config: Add VLLM_DISABLE_COMPILE_CACHE=1 to environment variables across VLLM scripts.

benchmarks/find_max_context.py

@@ -137,6 +137,7 @@ def get_vllm_server_cmd(model, tp_size, util, max_len, max_seqs):
     
     # Env Setup
     env = os.environ.copy()
+    env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
     env.update(config.get("env", {}))
 
     # CLUSTER / RAY LOGIC

benchmarks/run_vllm_bench.py

@@ -163,6 +163,7 @@ def run_throughput(model, tp_size, backend_name="Default", output_dir=RESULTS_DI
 
     # ENV Setup: Global + Model Specific
     env = os.environ.copy()
+    env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
     
     # Inject model specific env vars (e.g. for AWQ)
     model_env = MODEL_TABLE[model].get("env", {})

benchmarks/vllm_cluster_bench.py

@@ -140,6 +140,7 @@ def get_cluster_env():
     host_ip = get_local_ip(rdma_iface)
     
     env = os.environ.copy()
+    env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
     
     # Critical Cluster Envs (Match start_vllm_cluster.py)
     env["RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES"] = "1"

scripts/start_vllm.py

@@ -322,6 +322,7 @@ def configure_and_launch(model_idx, gpu_count):
     
     # Env Vars
     env = os.environ.copy()
+    env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
     env.update(config.get("env", {}))
     
     if use_rocm_attn:

scripts/start_vllm_cluster.py

@@ -263,6 +263,7 @@ def configure_and_launch_vllm(model_idx, head_ip):
     print(f"Detected RDMA Interface: {rdma_iface}")
     
     env = os.environ.copy()
+    env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
     env["RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES"] = "1"
     env["VLLM_HOST_IP"] = head_ip
     env["NCCL_SOCKET_IFNAME"] = rdma_iface