xlnstorch.benchmark.BenchResult

class xlnstorch.benchmark.BenchResult(prof, wall_ms, p50, p90, p99, cpu_us=0.0, cpu_mem_mb=0.0, cuda_us=0.0, cuda_mem_mb=0.0)

A container holding the aggregate numbers of a single benchmark run.

Parameters:

• prof (profile | None)

• wall_ms (float)

• p50 (float)

• p90 (float)

• p99 (float)

• cpu_us (float)

• cpu_mem_mb (float)

• cuda_us (float)

• cuda_mem_mb (float)

prof

The profiler output, if profiling was enabled. Otherwise None.

Type:

Optional[torch.profiler.profile]

wall_ms

Mean wall-clock latency in milliseconds.

Type:

float

p50 / p90 / p99

Percentiles (ms) of the wall-clock distribution.

Type:

float

cpu_us / cuda_us

Self CPU / CUDA time reported by torch.profiler in micro-seconds. Values are zero when profile=False or the respective device is unavailable.

Type:

float

cpu_mem_mb / cuda_mem_mb

Peak memory (mb) reported by torch.profiler. Same fallback rules as above apply.

Type:

float

__init__(prof, wall_ms, p50, p90, p99, cpu_us=0.0, cpu_mem_mb=0.0, cuda_us=0.0, cuda_mem_mb=0.0)

Parameters:

• prof (profile | None)

• wall_ms (float)

• p50 (float)

• p90 (float)

• p99 (float)

• cpu_us (float)

• cpu_mem_mb (float)

• cuda_us (float)

• cuda_mem_mb (float)

Return type:

None

Methods

__init__(prof, wall_ms, p50, p90, p99[, ...])
print()	Nicely format the dataclass to stdout (monospaced columns).
save_full_profile(path[, group_by_stack_n, ...])

Attributes

cpu_mem_mb
cpu_us
cuda_mem_mb
cuda_us
prof
wall_ms
p50
p90
p99