Initialize vLLM config for Frontier profiling

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

@@ -9,6 +9,7 @@ unchanged and adapt only that call boundary for the profiling experiment.
 from __future__ import annotations
 
 import inspect
+import os
 from typing import Any
 
 import torch
@@ -51,8 +52,20 @@ def main() -> None:
     frontier_rope._VLLM_GET_ROPE_IMPORT_ERROR = None
 
     from frontier.profiling.linear_op.main import main as frontier_main
+    from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
 
-    frontier_main()
+    model_root = os.environ.get("MODEL_ROOT")
+    if not model_root:
+        raise RuntimeError("MODEL_ROOT must point to the profiled Qwen checkpoint")
+    model_config = ModelConfig(
+        model=model_root,
+        dtype="bfloat16",
+        max_model_len=8192,
+        skip_tokenizer_init=True,
+        generation_config="vllm",
+    )
+    with set_current_vllm_config(VllmConfig(model_config=model_config)):
+        frontier_main()
 
 
 if __name__ == "__main__":