Files
aituner/runs/frontier-qwen30-vllm020-profile-v1/frontier_vllm020_compat.py

60 lines
1.7 KiB
Python

#!/usr/bin/env python3
"""Run Frontier's linear profiler against the vLLM 0.20 RoPE API.
Frontier passes the pre-v0.20 ``get_rope`` arguments separately, while vLLM
0.20 carries the same values in ``rope_parameters``. Keep both repositories
unchanged and adapt only that call boundary for the profiling experiment.
"""
from __future__ import annotations
import inspect
from typing import Any
import torch
from vllm.model_executor.layers.rotary_embedding import get_rope as vllm_get_rope
from frontier.profiling.common.layers import rotary_embedding as frontier_rope
def _vllm020_get_rope_adapter(
*,
head_size: int,
rotary_dim: int,
max_position: int,
base: int | float,
is_neox_style: bool,
rope_scaling: dict[str, Any] | None,
dtype: torch.dtype | None = None,
) -> Any:
rope_parameters = dict(rope_scaling or {})
rope_parameters["rope_theta"] = base
rope_parameters["rope_dim"] = rotary_dim
return vllm_get_rope(
head_size=head_size,
max_position=max_position,
is_neox_style=is_neox_style,
rope_parameters=rope_parameters,
dtype=dtype,
)
def main() -> None:
parameters = inspect.signature(vllm_get_rope).parameters
if "rope_parameters" not in parameters or "rotary_dim" in parameters:
raise RuntimeError(
"Expected the vLLM 0.20 get_rope API with rope_parameters; "
f"found {inspect.signature(vllm_get_rope)}"
)
frontier_rope._VLLM_GET_ROPE = _vllm020_get_rope_adapter
frontier_rope._VLLM_GET_ROPE_IMPORT_ERROR = None
from frontier.profiling.linear_op.main import main as frontier_main
frontier_main()
if __name__ == "__main__":
main()