docs-benchmarks
/
experts-backends
Batch Size
int64
1
1
Seq Length
int64
16
16
New Tokens
int64
16
64
Torch Compile
stringclasses
3 values
Implementation
stringclasses
4 values
Mean Generation Latency (ms)
float64
144
4.91k
Mean Prefill Latency (ms)
float64
52.2
289
Mean Decode Latency (ms)
float64
92
4.62k
Peak Mem (MB)
float64
27.3k
28.4k
1
16
16
False
eager
1,193.7
269.94
923.76
27,333.27
1
16
16
max-autotune-no-cudagraphs
eager
1,332.96
269.72
1,063.24
27,333.27
1
16
16
False
grouped_mm
814.61
64.32
750.29
27,333.27
1
16
16
max-autotune-no-cudagraphs
grouped_mm
476.82
63.42
413.41
27,333.27
1
16
16
False
batched_mm
535.34
52.18
483.17
28,386.68
1
16
16
max-autotune
batched_mm
144.21
52.25
91.97
28,386.68
1
16
16
False
grouped_prefill+batched_decode
579.92
64.83
515.1
27,396.05
1
16
16
max-autotune
grouped_prefill+batched_decode
162.56
64.71
97.84
27,332.98
1
16
64
False
eager
4,303.03
274.71
4,028.32
27,342.27
1
16
64
max-autotune-no-cudagraphs
eager
4,908.6
288.97
4,619.63
27,343.43
1
16
64
False
grouped_mm
3,300.09
64.67
3,235.42
27,343.43
1
16
64
max-autotune-no-cudagraphs
grouped_mm
1,801.67
64.25
1,737.41
27,342.27
1
16
64
False
batched_mm
2,151.96
52.23
2,099.73
28,395.68
1
16
64
max-autotune
batched_mm
434.84
52.25
382.58
28,395.68
1
16
64
False
grouped_prefill+batched_decode
2,217.35
64.99
2,152.36
27,405.06
1
16
64
max-autotune
grouped_prefill+batched_decode
465.11
64.5
400.62
27,341.98