences": 1, |
|
"prefill": { |
|
"memory": { |
|
"max_vram_used(MB)": 7230, |
|
"max_memory_reserved(MB)": 4332, |
|
"max_memory_allocated(MB)": 4031 |
|
}, |
|
"energy": { |
|
"cpu_energy(kHh)": 3.9784163236618045e-06, |
|
"gpu_energy(kHh)": 2.98222460770603e-06, |
|
"ram_energy(kHh)": 1.6811580245151466e-08, |
|
"total(kHh)": 6.977452511612986e-06 |
|
} |
|
}, |
|
"decode": { |
|
"memory": { |
|
"max_vram_used(MB)": 7358, |
|
"max_memory_reserved(MB)": 4460, |
|
"max_memory_allocated(MB)": 4288 |
|
}, |
|
"energy": { |
|
"cpu_energy(kHh)": 0.00012946108728647233, |
|
"gpu_energy(kHh)": 0.00023118935161825505, |
|
"ram_energy(kHh)": 5.490961296816485e-07, |
|
"total(kHh)": 0.000361199535034409 |
|
} |
|
} |
|
} |