[ROCm][CI] Fix tool use test stability - disable skinny GEMM, prefix caching, eliminate batch variance (#35553)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

tests/utils.py

@@ -109,6 +109,20 @@ else:
 VLLM_PATH = Path(__file__).parent.parent
 """Path to root of the vLLM repository."""
 
+# ROCm: disable skinny GEMM to avoid non-deterministic results from
+# atomic reductions in wvSplitKrc kernel.
+# See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3906083975
+ROCM_ENV_OVERRIDES = (
+    {"VLLM_ROCM_USE_SKINNY_GEMM": "0"} if current_platform.is_rocm() else {}
+)
+# ROCm: disable prefix caching and eliminate batch variance to reduce
+# test flakiness.
+ROCM_EXTRA_ARGS = (
+    ["--no-enable-prefix-caching", "--max-num-seqs", "1"]
+    if current_platform.is_rocm()
+    else []
+)
+
 
 class RemoteVLLMServer:
     """Base class for launching vLLM server subprocesses for testing.