Merge pull request #176 from FluffyAIcode/AgentMemory/agent-gan-safe-default-0716

fix(agents): fit default GAN cycle to worker memory

Author: FluffyAIcode

docs/ops/distributed-prefill-kv-network.md

@@ -308,7 +308,7 @@ Run two logical agents through real multi-round model inference:
 
 ```bash
 bash scripts/run_agent_gan_demo.sh \
-  --rounds 2 \
+  --rounds 1 \
   --output-tokens 64 \
   --report /tmp/kakeya-agent-gan-demo.json
 ```
@@ -320,6 +320,10 @@ persisted reports contain only output length/hash and metrics. The report
 separates inference-only KV hit rate from whole-workload hit rate including
 warmup, plus per-agent and aggregate token throughput/latency.
 
+One Generator/Critic cycle is the safe default for the current 16GB allens
+Gemma worker. Additional rounds grow agent histories and may exceed the strict
+remote-prefill timeout; increase worker memory/timeout before enabling them.
+
 ## Rollback
 
 The cache is an optimization; inference correctness does not depend on it.

scripts/agent_gan_inference_demo.py

@@ -68,7 +68,13 @@ def main() -> int:
     parser.add_argument("--dashboard", default="http://127.0.0.1:8090")
     parser.add_argument("--api-key-file", default="~/.kakeya/network_api_key")
     parser.add_argument("--tokenizer-id", required=True)
-    parser.add_argument("--rounds", type=int, default=2)
+    parser.add_argument(
+        "--rounds",
+        type=int,
+        default=1,
+        help="Generator/Critic cycles. The 16GB Gemma worker supports one "
+             "reliably; larger values require more worker memory or timeout.",
+    )
     parser.add_argument("--output-tokens", type=int, default=64)
     parser.add_argument("--report", default="/tmp/kakeya-agent-gan-demo.json")
     parser.add_argument("--skip-ensure", action="store_true")