[cuda-graphs] Add support for --cuda-graphs argument and update default behavior for CUDA Graphs

Author: orionpapadakis

llama-tornado

@@ -93,7 +93,9 @@ class LlamaRunner:
         if args.batch_prefill_size is not None:
             cmd.append(f"-Dllama.prefillBatchSize={args.batch_prefill_size}")
 
-        if args.no_cuda_graphs:
+        if args.cuda_graphs:
+            cmd.append("-Dllama.cudaGraphs=true")
+        elif args.no_cuda_graphs:
             cmd.append("-Dllama.cudaGraphs=false")
 
         # Debug options
@@ -505,11 +507,17 @@ def create_parser() -> argparse.ArgumentParser:
             "N>1: batched prefill processing N tokens per chunk (llama.prefillBatchSize=N)."
         ),
     )
+    prefill_group.add_argument(
+        "--cuda-graphs",
+        dest="cuda_graphs",
+        action="store_true",
+        help="Enable CUDA graph capture/replay (llama.cudaGraphs=true); PTX backend only",
+    )
     prefill_group.add_argument(
         "--no-cuda-graphs",
         dest="no_cuda_graphs",
         action="store_true",
-        help="Disable CUDA graph capture/replay (llama.cudaGraphs=false); useful for debugging",
+        help="Disable CUDA graph capture/replay (llama.cudaGraphs=false); no-op, disabled by default",
     )
 
     # Advanced options

src/main/java/org/beehive/gpullama3/tornadovm/TornadoVMMasterPlan.java

@@ -26,9 +26,9 @@ public interface TornadoVMMasterPlan {
     boolean ENABLE_TORNADOVM_INIT_TIME = Boolean.parseBoolean(
             System.getProperty("llama.EnableTimingForTornadoVMInit", "False"));
 
-    /** When {@code false}, {@code withCUDAGraph()} is never called — useful for debugging. */
+    /** When {@code true}, {@code withCUDAGraph()} is called — PTX/CUDA backend only. */
     boolean CUDA_GRAPHS = Boolean.parseBoolean(
-            System.getProperty("llama.cudaGraphs", "true"));
+            System.getProperty("llama.cudaGraphs", "false"));
 
     boolean WITH_PREFILL_DECODE = Boolean.getBoolean("llama.withPrefillDecode");