"""
Replay Slice cases from the performance database on Ascend NPU.
Purpose:
Read Slice rows from
profiling_database/data/{device}/vllm_ascend/{version}/Slice.csv,
rebuild the recorded source tensor, infer zero offsets from the current
dataset, then execute torch_npu.npu_slice().
"""
from __future__ import annotations
try:
from .common import parse_list_field, parse_shape
from .replay_framework import OpReplay
except ImportError:
from common import parse_list_field, parse_shape
from replay_framework import OpReplay
def build_case(row: dict[str, str]):
inputs = op.build_inputs(row)
if len(inputs) < 1:
raise ValueError("Slice requires at least one recorded input tensor")
output_shapes = [parse_shape(item) for item in parse_list_field(row["Output Shapes"])]
if not output_shapes:
raise ValueError("Slice requires at least one recorded output shape")
output_shape = output_shapes[0]
return {
"inputs": [inputs[0]],
"kwargs": {
"offsets": [0] * len(output_shape),
"sizes": list(output_shape),
},
"api": op.resolve_api(),
}
def run_case(case):
source = case["inputs"][0]
return case["api"](source, case["kwargs"]["offsets"], case["kwargs"]["sizes"])
def format_success(csv_path, row_index: int, _row: dict[str, str], case, result) -> str:
source = case["inputs"][0]
offsets = case["kwargs"]["offsets"]
sizes = case["kwargs"]["sizes"]
return (
f"[OK] {csv_path}:{row_index} "
f"source={tuple(source.shape)} output={tuple(result.shape)} "
f"offsets={offsets} sizes={sizes}"
)
op = OpReplay(
kernel_type="Slice",
api_path="torch_npu.npu_slice",
description=(
"Run Slice workload replay on Ascend NPU.\n"
"The script reads Slice.csv under the selected device and\n"
"vllm_ascend version directory, reconstructs the source tensor,\n"
"infers zero-based offsets from the current dataset, then runs\n"
"torch_npu.npu_slice(input, offsets, sizes)."
),
usage_examples=[
"py -3 tools/perf_data_collection/op_replay/Slice_run.py "
"--device ATLAS_800_A3_752T_128G_DIE --vllm-version 0.13.0",
],
version_help="vLLM-Ascend version, e.g. 0.13.0.",
build_case=build_case,
run_case=run_case,
format_success=format_success,
)
def main() -> None:
op.main()
if __name__ == "__main__":
main()