feat: Configure ROCm attention via --attention-backend CLI argument, disable the Ray dashboard, and make eager mode configurable for cluster benchmarks.

scripts/cluster_manager.py

@@ -59,7 +59,7 @@ def setup_worker_node(worker_ip, head_ip):
     export NCCL_IB_TIMEOUT=23
     export NCCL_IB_RETRY_CNT=7
     echo "Starting Ray Worker on {worker_ip} connecting to {head_ip}..."
-    ray start --address='{head_ip}:6379' --num-gpus=1 --num-cpus=8 --disable-usage-stats
+    ray start --address='{head_ip}:6379' --num-gpus=1 --num-cpus=8 --disable-usage-stats --include-dashboard=false
     """
     
     print(f"Setting up Worker Node ({worker_ip})...")
@@ -97,7 +97,7 @@ def setup_head_node(head_ip):
     export NCCL_IB_TIMEOUT=23
     export NCCL_IB_RETRY_CNT=7
     echo "Starting Ray Head on {head_ip}..."
-    ray start --head --port=6379 --node-ip-address={head_ip} --num-gpus=1 --num-cpus=8 --disable-usage-stats
+    ray start --head --port=6379 --node-ip-address={head_ip} --num-gpus=1 --num-cpus=8 --disable-usage-stats --include-dashboard=false
     """
     
     try:

scripts/start_vllm.py

@@ -309,9 +309,7 @@ def configure_and_launch(model_idx, gpu_count):
     env.update(config.get("env", {}))
     
     if use_rocm_attn:
-        env["VLLM_V1_USE_PREFILL_DECODE_ATTENTION"] = "1"
-        env["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
-        # Optional: Explicitly mention these in print
+        cmd.extend(["--attention-backend", "ROCM_ATTN"])
         
     
     print("\n" + "="*60)