Adapt Frontier RoPE profiling to vLLM 0.20

2026-07-16 22:23:56 +08:00
parent 0e9b64893f
commit 9b54075e75
4 changed files with 68 additions and 7 deletions
--- a/runs/frontier-qwen30-vllm020-profile-v1/frontier_vllm020_compat.py
+++ b/runs/frontier-qwen30-vllm020-profile-v1/frontier_vllm020_compat.py
@@ -0,0 +1,59 @@
+#!/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()
--- a/runs/frontier-qwen30-vllm020-profile-v1/jobs_frontier_linear_smoke.toml
+++ b/runs/frontier-qwen30-vllm020-profile-v1/jobs_frontier_linear_smoke.toml
@@ -1,18 +1,18 @@
 version = 1

 [[jobs]]
-name = "qwen30-vllm020-frontier-linear-smoke-20260716-v1"
+name = "qwen30-vllm020-frontier-linear-smoke-20260716-v2-rope-api"
 gpus = 1
 gpu_model = "H20"
 hosts = ["dash0"]
 command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-qwen30-vllm020-profile-v1 && timeout --signal=TERM --kill-after=30s 1020 bash run_frontier_linear_smoke.sh"
-artifacts = ["artifacts/frontier-linear-smoke-v1"]
+artifacts = ["artifacts/frontier-linear-smoke-v2"]

 [jobs.env]
 HOME = "/tmp/wjh"
 XDG_CACHE_HOME = "/tmp/wjh/.cache"
 VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
-OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-fleet/artifacts/frontier-linear-smoke-v1"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-fleet/artifacts/frontier-linear-smoke-v2"
 VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
 FRONTIER_ROOT = "/home/admin/cpfs/wjh/frontier-qwen30-vllm020-profile-v1/Frontier"
 VLLM_SOURCE = "/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build"
--- a/runs/frontier-qwen30-vllm020-profile-v1/run_frontier_linear_full.sh
+++ b/runs/frontier-qwen30-vllm020-profile-v1/run_frontier_linear_full.sh
@@ -6,6 +6,7 @@ OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT must be set}"
 VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
 FRONTIER_ROOT="${FRONTIER_ROOT:-/home/admin/cpfs/wjh/frontier-qwen30-vllm020-profile-v1/Frontier}"
 VLLM_SOURCE="${VLLM_SOURCE:-/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build}"
+RUN_DIR="$(pwd -P)"
 TOKENS=(1 8 16 32 64 128 256 512 1024 2048 4096 8192)
 mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/profiles"
 exec > >(tee -a "${OUTPUT_ROOT}/logs/full.log") 2>&1
@@ -25,7 +26,7 @@ test "$(git -C "${VLLM_SOURCE}" rev-parse HEAD)" = "88d34c6409e9fb3c7b8ca0c04756
 git rev-parse HEAD > "${OUTPUT_ROOT}/provenance/aituner.commit"
 git -C "${FRONTIER_ROOT}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/frontier.commit"
 git -C "${VLLM_SOURCE}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/vllm-source.commit"
-sha256sum run_frontier_linear_full.sh \
+sha256sum run_frontier_linear_full.sh frontier_vllm020_compat.py \
  > "${OUTPUT_ROOT}/provenance/source.sha256"
 uv pip freeze --python "${VENV_ROOT}/bin/python" \
  > "${OUTPUT_ROOT}/provenance/pip-freeze.txt"
@@ -35,7 +36,7 @@ printf '%s\n' "${TOKENS[@]}" > "${OUTPUT_ROOT}/provenance/tokens.txt"

 cd "${FRONTIER_ROOT}"
 timeout --signal=TERM --kill-after=30s 1380 \
-  "${VENV_ROOT}/bin/python" -m frontier.profiling.linear_op.main \
+  "${VENV_ROOT}/bin/python" "${RUN_DIR}/frontier_vllm020_compat.py" \
  --disable_ray \
  --num_gpus 1 \
  --output_dir "${OUTPUT_ROOT}/profiles" \
--- a/runs/frontier-qwen30-vllm020-profile-v1/run_frontier_linear_smoke.sh
+++ b/runs/frontier-qwen30-vllm020-profile-v1/run_frontier_linear_smoke.sh
@@ -6,6 +6,7 @@ OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT must be set}"
 VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
 FRONTIER_ROOT="${FRONTIER_ROOT:-/home/admin/cpfs/wjh/frontier-qwen30-vllm020-profile-v1/Frontier}"
 VLLM_SOURCE="${VLLM_SOURCE:-/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build}"
+RUN_DIR="$(pwd -P)"
 mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/profiles"
 exec > >(tee -a "${OUTPUT_ROOT}/logs/smoke.log") 2>&1

@@ -24,7 +25,7 @@ test "$(git -C "${VLLM_SOURCE}" rev-parse HEAD)" = "88d34c6409e9fb3c7b8ca0c04756
 git rev-parse HEAD > "${OUTPUT_ROOT}/provenance/aituner.commit"
 git -C "${FRONTIER_ROOT}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/frontier.commit"
 git -C "${VLLM_SOURCE}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/vllm-source.commit"
-sha256sum run_frontier_linear_smoke.sh \
+sha256sum run_frontier_linear_smoke.sh frontier_vllm020_compat.py \
  > "${OUTPUT_ROOT}/provenance/source.sha256"
 uv pip freeze --python "${VENV_ROOT}/bin/python" \
  > "${OUTPUT_ROOT}/provenance/pip-freeze.txt"
@@ -33,7 +34,7 @@ nvidia-smi --query-gpu=index,uuid,name,driver_version,memory.total \

 cd "${FRONTIER_ROOT}"
 timeout --signal=TERM --kill-after=30s 780 \
-  "${VENV_ROOT}/bin/python" -m frontier.profiling.linear_op.main \
+  "${VENV_ROOT}/bin/python" "${RUN_DIR}/frontier_vllm020_compat.py" \
  --disable_ray \
  --num_gpus 1 \
  --output_dir "${OUTPUT_ROOT}/profiles" \