"""
Replay SoftmaxV2 cases from the performance database on Ascend NPU.
Purpose:
Read SoftmaxV2 rows from
profiling_database/data/{device}/vllm_ascend/{version}/SoftmaxV2.csv,
rebuild the recorded tensor inputs, then execute
torch.nn.functional.softmax() along the last dimension.
"""
from __future__ import annotations
try:
from .replay_framework import OpReplay
except ImportError:
from replay_framework import OpReplay
def format_success(csv_path, row_index: int, row: dict[str, str], case, result) -> str:
input_tensor = case["inputs"][0]
return (
f"[OK] {csv_path}:{row_index} "
f"shape={tuple(input_tensor.shape)} dtype={row['Input Data Types']} "
f"format={row['Input Formats']} output={tuple(result.shape)} dim=-1"
)
op = OpReplay(
kernel_type="SoftmaxV2",
api_path="torch.nn.functional.softmax",
description=(
"Run SoftmaxV2 workload replay on Ascend NPU.\n"
"The script reads SoftmaxV2.csv under the selected device and\n"
"vllm_ascend version directory, reconstructs input tensors from\n"
"Input Shapes / Input Formats / Input Data Types, then runs\n"
"torch.nn.functional.softmax(input, dim=-1)."
),
usage_examples=[
"py -3 tools/perf_data_collection/op_replay/SoftmaxV2_run.py "
"--device ATLAS_800_A3_752T_128G_DIE --vllm-version 0.13.0",
],
version_help="vLLM-Ascend version, e.g. 0.13.0.",
input_count=1,
fixed_kwargs={"dim": -1},
format_success=format_success,
)
def main() -> None:
op.main()
if __name__ == "__main__":
main()