add perf scripts

requirements/rocm.txt

@@ -16,6 +16,7 @@ tensorizer>=2.9.0
 runai-model-streamer==0.11.0
 runai-model-streamer-s3==0.11.0
 setuptools_scm>=8
+cmake==3.29
 
 torch == 2.4.1
 triton == 3.0.0

setup.py

@@ -743,7 +743,7 @@ package_data = {
         "py.typed",
         "model_executor/layers/fused_moe/configs/*.json",
         "model_executor/layers/quantization/utils/configs/*.json",
-        "benchmarks/*.py",
+        "perf/*.py",
         "attention/backends/configs/*.json",
         "model_executor/layers/quantization/configs/awq/*.json"
     ]