"""
Replay ReshapeAndCacheNdKernel cases from the performance database on Ascend NPU.
Purpose:
Read ReshapeAndCacheNdKernel rows from
profiling_database/data/{device}/vllm_ascend/{version}/ReshapeAndCacheNdKernel.csv,
rebuild the recorded tensor inputs, construct a legal slot_mapping tensor,
then execute torch_npu._npu_reshape_and_cache().
"""
from __future__ import annotations
try:
from .common import get_runtime_modules, parse_list_field, parse_shape
from .replay_framework import OpReplay
except ImportError:
from common import get_runtime_modules, parse_list_field, parse_shape
from replay_framework import OpReplay
def build_slot_mapping_tensor(
slot_mapping_shape: tuple[int, ...],
key_cache_shape: tuple[int, ...],
):
runtime_torch, _ = get_runtime_modules()
if len(slot_mapping_shape) != 1:
raise ValueError(f"slot_mapping must be 1D, got shape={slot_mapping_shape}")
if len(key_cache_shape) < 2:
raise ValueError(f"key_cache rank must be >= 2, got shape={key_cache_shape}")
token_count = slot_mapping_shape[0]
total_slots = key_cache_shape[0] * key_cache_shape[1]
if token_count > total_slots:
raise ValueError(
"slot_mapping token count exceeds cache capacity: "
f"tokens={token_count}, total_slots={total_slots}, key_cache_shape={key_cache_shape}"
)
permutation = runtime_torch.randperm(total_slots, dtype=runtime_torch.int64)[:token_count]
return permutation.to(dtype=runtime_torch.int32, device="npu")
def build_case(row: dict[str, str]):
inputs = op.build_inputs(row)
input_shapes = [parse_shape(item) for item in parse_list_field(row["Input Shapes"])]
if len(inputs) != 5 or len(input_shapes) != 5:
raise ValueError("ReshapeAndCacheNdKernel expects exactly five inputs")
return {
"inputs": inputs[:4] + [
build_slot_mapping_tensor(
slot_mapping_shape=input_shapes[4],
key_cache_shape=input_shapes[2],
)
],
"kwargs": {},
"api": op.resolve_api(),
}
def format_success(csv_path, row_index: int, row: dict[str, str], case, _result) -> str:
key_cache = case["inputs"][2]
value_cache = case["inputs"][3]
return (
f"[OK] {csv_path}:{row_index} "
f"shapes={row['Input Shapes']} formats={row['Input Formats']} "
f"dtypes={row['Input Data Types']} "
f"key_cache={tuple(key_cache.shape)} value_cache={tuple(value_cache.shape)}"
)
op = OpReplay(
kernel_type="ReshapeAndCacheNdKernel",
api_path="torch_npu._npu_reshape_and_cache",
description=(
"Run ReshapeAndCacheNdKernel workload replay on Ascend NPU.\n"
"The script reads ReshapeAndCacheNdKernel.csv under the selected\n"
"device and vllm_ascend version directory, reconstructs input\n"
"tensors from Input Shapes / Input Formats / Input Data Types,\n"
"builds a legal slot_mapping tensor, then runs\n"
"torch_npu._npu_reshape_and_cache()."
),
usage_examples=[
"py -3 tools/perf_data_collection/op_replay/ReshapeAndCacheNdKernel_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,
format_success=format_success,
)
def main() -> None:
op.main()
if __name__ == "__main__":
main()