more fixes :(

Signed-off-by: jthomson04 <jwillthomson19@gmail.com>

Author: jthomson04

3rdparty/Gym-workspace/Gym

@@ -192,30 +192,25 @@ def _postprocess_nemo_gym_to_nemo_rl_result(
             if "generation_token_ids" not in output_item_dict:
                 continue
 
-            assert (
-                seen_token_ids
-                == output_item_dict["prompt_token_ids"][: len(seen_token_ids)]
-            ), f"""Non-contiguous messages found! This may be a tokenization issue where certain tokens are combined when messages are concatenated, or it may be due to part of the chat history being truncated (like if super long history is truncated or if reasoning is stripped out).
-Seen token IDs: {seen_token_ids}
-Output prompt token IDs: {output_item_dict["prompt_token_ids"]}
-"""
-
             nemo_rl_message_log.append(
                 {
                     "role": "user",
                     "content": "",
                     "token_ids": torch.tensor(
-                        output_item_dict["prompt_token_ids"][len(seen_token_ids) :]
+                        output_item_dict["prompt_token_ids"][len(seen_token_ids) :],
+                        dtype=torch.long,
                     ),
                 }
             )
             nemo_rl_message_log.append(
                 {
                     "role": "assistant",
                     "content": "",
-                    "token_ids": torch.tensor(output_item_dict["generation_token_ids"]),
+                    "token_ids": torch.tensor(
+                        output_item_dict["generation_token_ids"], dtype=torch.long
+                    ),
                     "generation_logprobs": torch.tensor(
-                        output_item_dict["generation_log_probs"]
+                        output_item_dict["generation_log_probs"], dtype=torch.float32
                     ),
                 }
             )

nemo_rl/environments/nemo_gym.py

@@ -198,6 +198,7 @@ def _dynamo_env_vars(self) -> dict[str, str]:
             "NATS_SERVER": f"nats://{self._host}:{self._nats_port}",
             "DYN_NAMESPACE": self._namespace,
             "DYN_DISCOVERY_BACKEND": "etcd",
+            "DYN_SDK_DISABLE_ANSI_LOGGING": "1",
             "DYN_LOG": "dynamo_llm::http::service::metrics=warn,dynamo_runtime::pipeline::network::ingress::push_handler=warn,dynamo_llm::http::service::service_v2=warn,info"
         }
 
@@ -298,7 +299,7 @@ def _start_planner(self) -> None:
 
         planner_config = {
             "environment": "virtual",
-            "mode": "decode",
+            "mode": "agg",
             "backend": "vllm",
             "namespace": self._namespace,
             "model_name": self.cfg.get("model", "unknown"),

nemo_rl/models/generation/dynamo/dynamo_generation.py

@@ -276,6 +276,8 @@ def __init__(
                 kv_events_config_json=kv_events_json,
                 seed=seed,
             ),
+            "--dyn-tool-call-parser", "hermes",
+            "--dyn-reasoning-parser", "qwen3"
         ]
 
         # --- Subprocess environment ---